Understanding Linux On System Z Performance Tools

Christian Ehrhardt
Linux Performance know-it-all series
How to Surprise by being a

Linux Performance "know-it-all"
Christian Ehrhardt, IBM R&D Germany,

System Performance Analyst
© 2013 IBM Corporation

Linux on System z Performance Evaluation
Trademarks
IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International
Business Machines Corp., registered in many jurisdictions worldwide. Other product and service
names might be trademarks of IBM or other companies. A current list of IBM trademarks is
available on the Web at www.ibm.com/legal/copytrade.shtml.
2 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

4
Agenda
 Your swiss army knife for the complex cases – Netstat – network statistics and overview
– Pidstat – per process statistics – Socket Statistics – extended socket statistics
– Slabtop – kernel memory pool consumption – top / ps – process overview
– Lsof – check file flags of open files – Icastats / lszcrypt – check usage of crypto hw support
– Blktrace – low level disk I/O analysis – Lsluns / multipath – check multipath setup
– Hyptop – cross guest cpu consumption monitor – Lsqeth – check hw checksumming and buffer count
– Iptraf - network traffic monitor – Ethtool – check offloading functions
– Dstat – very configurable live system overview – Collectl – full system monitoring
– Irqstats – check irq amount and cpu distribution – Ftrace – kernel function tracing
– Smem – per process/per mapping memory overview – Lttng – complex latency tracing infrastructure
– Jinsight – Java method call stack analysis – Ziomon – Analyze FCP setup and I/O
– Htop – top on steroids – Systemtap – another kernel tracing infrastructure
– Strace – system call statistics – Wireshark / Tcpdump – analyze network traffic in depth
– Ltrace – library call statistics – Iotop – order processes by disk I/O
– Kernel tracepoints – get in-depth timing inside the kernel – Iftop - per connection traffic overview
– Vmstat – virtual memory statistics … ever growing
– Sysstat – full system overview
– Iostat – I/O related statistics
– Dasdstat – disk statistics
– scsi statistics – disk statistics
– Perf – hw counters, tracepoint based evaluations, profiling to find hotspots
– Valgrind – in depth memory/cache analysis and leak detection
– Java Health Center – high level java overview and monitoring
– Java Garbage Collection and Memory visualizer – in depth gc analysis
4
Structured Agenda
Basic Intermediate Advanced Master Elite

– Utilization – General – Strace – Perf – Cachestat
– Scheduling thoughts – Ltrace – slabtop – Smem
– Page Cache – Sysstat – Lsof – Blktrace – Valgrind
– Swapping – Dasdstat – Lsluns – Ziomon – Irqstats
– Scsi I/O – Multipath – Tcpdump – Wireshark
– top statistics – hyptop – Java Health Center – Kernel
– ps – iotop – Dstat – Java Garbage Tracepoints
– vmstat – Lszcrpt – Htop Collection and – Systemtap
– icastats – Netstat Memory visualizer
– Lsqeth – Socket – Jinsight
– Ethtool Statistics
– Preparation – Iptraf

4
Utilization
 Utilization means that a cpu core is used
 Categories qualify what the core was used for

– System, IRQ, SoftIRQ
– Userspace, Guest
– Idle, IOWait, Steal
– Nice
 Accounting unit is Jiffy, reports usually as percentage

– Percentage is better to express relative ratios
 The majority of those basics is usually known, but it still has a purpose
– Clarify and synchronize the understanding between everybody
– Provide metaphors that allow to explain it more easily next time
4
Utilization - Metaphor
 For all the clarifications on basic terms I will use metaphors
– Based on well known real world examples
– At the beginning of a topic the matching metaphor is provided
 Imagine a laptop is a CPU core
 People using that laptop are different Programs, that could be

– Application(s)
– Kernel
– Hypervisor
*Roles are defined by clothes and equipment

not people (bad actors)

4
Utilization - USR
CPU
Application
 If a userspace application is running it is accounted as USR

– This is usually what you want
– Also known as problem state, because there your problems are solved
– If this “application” is actually a virtualized guest it is accounted to Guest
instead

4
Utilization – SYS, (H)IRQ, SIRQ
Application
Kernel
System Call
 For some tasks you need certain privileges
– so you call an administrator (System Call to the Kernel)
– He executes the privileged stuff for you (accounted as SYStem time)

4
Utilization – SYS, (H)IRQ, SIRQ

Other invocations of
the Kernel → IRQ/SIRQ
Application
Kernel
System Call
 For some tasks you need certain privileges
– so you call an administrator (System Call to the Kernel)
– He executes the privileged stuff for you (accounted as SYStem time)
– There are subcategories
• (H)IRQ: privileged work driven by interrupts (instead of the user)
• SIRQ: privileged work driven by soft interrupts and tasklets (instead of the user)
4
Utilization - Idle
Application not
CPU Idle requesting work
This is the most simple case of “doing nothing”

 The CPU executes nothing because it is not requested to do so
 This is accounted as idle

4
Utilization - IOWait
Waiting for I/O
CPU is Idle Application

accounted as iowait waiting for I/O
Again “doing nothing”, but no more that simple

 The System was requested to do some synchronous I/O
– Still the CPU executes nothing because it is not requested to do so
– But it knows it “could” do some work if that I/O would complete
 This is accounted as iowait
4
Utilization – IOWait
Waiting for I/O
CPU is Idle, but No one waiting for I/O

accounted as idle
 If no one is waiting for I/O, (asynchronous)

– Example 1: real Linux AIO
– Example 2: writes via page cache
 Not accounted as iowait, but idle
4
Utilization - Steal
Application thought
it could work
(scheduled)
CPU seems to
be non existent
 The CPU is doing something, just not for you

 CPU doesn't exist for you, but you'd need a CPU to realize that
– Accounted as steal time
– Based on Virtual vs Real timers
 Imagine the laptop is used for multiple groups of people and switched
between their docking stations
– A group of people in front of one laptop causes Context switches (later)

4
Utilization - Steal
If the Application
doesn't work
Nobody cares about

the CPU being
“non existant”
 In case a CPU shouldn't run anyway the stealing isn't even recognized
 Still accounted as idle or iowait

4
Utilization - Steal
Thought it would
transmit network packets
It got from the App
Asked the kernel
for I/O handling
CPU is still busy but
for other things
than Linux thought
 A Linux does not know for which purpose the cpu was stolen
 Steal can indicate issues
– Too high cpu or memory overcommitment
 But steal also isn't always bad
– It could be work you requested, just like “USR->SYS in Linux alone”
4
Utilization – Steal Quiz

 I: Driving I/O synchronously e.g. causing a vswitch to work for your
submission
=> Steal?

4

submission
=> Steal?
Yes

4

 II: Driving sync reads from a file, mdisk does work for you
=> Steal?

4

=> Steal?
No, IOWait

4

 III: Driving async writes (AIO) by a DB, causing mdisk work in the HV
=> Steal?

4

=> Steal?
Kind of, Idle if idle without steal

4

 IV: Paging in the HV takes place while you were running a Java based
BI load utilizing all cores
=> Steal?

4

 IV: Paging in the HV takes place while you were running a Java based
BI load utilizing all cores
=> Steal?
Yes

4

 V: Paging in the HV takes place while your system is an idling Dev
testbed
=> Steal?

4

 V: Paging in the HV takes place while your system is an idling Dev
testbed
=> Steal?
No, Idle

4
Utilization – Steal Quiz summary

submission
=> Steal? - Yes
=> Steal? - No, IOWait
=> Steal? - Kind of, Idle if idle without steal
 IV: Paging in the HV takes place while you were running a Java based BI load
utilizing all cores
=> Steal? - Yes
 V: Paging in the HV takes place while your system is an idling Dev testbed
=> Steal? - No, Idle
4
Scheduling on Cores
 Most systems have more programs than CPUs
 So the OS will have to schedule them (time multiplexing)
 Such scheduling is called Context switching
 So in our metaphor we have multiple people of the same privilege level

using a laptop …

4
Scheduling on Cores
Application
CPU busy working
 Well, that is easy a single program means no context switches

4
Scheduling on Cores
CPU busy, for work and overhead
Application
Application working
working
 With two programs the OS has to switch them every now and then
– Every program shall get its fair share
– Switching causes overhead
• Some CPU time is no more used for the actual work done by the programs (red)

4
Scheduling on Cores
Applications trying
to work
CPU busy, but primarily for overhead
Add steal time if you want real trouble
 With too much runnable programs per CPU the OS gets in trouble
– Actually this happens with any resource being shared a lot
 Eventually there are two bad options to choose from
– Real throughput converging to zero (latency optimized)
– Individual applications have to wait longer (throughput optimized)
4
Scheduling on Cores
cooperative
Applications
Voluntary
context switch
 There are ways to switch cooperatively

– On all blocking system calls like reads, timer sleeps
– On explicit generic or directed context switch
• local (yield / yield_to)
• virtual (diagnose X'44'/ diagnose X'9C')

4
Scheduling on Cores
greedy
Applications
OS has to interrupt
and switch
→ non voluntary
 There are ways to switch cooperatively … or not

– OS can always interrupt and switch
– Actually it looks more like one works and one has to wait, but we fight ...

4
Scheduling between Cores

 Most systems also have more than one CPU
 So the OS will have to schedule/dispatch programs on them
– most “classic” System z Operating Systems use single queue dispatchers
(z/OS, z/VM)
– Linux has a multi queue scheduler (one queue per CPU)
• Tasks are migrated to or pulled from cpus
 So in our metaphor we have

– multiple people and a scheduler
– in our office being a 4 laptop (CPU core)

4
4 Applications Dispatcher
in queue
 Single Queue scheduling

– One scheduler instance directs programs to the CPU cores
– Benefit of single control point and easy synchronization

4

No need for the
scheduler(s) to do anything
Applications
in local runqueue
Applications
in local runqueue
 Multi queue scheduler

– Here with the usual optimum of queues with 1 Program each
– There are cases where it is better to leave one CPU idle
• When two task are bound by the speed of their communication
– You can also see topology here, as some cores are more “remote” than
others
4
Scheduler
migrates task
3 Applications
in local runqueue
 Multi queue scheduler

– Here one queue got rather full and a scheduler starts migrating tasks

4
Scheduling / Utilization – Combo Quiz

 Based on a real customer case
– Database with a lot of stored procedures that does parallelization
– 0-75 runnable processes varying a lot in a “spiky” fashion
– They tried various setups, initially 10, later 4 to 20 real CPU cores
– They didn't achieve their expected target utilization of 85%+ USR
 Question 1: Why was the utilization with 10 CPUs so low at an average

of ~45% despite up to 75 runnable processes?

4

 Question 1: why was the utilization so low at ~45%?
 Eventually one can never get >100%, but easily less
– Could the scheduler do anything about it? → No
– Would it be different with a single queue scheduler? → No
CPU Utilization of spiky applications

10 CPUs - 0-75 Processes - oversimplified for illustarional purposes
120.00%
100.00%
80.00%
Utilization
Utilization
60.00% Mean (Utilization)
40.00%
20.00%
0.00%
1 3 70 65 60 50 20 3 1 0 0 2 4 67 57 45 2 3 2 0 1
# of runnable processes

4

 Question 2: Why did they achieve the following by changing cpu count?
– low # of CPUs: fully utilized, but now with a lot of SYS overhead
– high # of CPUs: even more underutilized but with almost no SYS overhead

10 CPUs - 0-75 Processes - oversimplified for illustarional purposes
120.00%
100.00%
80.00%
Utilization
Utilization
60.00% Mean (Utilization)
40.00%
20.00%
0.00%
1 3 70 65 60 50 20 3 1 0 0 2 4 67 57 45 2 3 2 0 1

4

 Question 2: Why did they achieve the following by changing cpu count?
– low # of CPUs: fully utilized, but now with a lot of SYS overhead
• Context switch overhead
– high # of CPUs: even more underutilized but with almost no SYS overhead
• Even more times of #runnable << #CPUs
10 CPUs - 0.75 Processes - oversimplified for illustrational purposes
120.00%
100.00%
Utilization
80.00%
Mean (Utilization)
Utiization
60.00% SYS
USR
40.00%
20.00%
0.00%
1 3 70 65 60 50 20 3 1 0 0 2 4 67 57 45 2 3 2 0 1

4

– Runnable processes varied a lot in a “spiky” fashion (range 0-75)
– They tried various setups, initially 10, later 4 to 20 CPU cores
 Question 3: Real fix approaches?

4

– Runnable processes varied a lot in a “spiky” fashion (range 0-75)
– They tried various setups, initially 10, later 4 to 20 CPU cores
 Question 3: Real fix approaches?

– Application design (recommended)
– Try to let the Hipervisor make underutilized resources otherwise
usable (needs lower priority workload)

4
Page Cache
 Keeping disk data available in memory is caching
– Certain strategies take place
• What should be cached
• Read ahead of data that will likely be used
• Coalesce writes to issue a single disk write for several memory writes
– Proper management of caches can be complex (read cpu intensive)
 Imagine processes are kids

– There is obviously some privileged person (kernel) watching
– If I learned something they surely will make a mess over time (dirty pages)
– As long as the kids just watch all their toys (read) things stay clean
– But when they really play with things (write) the room gets messy (dirty)
– How are things cleaned up?

4
Page Cache – Cleaning I

 If things stay relatively clean nobody cleans anything
Linux tunable: (% dirty pages < dirty_background_ratio)
– Only long unused items are put back where they belong
Linux tunable: (dirty_expire_centisecs)
All pages
dirty_ratio
dirty_background_ratio
Dirty pages

4
Page Cache – Cleaning II

 As long as the amount of dirtiness is in a sane range it is likely that the
parents will clean a bit in background
Linux tunable: (dirty pages in % > dirty_background_ratio)
– Cleaning of dirty pages consumes CPU, done by the kernel
– Run by the kswap thread(s) and accounted as SYS
– Kernel tries to be nice and stay in background
(as you don't bother the kids too much the kernel tries not to take away cpu)
All pages
dirty_ratio
Dirty pages dirty_background_ratio

4
Page Cache – Cleaning III

 If the amount of dirtiness rises too a really high level the parents will force
the kids to help cleaning up
Linux tunable: (dirty pages in % > dirty_ratio)
– Now writing processes have to contribute parts of their time slices to the kernel
– No more nice, but trying to stall those who make pages dirty
All pages
dirty_ratio
Dirty pages
dirty_background_ratio

4
Page Cache – further details

 Another complex topic is the “proper” size of cache
– Ever realized your flat/house is always too small except for cleaning it
– If you use the kids room as office from 9am-6pm obviously toys have to be
moved aside (cache shrink due to other workload)
– On the other hand if you organize a party it is likely that they will consume more
than just the kids room (cache grow due to more I/O)
 Even cleaning up before going to bed exists in IT
– for actions like hibernate cache has to be cleaned up before going to sleep
Your system, just like cleaning

Parents sometimes feels like
being in a wasteland

4
Paging/Swapping
 Spending more memory than available is overcommitment
 In case the accessed memory exceeds the real memory paging has to
take place
– Paging (z/VM) is the same as swapping (Linux)
 As metaphor imagine a notebook page to be your memory ...

4
Paging/Swapping
Writing
Application
Memory Page
 Initially a page is empty, so a process might write onto it

4
Paging/Swapping
Reading
Application
Memory Page
 Later on the process (or someone else) might read from it

4
Paging/Swapping
Application wants
to write
Memory has no
free page left
 later the process might have something in mind that he wants to write
to the next page
 But all pages the OS could provide are full
– This now requires swapping (OS level)
– Or paging (z/VM level)
4
Paging/Swapping
Kernel takes over

Literally “back breaking”
Swapping can push memory

pages to disk (floor) or
from disk into memory
 The kernel takes over

– Swaps out a page (based on least recently used plus some extras)
– This makes room for a clean new page in real memory
– Impact high, due to the orders of magnitude between disk and memory
 As you see this burden can literally break the back of your system :-)
4
Top
 Characteristics: Easy to use

 Objective: Shows resource usage on process level
 Usage: top b d [interval in sec] > [outfile]
 Package: RHEL: procps SLES: procps
 Shows
– CPU utilization
– Detailed memory usage
 Hints
– Parameter -b enables to write the output for each interval into a file
– Use -p [pid1, pid2,...] to reduce the output to the processes of interest
– Configure displayed columns using 'f' key on the running top program
– Use the 'W' key to write current configuration to ~/.toprc
→ becomes the default

4
top (cont.)
 Output
top 11:12:52 up 1:11, 3 users, load average: 1.21, 1.61, 2.03
Tasks: 53 total, 5 running, 48 sleeping, 0 stopped, 0 zombie
Cpu(s): 3.0%us, 5.9%sy, 0.0%ni, 79.2%id, 9.9%wa, 0.0%hi, 1.0%si, 1.0%st
Mem: 5138052k total, 801100k used, 4336952k free, 447868k buffers
Swap: 88k total, 0k used, 88k free, 271436k cached
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ P SWAP DATA WCHAN COMMAND
3224 root 18 0 1820 604 444 R 2.0 0.0 0:00.56 0 1216 252 dbench
3226 root 18 0 1820 604 444 R 2.0 0.0 0:00.56 0 1216 252 dbench
2737 root 16 0 9512 3228 2540 R 1.0 0.1 0:00.46 0 6284 868 sshd
3225 root 18 0 1820 604 444 R 1.0 0.0 0:00.56 0 1216 252 dbench
3230 root 16 0 2652 1264 980 R 1.0 0.0 0:00.01 0 1388 344 top
1 root 16 0 848 304 256 S 0.0 0.0 0:00.54 0 544 232 select init
2 root RT 0 0 0 0 S 0.0 0.0 0:00.00 0 0 0 migration migration/0
3 root 34 19 0 0 0 S 0.0 0.0 0:00.00 0 0 0 ksoftirqd ksoftirqd/0
4 root 10 5 0 0 0 S 0.0 0.0 0:00.13 0 0 0 worker_th events/0
5 root 20 5 0 0 0 S 0.0 0.0 0:00.00 0 0 0 worker_th khelper

 Hints
– virtual memory: VIRT = SWAP + RES unit KB
– physical memory used: RES = CODE + DATA unit KB
– shared memory SHR unit KB
4
Linux ps command
 Characteristics: very comprehensive, statistics data on process level
 Objective: reports a snapshot of the current processes
 Usage: “ps axlf”
 Shows
– IDs: Pid, Tid, User, …
– Status: stat and wchan
– Details: command, memory consumption and accumulated cpu time
PID TID NLWP POL USER TTY NI PRI PSR P STAT WCHAN START TIME %CPU %MEM VSZ SZ RSS COMMAND
871 871 1 TS root ? 5 29 0 * S< kauditd_thre 10:01 00:00:00 0.0 0.0 0 0 0 [kauditd]
2835 2835 1 TS root pts/2 0 23 0 * Ss+ read_chan 10:38 00:00:00 0.0 0.0 5140 824 2644 bash
3437 3437 1 TS root pts/1 0 23 0 * S+ wait4 11:39 00:00:00 0.0 0.0 1816 248 644 dbench 3
3438 3438 1 TS root pts/1 0 20 0 0 R+ 11:39 00:00:24 33.1 0.0 1820 252 604 dbench 3
…
 Hints
– Do not specify blanks inside the -o format string
– Status is a one time shot, most interactive or I/O bound processes might sleep
4
Linux ps command
 cron that should not appear multiple times
– often long jobs that stack up and make it worse
 Understand what is on there while healthy
 One should be able to name all processes and what they do
#> ps axlf
F UID PID PPID PRI NI VSZ RSS WCHAN STAT TTY TIME COMMAND
1 0 2 0 20 0 0 0 kthrea S ? 0:00 [kthreadd]
1 0 3 2 20 0 0 0 smpboo S ? 0:00 \_ [ksoftirqd/0]
5 0 4 2 20 0 0 0 worker S ? 0:00 \_ [kworker/0:0]
1 0 5 2 0 -20 0 0 worker S< ? 0:00 \_ [kworker/0:0H]
...
4 0 1 0 20 0 9672 5096 SyS_ep Ss ? 0:01 /sbin/init
4 0 2023 1 20 0 3584 1400 hrtime Ss ? 0:00 /usr/sbin/crond -n
4 0 2024 1 20 0 2840 892 pause Ss ? 0:00 /usr/sbin/atd -f
...
4 0 2280 1 20 0 10892 3344 poll_s Ss ? 0:00 /usr/sbin/sshd -D
4 0 2316 2280 20 0 14472 4700 poll_s Ss ? 0:00 \_ sshd: root@pts/0
4 0 2318 2316 20 0 106392 2216 wait Ss pts/0 0:00 \_ -bash
0 0 2350 2318 20 0 105576 1056 - R+ pts/0 0:00 \_ ps axlf
1 0 2351 2318 20 0 106392 796 sleep_ D+ pts/0 0:00 \_ -bash
...
0 0 2368 2318 20 0 2192 360 - R pts/0 110:04 /usr/bin/app
...

4
vmstat
 Characteristics: Easy to use, high-level information
 Objective: First and fast impression of the current state
 Usage: vmstat [interval in sec]
 Package: RHEL: sysstat.s390x SLES: sysstat
 Output sample:
vmstat 1
procs memory swap io system cpu
r b swpd free buff cache si so bi bo in cs us sy id wa st
2 2 0 4415152 64068 554100 0 0 4 63144 350 55 29 64 0 3 4
3 0 0 4417632 64832 551272 0 0 0 988 125 60 32 67 0 0 1
3 0 0 4411804 72188 549592 0 0 0 8984 230 42 32 67 0 0 1
3 0 0 4405232 72896 555592 0 0 0 16 105 52 32 68 0 0 0
 Shows
– Data per time interval
– CPU utilization
– Disk I/O
– Memory usage/Swapping
 Hints
– Shared memory usage is listed under 'cache'
4
Utilization - USR
CPU
Application
 If a userspace application is running it is accounted as USR

– This is usually what you want
– Also known as problem state, because there your problems are solved
– If this “application” is actually a virtualized guest it is accounted to Guest
instead

4
End of Part I
 The one you should always have → IBM System z Enterprise

4
Agenda


4
Non-legal Disclaimer
 This is an introduction and cheat sheet
– Know what is out there
– What could be useful in which case
– How could I debug even further
 These descriptions are not full explanations

– Most tools could get at least 1-2 presentations on their own
– Don't start using them without reading howtos / man pages
 This is not about monitoring

– Some tools used to start performance analysis CAN be monitors, but thats
not part of the presentation

4
General thoughts on performance tools
 Things that are always to consider

– Monitoring can impact the system
– Most data gathering averages over a certain period of time
→ this flattens peaks
– Start with defining the problem
• which parameter(s) from the application/system indicates the problem
• which range is considered as bad, what is considered as good
– monitor the good case and save the results
• comparisons when a problem occurs can save days and weeks
 Work with the tools before an issue occurs

– Learn what should be normal on your system

4
Formulate a theory
 Staged approach saves a lot of work

– Try to use general tools to isolate the area of the issue
– Create theories and try to quickly verify/falsify them
– Use advanced tools to debug the identified area
– Usually a lot of data is generated that needs to be analysed
– Monitoring will impact your system
 Identify the most likely problem area

– CPU related
– Memory related
– I/O related
– Networking related
– or a combination of several factors

4
Orientation - where to go if something is broken

Over-
Tool CPU cons. lat. Hot spots Disk Mem Net
View
top / ps X X n/a n/a n/a n/a n/a
sysstat X X n/a n/a X X n/a
vmstat X X n/a n/a n/a X n/a
iostat X n/a n/a n/a X n/a n/a
dasdstat n/a n/a n/a n/a X n/a n/a
scsistat n/a n/a n/a n/a X n/a n/a
netstat / ss X n/a n/a n/a n/a n/a X
htop / dstat / pidstat X X X n/a X n/a n/a
irqstats X X X n/a n/a n/a n/a
strace / ltrace n/a n/a X n/a n/a n/a n/a
hyptop n/a X n/a n/a n/a n/a n/a
perf n/a X X X X X X
jinsight n/a X X n/a n/a n/a n/a
Health Center X n/a n/a n/a n/a n/a n/a
GMVC n/a n/a X n/a n/a X n/a
blktrace / ziomon / lsof n/a n/a n/a n/a X n/a n/a
Valgrind / smem / slabtop n/a n/a n/a n/a n/a X n/a
iptraf X n/a n/a n/a n/a n/a X
tracepoints n/a n/a X X X X X
iotop n/a n/a n/a n/a X n/a n/a
lszcrypt / icastat n/a X n/a n/a n/a n/a n/a
lsqeth / ethtool / iftop / iptraf
n/a n/a n/a n/a n/a n/a X
/ tcpdump / wireshark
lsluns / multipath n/a n/a n/a n/a X n/a n/a

4
Sysstat - sadc/sar
 Characteristics: Very comprehensive, statistics data on device level
 Objective: Suitable for permanent system monitoring and detailed analysis
 Usage (recommended):
– monitor /usr/lib64/sa/sadc [S XALL] [interval in sec] [outfile]
– View sar A f [outfile]
 Shows
– CPU utilization
– Disk I/O overview and on device level
– Network I/O and errors on device level
– Memory usage/Swapping
– … and much more
– Reports statistics data over time and creates average values for each item
 Hints
– sadc parameter “-S XALL” enables the gathering of further optional data
– Shared memory is listed under 'cache'
– [outfile] is a binary file, which contains all values. It is formatted using sar
• enables the creation of item specific reports, e.g. network only
• enables the specification of a start and end time → time of interest

4
SAR - Processes created
11:47:57 proc/s cswch/s
11:48:05 7.98 53189.03
11:48:09 8.00 54030.25
11:48:13 8.75 53599.25
11:48:17 7.25 54208.75
11:48:21 1.25 54054.75
11:48:25 8.02 54390.48
Context switches per second usually < 1000 per cpu

except during startup or while running a benchmark
if > 10000 your application might have an issue.
Processes created per second usually small except during startup.

If constantly at a high rate your application likely has an issue.
Be aware – the numbers scale with your system size and setup.

4
SAR - CPU utilization

Per CPU values:
watch out for
system time (kernel)
user (applications)
irq/soft (kernel, interrupt handling)
idle (nothing to do)
iowait time (runnable but waiting for I/O)
steal time (runnable but utilized somewhere else)
11:47:57 CPU %usr %sys %iowait %steal %irq %soft %guest %idle
11:48:05 all 70.66 23.47 0.00 0.00 0.12 5.74 0.00 0.00
11:48:05 0 70.32 23.44 0.00 0.00 0.25 5.74 0.00 0.25
11:48:05 1 70.75 23.50 0.00 0.00 0.25 5.50 0.00 0.00
11:48:09 all 71.84 22.53 0.00 0.00 0.00 5.51 0.00 0.13
11:48:09 0 72.00 22.25 0.00 0.25 0.00 5.50 0.00 0.00
11:48:09 1 71.75 22.50 0.00 0.00 0.00 5.50 0.00 0.25
11:48:13 all 71.38 23.12 0.00 0.00 0.00 5.50 0.00 0.00
11:48:13 0 70.75 23.75 0.00 0.00 0.00 5.50 0.00 0.00
11:48:13 1 71.93 22.56 0.00 0.00 0.00 5.51 0.00 0.00

4
SAR - Network traffic
11:47:57 IFACE rxpck/s txpck/s rxkB/s txkB/s rxcmp/s txcmp/s rxmcst/s
11:48:05 lo 52136.16 52136.16 45319.64 45319.64 0.00 0.00 0.00
11:48:05 eth1 0.00 0.00 0.00 0.00 0.00 0.00 0.00
11:48:05 eth0 0.00 0.00 0.00 0.00 0.00 0.00 0.00
11:48:09 lo 53254.25 53254.25 46308.95 46308.95 0.00 0.00 0.00
11:48:09 eth1 0.25 0.00 0.08 0.00 0.00 0.00 0.25
11:48:09 eth0 0.25 0.00 0.08 0.00 0.00 0.00 0.25
Per interface statistic of packets/bytes

You can easily derive average packet sizes from that.
Sometimes people expect - and planned for – different sizes.
Has another panel for errors, drops and such events.

4
SAR – Disk I/O I – overall
11:47:57 tps rtps wtps bread/s bwrtn/s
11:48:01 1023.06 1020.30 2.76 35689.22 92.23
11:48:05 2588.53 2587.03 1.50 89208.98 31.92
11:48:09 491.00 425.50 65.50 15352.00 880.00
11:48:13 278.50 276.00 2.50 9016.00 44.00
11:48:17 135.25 133.00 2.25 4560.00 42.00
11:48:21 698.75 152.00 546.75 5376.00 4566.00
Overview of
- operations per second
- transferred amount

4
SAR – Disk I/O II – per device
11:47:57 DEV tps rd_sec/s wr_sec/s avgrqsz avgqusz await svctm %util
11:48:05 dev2530 165.34 5564.09 0.00 33.65 0.06 0.38 0.21 3.49
11:48:05 dev2531 170.32 5564.09 0.00 32.67 0.06 0.37 0.20 3.49
11:48:05 dev8128 18.95 576.56 0.00 30.42 0.01 0.39 0.39 0.75
11:48:05 dev8144 19.20 674.31 0.00 35.12 0.00 0.26 0.26 0.50
11:48:05 dev8160 19.20 724.19 0.00 37.71 0.01 0.39 0.39 0.75
11:48:05 dev8176 18.95 652.37 0.00 34.42 0.01 0.53 0.53 1.00
Is your I/O balanced across devices?

Imbalances can indicate issues wit a LV setup.
tps and avgrq-sz combined can be important.

Do they match your sizing assumptions?
Await shows the time the application has to wait.

4
SAR - Memory statistics - the false friend
11:47:57 kbmemfree kbmemused %memused kbbuffers kbcached kbcommit %commit
11:48:01 41324 8211980 99.50 16992 7776084 145764 0.94
11:48:05 40844 8212460 99.51 17008 7776388 145964 0.94
11:48:09 40816 8212488 99.51 17036 7776800 145964 0.94
11:48:13 41088 8212216 99.50 17052 7777028 145968 0.94
11:48:17 42592 8210712 99.48 17072 7775396 146100 0.9
Be aware that high %memused and low kbmemfree

is no indication of a memory shortage (common mistake).
Same for swap – to use swap is actually good,

but to access it (swapin/-out) all the time is bad.

4
SAR - Memory pressure - Swap
11:47:57 pswpin/s pswpout/s
11:48:01 2853.95 2658.26
11:48:05 2003.26 5399.80
11:48:09 88.59 9921.92
11:48:13 3199.30 53.15
The percentage seen before can be high,

But the swap rate shown here should be low.
Ideally it is near zero after a rampup time.
High rates can indicate memory shortages.

4
SAR - Memory pressure – faults and reclaim
11:47:57 pgpgin/s pgpgout/s fault/s majflt/s pgfree/s pgscank/s pgscand/s pgsteal/s %vmeff
11:48:01 4799.00 46.12 55571.93 1.00 12821.05 1859.40 0.00 1859.40 100.00
11:48:05 11153.12 15.96 22226.68 0.00 13414.46 2767.08 0.00 2767.08 100.00
11:48:09 1919.00 125.00 825.00 0.00 11345.75 461.75 0.00 461.75 100.00
11:48:13 1127.00 22.00 598.50 0.00 11227.00 262.25 0.00 262.25 100.00
11:48:17 570.00 21.00 52.75 0.00 11184.50 254.25 0.00 254.25 100.00
Don't always trust pgpgin/-out absolute values

Faults populate memory
Major faults need I/O
Scank/s is background reclaim by kswap/flush (modern)
Scand/s is reclaim with a “waiting” allocation
Steal is the amount reclaimed by those scans
vmefficiency is the ratio of scans vs reused (stolen) pages

4
SAR - System Load
11:47:57 runqsz plistsz ldavg1 ldavg5 ldavg15
11:48:01 17 172 0.72 0.35 0.14
11:48:05 15 172 1.95 0.61 0.23
11:48:09 15 172 2.91 0.84 0.30
11:48:13 5 172 3.88 1.07 0.38
11:48:17 15 172 4.21 1.19 0.42
11:48:21 17 171 4.21 1.19 0.42
Runqueue size are the currently runnable programs.

It's not bad to have many, but if they exceed the amount
of CPUs you could do more work in parallel.
Plist-sz is the overall number of programs, if that is always

growing you have likely a process starvation or connection issue.
Load average is a runqueue length average for 1/5/15 minutes.

4
Sysstat - iostat
 Characteristics: Easy to use, information on disk device level

 Objective: Detailed input/output disk statistics
 Usage: iostat -xtdk [interval in sec]
 Shows
– Throughput
– Request merging
– Device queue information
– Service times
 Hints
– Most critical parameter often is await
• average time (in milliseconds) for I/O requests issued to the device to be served.
• includes the time spent by the requests in queue and the time spent servicing them.
– Also suitable for network file systems
4
iostat
 Output sample:
Time: 10:56:35 AM
Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrqsz avgqusz await svctm %util
dasda 0.19 1.45 1.23 0.74 64.43 9.29 74.88 0.01 2.65 0.80 0.16
dasdb 0.02 232.93 0.03 9.83 0.18 975.17 197.84 0.98 99.80 1.34 1.33
Time: 10:56:36 AM
dasda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dasdb 0.00 1981.55 0.00 339.81 0.00 9495.15 55.89 0.91 2.69 1.14 38.83
Time: 10:56:37 AM
dasda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dasdb 0.00 2055.00 0.00 344.00 0.00 9628.00 55.98 1.01 2.88 1.19 41.00
Recent versions are improved by reporting

reads/writes separately which is great as
they have vastly different characteristics

4
Sysstat - PIDSTAT
 Characteristics: Easy to use extended per process statistics

 Objective: Identify processes with peak activity
 Usage: pidstat [w|r|d]
 Package: RHEL: sysstat SLES: sysstat
 Shows
– w context switching activity and if it was voluntary
– r memory statistics, especially minor/major faults per process
– d disk throughput per process
 Hints
– Also useful if run as background log due to its low overhead
• Good extension to sadc in systems running different applications/services
– p <pid> can be useful to track activity of a specific process

4
Pidstat examples
12:46:18 PM PID cswch/s nvcswch/s Command
12:46:18 PM 3 2.39 0.00 smbd
12:46:18 PM 4 0.04 0.00 sshd
12:46:18 PM 1073 123.42 180.18 Xorg
cswch/s means voluntary context switch; nvcswch non-voluntary
12:47:51 PM PID minflt/s majflt/s VSZ RSS %MEM Command
12:47:51 PM 985 0.06 0.00 15328 3948 0.10 smbd
12:47:51 PM 992 0.04 0.00 5592 2152 0.05 sshd
12:47:51 PM 1073 526.41 0.00 1044240 321512 7.89 Xorg
minflt & majflt indicate faults per process
12:49:18 PM PID kB_rd/s kB_wr/s kB_ccwr/s Command
12:49:18 PM 330 0.00 1.15 0.00 sshd
12:49:18 PM 2899 4.35 0.09 0.04 notes2
12:49:18 PM 3045 23.43 0.01 0.00 audacious2
kB_rd & kb_wr represent disk I/O per process

4
Sysstat - mpstat
 Characteristics: Show statistics per processor

 Objective: Identify imbalanced utilization or interrupt peaks
 Usage: mpstat A <interval>
 Package: RHEL: sysstat SLES: sysstat
 Shows
– u utilization
– I <CPU|SCPU|ALL> Interrupts
 Hints
– Can be restricted to selected processor(s) (P)

4
Sysstat – mpstat example
 As one can see there are plenty of different (s)irq sources these days
– Ordered horizontally per type and vertically per cpu
 IRQs
10:40:12 CPU EXT/s I/O/s AIO/s CLK/s EXC/s EMS/s TMR/s TAL/s
PFL/s DSD/s VRT/s SCP/s IUC/s CMS/s CMC/s CMR/s CIO/s QAI/s
DAS/s C15/s C70/s TAP/s VMR/s LCS/s CLW/s CTC/s APB/s ADM/s
CSC/s PCI/s MSI/s VIR/s VAI/s NMI/s RST/s
10:40:17 0 2.40 0.00 0.00 2.00 0.20 0.20 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00
10:40:17 1 1.40 0.00 0.00 1.40 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00
[...]
CPU IRQ type Rate per second
 SoftIRQs
10:40:26 CPU HI/s TIMER/s NET_TX/s NET_RX/s BLOCK/s BLOCK_IOPOLL/s TASKLET/s SCHED/s
HRTIMER/s RCU/s
10:40:31 0 0.00 0.60 0.00 0.00 0.00 0.00 0.00 1.00
0.00 0.60
10:40:31 1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.00 0.00
[...]
4
DASD statistics
 Characteristics: Easy to use, very detailed
 Objective: Collects statistics of I/O operations on DASD devices
 Usage:
– enable: echo on > /proc/dasd/statistics
– show:
• Overall cat /proc/dasd/statistics
• for individual DASDs tunedasd P /dev/dasda
 Package: n/a for kernel interface, s390-tools for dasdstat
 Shows:
Histogram of I/O between Histogram between
Histogram of I/O till ssch ssch and IRQ I/O and End
Start Build channel program Processing data transfer Tell block dev layer
End
wait till subchannel is from/to storage server Data has arrived
free
Histogram of I/O times
 Hints
– The tools dasdstat is available in s390-utils to handle all that in one place

4
DASD statistics – report

 Sample:
8*512b = 4KB <= request size < 16*512b =8KB 1ms <= response time < 2 ms
29432 dasd I/O requests
with 6227424 sectors(512B each)
__<4 ___8 __16 __32 __64 _128 _256 _512 __1k __2k __4k __8k _16k _32k _64k 128k
_256 _512 __1M __2M __4M __8M _16M _32M _64M 128M 256M 512M __1G __2G __4G _>4G
Histogram of sizes (512B secs)
0 0 9925 3605 1866 4050 4102 933 2700 2251 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Histogram of I/O times (microseconds)
0 0 0 0 0 0 0 1283 1249 6351 7496 3658 8583 805 7 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Histogram of I/O time till ssch ← look here for subchannel busy
2314 283 98 34 13 5 16 275 497 8917 5567 4232 7117 60 4 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Histogram of I/O time between ssch and irq ← look here for slow SAN
0 0 0 0 0 0 0 14018 7189 2402 1031 4758 27 4 3 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Histogram of I/O time between irq and end
2733 6 5702 9376 5781 940 1113 3781 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
# of req in chanq at enqueuing (1..32)
0 2740 628 1711 1328 23024 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 Hints
– Also shows data per sector which usually is only confusing

4
DASD statistics – look for subchannel busy issues

__<4 ___8 __16 __32 __64 _128 _256 _512 __1k __2k __4k __8k _16k _32k _64k 128k
_256 _512 __1M __2M __4M __8M _16M _32M _64M 128M 256M 512M __1G __2G __4G _>4G
[...]
Histogram of I/O time till ssch
2314 283 98 34 13 5 16 275 497 8917 5567 4232 7117 60 4 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 Time consists of
– “Build subchannel program (usually very fast)
– wait for a free subchannel (can be long without or too few HPAV)
– Please be aware that all columns in dasdstat are scaled by 2n
Number of
I/O time till ssch
requests before
10000
the current one
9000
8000
7000
2+
6000
# of requests
5000
4000
3000 0
2000 1-2
1000
0
time in µs (in buckets)

4
FCP statistics
 Characteristics: Detailed latency information for FCP I/O

 Objective: Collect details of I/O operations on FCP devices
 Package: n/a (Kernel interface)
 Usage:
– enable
• CONFIG_STATISTICS=y must be set in the kernel config file
• debugfs is mounted at /sys/kernel/debug/
• For a certain LUN in directory
/sys/kernel/debug/statistics/zfcp<devicebusid><WWPN><LUN>
issue echo on=1 > definition (turn off with on=0, reset with data=reset)
– view
• cat /sys/kernel/debug/statistics/zfcp<devicebusid><WWPN>
<LUN>/data
 Hint
– FCP and DASD statistics are not directly comparable, because in the FCP case many
I/O requests can be sent to the same LUN before the first response is given. There is
a queue at FCP driver entry and in the storage server
4
FCP statistics
 Shows:
– Request sizes in bytes (hexadecimal)
– Channel latency in ns Time spent on the FCP channel (internal transfer)
– Fabric latency in ns Time spent in the FCP fabric (outside transfer)
– (Overall) latencies whole time spent entry/exit of the zFCP layer in ms
SCSI stack zfcp I/O passes through I/O queued Completion IRQ handling Completion
Fabric
preps I/O preps I/O QDIO and z/VM for transm. handling by zfcp handled by SCSI
SCSI stack issues I/O
Resp. received
zfcp issues I/O
= blktrace dispatch
Receive IRQ
I/O received
I/O issued
by card
by card
by card
Received by
completed by
SCSI stack
SCSI stack
Channel Fabric Channel
Seen in Overall Latency Seen in

blktrace blktrace
4
FCP statistics
 On popular request – the “where to complain” guide
– █ App to zfcp driver: Linux developers / Distributor in general
– █ zfcp driver to hardware: zfcp/qdio (optional also z/VM) developers
– █ In system hardware to SAN: FCP card HW/FW stack
– █ External: SAN administrator
SCSI stack zfcp I/O passes through I/O queued Completion IRQ handling Completion
Fabric
preps I/O preps I/O QDIO and z/VM for transm. handling by zfcp handled by SCSI
SCSI stack issues I/O
Resp. received
zfcp issues I/O
= blktrace dispatch
Receive IRQ
I/O received
I/O issued
by card
by card
by card
Received by
completed by
SCSI stack
SCSI stack
Channel Fabric Channel
Seen in Overall Latency Seen in

blktrace blktrace
4
FCP statistics example – rather unreadable

cat /sys/kernel/debug/statistics/zfcp-0.0.1700-0x5005076303010482-0x4014400500000000/data
...
request_sizes_scsi_read 0x1000 1163
means: request size 4KB, 1163 occurrences
request_sizes_scsi_read 0x2d000 44
request_sizes_scsi_read 0x2a000 26
request_sizes_scsi_read 0x1e000 25
...
latencies_scsi_read <=1 1076
means: response time <= 1ms
...
channel_latency_read <=16000 0
means: Channel response time <= 32μs
= all below driver
...
fabric_latency_read <=1000000 1238 means: Fabric response time <= 1ms
fabric_latency_read <=2000000 328
fabric_latency_read <=4000000 522 = once leaving the card
...

4
FCP statistics example

 Some statistics are per device
 Some per adapter
– Adapter statistics can not be reseted
 Example beautifying the same data with a little script (not public yet)
per device latency statistics (f fabric; c channel)
disk rdfmin rdfmax rdfsum rdfavg rdcmin rdcmax rdcsum rdcavg rdcnt | wr* [..wr/cmd]
sda 92 130 1386 126.00 7 10 98 8.91 11
sdb 127 131 2072 129.50 7 10 140 8.75 16
sdc 126 140 2075 129.69 7 14 145 9.06 16
sdd 126 132 1160 128.89 7 8 74 8.22 9
[...]
sdbd n/a n/a 0 0.00 n/a n/a 0 0.00 0
sdbe n/a n/a 0 0.00 n/a n/a 0 0.00 0
per adapter statistics
adapter (subch/dev) rdcnt rdmb rdavgsz wrcnt wrmb wravgsz cmdcnt secactive
0.0.0004/0.0.1700 4899 18 3.76 11572 1249 110.52 240 17324
0.0.000c/0.0.1800 4901 16 3.34 11564 1265 112.02 236 17325
0.0.00d6/0.0.5100 4765 16 3.44 11595 1254 110.75 239 17318
0.0.00e2/0.0.5b00 1888 5 2.71 0 0 0.00 160 17309

4
iotop
 Characteristics: simple, top like I/O monitor
 Objective: Check which processes are doing I/O
 Usage: iotop
 Package: RHEL: iotop SLES: iotop
 Shows
– Read/Write per thread
– Can accumulate (-a) for updating summaries instead of live views
• Useful for Disk I/O tests that don't account on their own
– Separate accounting for swap
 Hints
– Can be restricted to certain processes via (p)
– Has a batch mode like top

4
Iotop - examples System wide values:

total: app<->kernel
actual: kernel<->disk
 Example I: disk I/O can be spread other than expected
Total DISK READ : 34.05 M/s | Total DISK WRITE : 45.20 B/s
Actual DISK READ: 24.02 B/s | Actual DISK WRITE: 31.19 M/s
TID PRIO USER DISK READ DISK WRITE SWAPIN IO> COMMAND
7204 be/4 qemu 378.16 K/s 0.00 B/s 0.00 % 6.16 % qemusystems390x
7225 be/4 qemu 343.08 K/s 19.49 K/s 0.00 % 6.00 % qemusystems390x
Disk Read / Write

values are per process
 Example II: even interesting for memory bound (overcommitted) loads

Total DISK READ : 971.07 M/s | Total DISK WRITE : 180.50 M/s
Actual DISK READ: 850.18 M/s | Actual DISK WRITE: 150.23 M/s
TID PRIO USER DISK READ DISK WRITE SWAPIN IO> COMMAND
7225 be/4 qemu 6496.12 K/s 0.00 K/s 28.59 % 2.86 % qemusystems390x
Swapin % of time IO % means time

while swapping in waiting for I/O
4
Lszcrypt / icastats
 Characteristics: overview of s390 crypto HW and libica usage
 Objective: am I really using my crypto hardware
 Usage: “lszcrypt VV[V]” “cat /proc/icastats”
 Package: RHEL: s390utils-base SLES: s390-tools
lszcrypt VV Cat/proc/icastats
card02: CEX3C online hwtype=9 depth=8 function | # hardware | # software
request_count=443 ++
card03: CEX3A offline hwtype=8 depth=8 SHA1 | 0 | 0
request_count=0 SHA224 | 0 | 0
SHA256 | 0 | 0
SHA384 | 0 | 0
SHA512 | 0 | 0
RANDOM | 187109 | 0
MOD EXPO | 0 | 0
RSA CRT | 93554 | 0
DES ENC | 0 | 0
DES DEC | 0 | 0
3DES ENC | 0 | 0
3DES DEC | 0 | 0
AES ENC | 2574106 | 0
AES DEC | 2075854 | 0
CMAC GEN | 0 | 0
CMAC VER | 0 | 0
 Never assume your HW correctly is used until you confirmed it

– If not going via libica (e.g. Java pkcs#11 you won't see it in icastat)

4
lsqeth
 Characteristics: overview of network devices
 Objective: check your network devices basic setup
 Usage: “lsqeth p”
 Package: RHEL: s390-utils-base SLES: s390-tools
lsqeth p
devices CHPID interface cardtype port chksum prioq'ing rtr4 rtr6 lay'2 cnt

0.0.e000/0.0.e001/0.0.e002 x84 eth1 OSD_10GIG 0 sw always_q_2 n/a n/a 1 64
0.0.e100/0.0.e101/0.0.e102 x85 eth2 OSD_10GIG 0 sw always_q_2 n/a n/a 1 64
0.0.f200/0.0.f201/0.0.f202 x6B eth0 OSD_1000 0 hw always_q_2 no no 0 64
 Check for layer, offload, and buffer counts

– More buffers are usually better especially for massive amounts of concurrent
connections

4
Ethtool I
 Characteristics: overview of network device capabilities / offload settings
 Objective: check your network device (offload) settings
 Usage: “ethtool <dev>”, “ethtool k <dev>”
 Package: RHEL: ethtool SLES: ethtool
ethtool eth1
Settings for eth1:
Supported ports: [ FIBRE ]
Supported link modes: 10baseT/Half 10baseT/Full
100baseT/Half 100baseT/Full
10000baseT/Full
Supported pause frame use: No
Supports autonegotiation: Yes
Advertised link modes: 10baseT/Half 10baseT/Full
10000baseT/Full
Advertised pause frame use: No
Advertised autonegotiation: Yes
Speed: 10000Mb/s
Duplex: Full
Port: FIBRE
PHYAD: 0
Transceiver: internal
Autonegotiation: on
Link detected: yes
 Check e.g. announced speeds

4
Ethtool II
 Offload Settings via “ethtool k <dev>”
 Changes via upper case “-K”
ethtool k eth1 […]
Features for eth1: ntuplefilters: off [fixed]
rxchecksumming: off [fixed] receivehashing: off [fixed]
txchecksumming: off highdma: off [fixed]
txchecksumipv4: off [fixed] rxvlanfilter: on [fixed]
txchecksumipgeneric: off [fixed] vlanchallenged: off [fixed]
txchecksumipv6: off [fixed] txlockless: off [fixed]
txchecksumfcoecrc: off [fixed] netnslocal: off [fixed]
txchecksumsctp: off [fixed] txgsorobust: off [fixed]
scattergather: off txfcoesegmentation: off [fixed]
txscattergather: off [fixed] txgresegmentation: off [fixed]
txscattergatherfraglist: off [fixed] txudp_tnlsegmentation: off [fixed]
tcpsegmentationoffload: off fcoemtu: off [fixed]
txtcpsegmentation: off [fixed] txnocachecopy: off
txtcpecnsegmentation: off [fixed] loopback: off [fixed]
txtcp6segmentation: off [fixed] rxfcs: off [fixed]
udpfragmentationoffload: off [fixed] rxall: off [fixed]
genericsegmentationoffload: off [requested on] txvlanstaghwinsert: off [fixed]
genericreceiveoffload: on rxvlanstaghwparse: off [fixed]
largereceiveoffload: off [fixed] rxvlanstagfilter: off [fixed]
rxvlanoffload: off [fixed]
txvlanoffload: off [fixed]
[...]
 In some cases external influences like OSA-layer2 prevent most offloads

(the example here)
4
Don't miss preparation
 Of all tools preparation is clearly

–The most important
–The most effective
 Prepare
– System and Workload descriptions Prepare Gather
– Healthy system data for comparison
 Gather lessons learned
problem occurs
need assistance
– In case of emergency analysis
 Report
– How to report a Problem Description Solve Report
 Solve
– Tools to start an analysis more info

4
End of Part II

4
Agenda


4
STRACE
 Characteristics: High overhead, high detail tool

 Objective: Get insights about the ongoing system calls of a program
 Usage: strace p [pid of target program]
 Package: RHEL: strace SLES: strace
 Shows
– Identify kernel entries called more often or taking too long
• Can be useful if you search for increased system time
– Time in call (T)
– Relative timestamp (r)
 Hints
– The option “c” allows medium overhead by just tracking counters and
durations

4
strace - example
The columns “seconds”, “usecs/call”,
“calls”, “errors” can reveal if one has
% time reports shares, useful a lot, slow or failing calls? syscall references are
to rate the importance of a call available as “man pages”
strace cf p 26802
Process 26802 attached interrupt to quit
^Process 26802 detached
% time seconds usecs/call calls errors syscall

58.43 0.007430 17 450 read
24.33 0.003094 4 850 210 access
5.53 0.000703 4 190 10 open
4.16 0.000529 3 175 write
2.97 0.000377 2 180 munmap
1.95 0.000248 1 180 close
1.01 0.000128 1 180 mmap
0.69 0.000088 18 5 fdatasync
0.61 0.000078 0 180 fstat
0.13 0.000017 3 5 pause

100.00 0.012715 2415 225 total

4
LTRACE
 Characteristics: High overhead, high detail tool

 Objective: Get insights about the ongoing library calls of a program
 Usage: ltrace p [pid of target program]
 Package: RHEL: ltrace SLES: ltrace
 Shows
– Identify library calls that are too often or take too long
• Good if you search for additional user time
• Good if things changed after upgrading libs
– Time in call (T)
– Relative timestamp (r)
 Hints
– The option “c” allows medium overhead by just tracking counters and durations
– The option “S” allows to combine ltrace and strace

4
ltrace - example
The columns “seconds”, “usecs/call”,
“calls”, “errors” can reveal if one has
% time reports shares, useful a lot, slow or failing calls? Most function references are
to rate the importance of a call available as “man pages”
ltrace cf p 26802
% time seconds usecs/call calls function

98.33 46.765660 5845707 8 pause
0.94 0.445621 10 42669 strncmp
0.44 0.209839 25 8253 fgets
0.08 0.037737 11 3168 __isoc99_sscanf
0.07 0.031786 20 1530 access
0.04 0.016757 10 1611 strchr
0.03 0.016479 10 1530 snprintf
0.02 0.010467 1163 9 fdatasync
0.02 0.008899 27 324 fclose
0.02 0.007218 21 342 fopen
0.01 0.006239 19 315 write
0.00 0.000565 10 54 strncpy

100.00 47.560161 59948 total

4
Strace / Ltrace – full trace

 Without -c both tools produce a full detail log
– Via -f child processes can be traced as well
– Extra options “Tr” are useful to search for latencies
follow time in call / relative timestamp
– Useful to “read” what exactly goes on when
Example strace'ing a sadc data gatherer
0.000028 write(3, "\0\0\0\0\0\0\0\17\0\0\0\0\0\0\0"..., 680) = 680 <0.000007>
0.000027 write(3, "\0\0\0\0\0\0\0\17\0\0\0\0\0\0\0"..., 680) = 680 <0.000007>
0.000026 fdatasync(3) = 0 <0.002673>
0.002688 pause() = 0 <3.972935>
3.972957 SIGALRM (Alarm clock) @ 0 (0)
0.000051 rt_sigaction(SIGALRM, {0x8000314c, [ALRM], SA_RESTART}, 8) = 0 <0.000005>
0.000038 alarm(4) = 0 <0.000005>
0.000031 sigreturn() = ? (mask now []) <0.000005>
0.000024 stat("/etc/localtime", {st_mode=S_IFREG|0644, st_size=2309, ...}) = 0 <0.000007>
0.000034 open("/proc/uptime", O_RDONLY) = 4 <0.000009>
0.000024 fstat(4, {st_mode=S_IFREG|0444, st_size=0, ...}) = 0 <0.000005>
0.000029 mmap(NULL, 4096, PROT_READ, MAP_PRIVATE|MAP_ANONYMOUS, 1, 0) = 0x3fffd20a000 <0.000006>
0.000028 read(4, "11687.70 24836.04\n", 1024) = 18 <0.000010>
0.000027 close(4) = 0 <0.000006>
0.000020 munmap(0x3fffd20a000, 4096) = 0 <0.000009>

4
lsof
 Characteristics: list of open files plus extra details

 Objective: which process accesses which file in which mode
 Usage: lsof +fg
 Package: RHEL: lsof SLES: lsof
 Shows
– List of files including sockets, directories, pipes
– User, Command, Pid, Size, Device
– File Type and File Flags
 Hints
– +fg reports file flags which can provide a good cross check opportunity

4
lsof - example
COMMAND PID TID USER FD TYPE FILEFLAG DEVICE SIZE/OFF NODE NAME
crond 16129 root mem REG 94,1 165000 881893
/usr/lib64/ld2.16.so
crond 16129 root 0r CHR LG 1,3 0t0 2051 /dev/null
crond 16129 root 1u unix RW 0x0000001f1ba02000 0t0 106645 socket
crond 16129 root 2u unix RW 0x0000001f1ba02000 0t0 106645 socket
crond 16129 root 4r a_inode 0x80000 0,9 0 6675 inotify
crond 16129 root 5u unix RW,0x80000 0x0000001f5d3ad000 0t0 68545 socket
dd 17617 root cwd DIR 94,1 4096 16321 /root
dd 17617 root rtd DIR 94,1 4096 2 /
dd 17617 root txt REG 94,1 70568 1053994 /usr/bin/dd
dd 17617 root mem REG 94,1 165000 881893
/usr/lib64/ld2.16.so
dd 17617 root 0r CHR LG 1,9 0t0 2055 /dev/urandom
dd 17617 root 1w REG W,DIR,LG 94,1 5103616 16423 /root/test
dd 17617 root 2u CHR RW,LG 136,2 0t0 5 /dev/pts/2
 You can filter that per application or per file

– Fd holds fdnumber, type, characteristic and lock information
• File descriptors can help to read strace/ltrace output
– Flags can be good to confirm e.g. direct IO, async IO
– Size (e.g. mem) or offset (fds), name, ...

4
lsluns
 Characteristics: overview of multipathing
 Objective: check your multipath setup hierarchy
 Usage: “lsluns a”
lsluns a
adapter = 0.0.1700
port = 0x500507630900c7c1
lun = 0x4020402100000000 /dev/sg0 Disk IBM:2107900
adapter = 0.0.1780
port = 0x500507630903c7c1
[...]
 Lsluns provides a hierarchical view which often easily identifies missing paths,
adapters or similar imbalances
 Adapter to WWPN associations can have concurring targets
– Low overhead, max fallback capability, best performance, ...
4
Multipath -ll
 Characteristics: overview of multipathing
 Objective: check your multipath setup configuration
 Usage: “mutlipath ll”
 Package: RHEL: device-mapper-multipath SLES: mutlipath-tools
multipath ll
swap3of6 (36005076309ffc7c10000000000002022) dm2 IBM ,2107900
size=256G features='0' hwhandler='0' wp=rw
`+ policy='servicetime 0' prio=0 status=active
| 0:0:20:1075986464 sdb 8:16 active ready running
| 1:0:22:1075986464 sdx 65:112 active ready running
| 2:0:21:1075986464 sdh 8:112 active ready running
| 3:0:20:1075986464 sdn 8:208 active ready running
| 4:0:26:1075986464 sdz 65:144 active ready running
| 5:0:19:1075986464 sdy 65:128 active ready running
| 7:0:25:1075986464 sdac 65:192 active ready running
` 6:0:24:1075986464 sdad 65:208 active ready running
[...]
 This also reports multipath.conf inconsitencies

 Check all reported parameters are what you thought them to be
– For example (in)famous rr_min_io renaming
4
Hyptop
 Characteristics: Easy to use Guest/LPAR overview

 Objective: Check CPU and overhead statistics of your and sibling images
 Usage: hyptop
 Shows
– CPU load & Management overhead
– Memory usage (only under zVM)
– Can show image overview or single image details
 Hints
– Good “first view” tool for linux admins that want to look “out of their linux”
– Requirements:
• For z/VM the Guest needs Class B
• For LPAR “Global performance data control” checkbox in HMC

4
Hyptop example memuse = resident
Why are exactly 4 CPUs used

in all 6 CPU guests
All these do not fully

utilize their 2 CPUs
service guest weights
No peaks in service guests
LPAR images would see other

LPARs

4
DSTAT
 Characteristics: Live easy to use full system information

 Objective: Flexible set of statistics
 Usage: dstat -tv –aio –disk-util -n –net-packets -i –ipc
 -D total,[diskname] –top-io [...] [interval]
 Short: dstat vtin
 Package: RHEL: dstat SLES: n/a WWW: http://dag.wieers.com/home-made/dstat/
 Shows
– Throughput
– Utilization
– Summarized and per Device queue information
– Much more … it more or less combines several classic tools like iostat and vmstat
 Hints
– Powerful plug-in concept
• “topio” for example identifies the application causing the most I/Os
– Colorization allows fast identification of deviations

4
Dstat – the limit is your screen width

system procs memoryusage paging dsk/total
time |run blk new| used buff cach free| in out | read writ
0402 10:52:50| 0 0 1.1| 291M 81.1M 160M 9382M| 0 0 | 93k 8591B
0402 10:53:20|0.1 0 0.3|9771M 400k 60.3M 83.5M| 637M 565M| 303k 29k
0402 10:53:50| 0 0 0|9785M 376k 60.2M 69.5M|1162M 166M| 229k 2458B
0402 10:54:20| 0 0 0|9786M 528k 60.7M 68.0M| 48k 0 | 50k 1092B
0402 10:54:50|0.1 0 0|9790M 428k 59.7M 64.5M|1102M 895k| 20k 1911B
new in live tool similar to iostat

similar to vmstat
(also per device)
totalcpuusage inter system virtualmemory
usr sys idl wai hiq siq| 1 | int csw |majpf minpf alloc free
many
0 0 100 0 0 0| 9 | 145 279 | 0 307 195 689 many
1 2 97 0 0 0| 4 | 12k 1119 | 22k 150k 308k 229k more ...
1 3 96 0 0 0| 3 | 892 960 | 69k 254k 304k 304k
0 0 100 0 0 0| 2 | 309 619 | 2 8 20 1
1 3 96 0 0 0| 1 | 594 875 | 79k 258k 282k 282k
4
HTOP
 Characteristics: Process overview with extra features

 Objective: Get a understanding about your running processes
 Usage: htop
 Package: RHEL: n/a SLES: n/a WWW: http://htop.sourceforge.net/
 Shows
– Running processes
– CPU and memory utilization
– Accumulated times
– I/O rates
– System utilization visualization
 Hints
– Htop can display more uncommon fields (in menu)
– Able to send signals out of its UI for administration purposes
– Processes can be sorted/filtered for a more condensed view

4
htop example
Configurable utilization visualization
Hierarchy
Common process info
Accumulated Usage
and IO rates
4
netstat
 Characteristics: Easy to use, connection information
 Objective: Lists connections
 Usage: netstat eeapn
 Package: RHEL: net-tools SLES: net-tools
 Shows
– Information about each connection
– Various connection states
 Hints
– Inodes and program names are useful to reverse-map ports to applications

4
netstat -s
 Characteristics: Easy to use, very detailed information
 Objective: Display summary statistics for each protocol
 Usage: netstat s
 Shows
– Information to each protocol
– Amount of incoming and outgoing packages
– Various error states, for example TCP segments retransmitted!
 Hints
– Shows accumulated values since system start, therefore mostly the differences
between two snapshots are needed
– There is always a low amount of packets in error or resets
– Retransmits occurring only when the system is sending data
When the system is not able to receive, then the sender shows retransmits
– Use sadc/sar to identify the device

4
netstat -s
 Output sample:
Tcp:
15813 active connections openings
35547 passive connection openings
305 failed connection attempts
0 connection resets received
6117 connections established
81606342 segments received
127803327 segments send out
288729 segments retransmitted
0 bad segments received.
6 resets sent

4
Socket statistics
 Characteristics: Information on socket level
 Objective: Check socket options and weird connection states
 Usage: ss aempi
 Package: RHEL: iproute-2 SLES: iproute2
 Shows
– Socket options
– Socket receive and send queues
– Inode, socket identifiers
 Sample output
ss aempi
State RecvQ SendQ Local Address:Port Peer Address:Port
LISTEN 0 128 :::ssh :::*
users:(("sshd",959,4)) ino:7851 sk:ef858000 mem:(r0,w0,f0,t0)
 Hints
– Inode numbers can assist reading strace logs
– Check long outstanding queue elements
4
IPTRAF
 Characteristics: Live information on network devices / connections

 Objective: Filter and format network statistics
 Usage: iptraf
 Package: RHEL: iptraf / iptraf-ng SLES: iptraf
 Shows
– Details per Connection / Interface
– Statistical breakdown of ports / packet sizes
– LAN station monitor
 Hints
– Can be used for background logging as well
• Use SIGUSR1 and logrotate to handle the growing amount of data
– Knowledge of packet sizes important for the right tuning
– There are various other tools: iftop, bmon, …
• like with net benchmarks no one seem to fit all
4
iptraf
 Questions that usually can be addressed
– Connection behavior overview
– Do you have peaks in your workload characteristic
– Who does your host really communicate with
 Comparison to wireshark
– Not as powerful, but much easier and faster to use
– Lower overhead and no sniffing needed (often prohibited)
Packet IF
sizes details

4
End of Part III


4
Agenda


4
Perf
 Characteristics: Easy to use profiling and kernel tracing

 Objective: Get detailed information where & why CPU is consumed
 Usage: perf (to begin with)
 Package: RHEL: perf SLES: perf
 Shows
– Sampling for CPU hotspots
• Annotated source code along hotspots
• list functions according to their usage
– CPU event counters
– Further integrated non-sampling tools
 Hints
– Without HW support only userspace can be reasonably profiled
– “successor” of oprofile that is available with HW support (SLES11-SP2)
– Perf HW support upstream but still experimental

4
Perf
 What profiling can and what it can't

+ Search hotspots of CPU consumption worth to optimize
+ List functions according to their usage
- Search where time is lost (I/O, Latency caused stalls)
 Perf is not just a sampling tool

– Integrated tools to evaluate tracepoints like
“perf sched”, “perf timechart”, …
• Other than real “sampling” this can help to search for stalls
– Counters provide even lower overhead and report HW and Software events

4
Perf profiling
 Perf example how-to

– Needs proper HW support to work well for the kernel (not yet in the field)
• Ignore and kernel profiling data until this is available!
– We had a case where new code caused cpus to scale badly
– perf record “workload”
• Creates a file called perf.data that can be analyzes
– We used “perf diff” on both data files to get a comparison
 “Myriad” of further options/modules

– Live view with perf top
– Perf sched for an integrated analysis of scheduler tracepoints
– Perf annotate to see samples alongside code
– Perf stat for a counter based analysis
– [...]

4
Perf profiling
 Perf example (perf diff)

– found a locking issue causing increased cpu consumption
# Baseline Delta Symbol
# ........ .......... ................... ......
#
12.14% +8.07% [kernel.kallsyms] [k] lock_acquire
8.96% +5.50% [kernel.kallsyms] [k] lock_release
4.83% +0.38% reaim [.] add_long
4.22% +0.41% reaim [.] add_int
4.10% +2.49% [kernel.kallsyms] [k] lock_acquired
3.17% +0.38% libc2.11.3.so [.] msort_with_tmp
3.56% 0.37% reaim [.] string_rtns_1
3.04% 0.38% libc2.11.3.so [.] strncat

4
Perf stat - preparation
 Activate the cpu measurement facility

– If not you'll encounter this
Error: You may not have permission to collect stats.
Consider tweaking /proc/sys/kernel/perf_event_paranoid
Fatal: Not all events could be opened.
– Check if its activated

• separate for counter and sampling
• Basic and/or Diagnostic mode
lscpumf i
CPUmeasurement counter facility
[...]
Sampling facility information for cpum_sf
[...]
Authorized sampling modes:
basic (sample size: 32 bytes)
diagnostic (sample size: 85 bytes)
[...]

4
Perf stat - usage
perf stat B event=cycles,instructions,r20,r21,r3,r5,sched:sched_wakeup find / iname
"*foobar*"
Performance counter stats for 'find / iname *foobar*':
3,623,031,935 cycles # 0.000 GHz
1,515,404,340 instructions # 0.42 insns per cycle
1,446,545,776 PROBLEM_STATE_CPU_CYCLES
757,589,098 PROBLEM_STATE_INSTRUCTIONS
705,740,759 L1I_PENALTY_CYCLES
576,226,424 L1D_PENALTY_CYCLES
40,675 sched:sched_wakeup
6.156288957 seconds time elapsed
 Still experimental
 Events
– Cycles/Instructions globally
– Note: counters are now readable, but aliases can still be used
• e.g. r20 = PROBLEM_STATE_CPU_CYCLES
• list of all existing events lscpumf C
• counters available to you lscpumf c
– Not only HW events, you can use any of the currently 163 tracepoints
4
Slabtop
 Characteristics: live profiling of kernel memory pools

 Objective: Analyze kernel memory consumption
 Usage: slabtop
 Shows
– Active / Total object number/size
– Objects per Slab
– Object Name and Size
– Objects per Slab
 Hints
– -o is one time output e.g. to gather debug data
– Despite slab/slob/slub in kernel its always slabtop

4
Slabtop - example
Active / Total Objects (% used) : 2436408 / 2522983 (96.6%)
Active / Total Slabs (% used) : 57999 / 57999 (100.0%)
Active / Total Caches (% used) : 75 / 93 (80.6%)
Active / Total Size (% used) : 793128.19K / 806103.80K (98.4%)
Minimum / Average / Maximum Object : 0.01K / 0.32K / 8.00K
OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME
578172 578172 100% 0.19K 13766 42 110128K dentry
458316 458316 100% 0.11K 12731 36 50924K sysfs_dir_cache
368784 368784 100% 0.61K 7092 52 226944K proc_inode_cache
113685 113685 100% 0.10K 2915 39 11660K buffer_head
113448 113448 100% 0.55K 1956 58 62592K inode_cache
111872 44251 39% 0.06K 1748 64 6992K kmalloc64
54688 50382 92% 0.25K 1709 32 13672K kmalloc256
40272 40239 99% 4.00K 5034 8 161088K kmalloc4096
39882 39882 100% 0.04K 391 102 1564K ksm_stable_node
38505 36966 96% 0.62K 755 51 24160K shmem_inode_cache
37674 37674 100% 0.41K 966 39 15456K dm_rq_target_io
 How is kernel memory managed by the sl[auo]b allocator used

– Named memory pools or Generic kmalloc pools
– Active/total objects and their size
– growth/shrinks of caches due to workload adaption

4
BLKTRACE
 Characteristics: High detail info of the block device layer actions

 Objective: Understand whats going with your I/O in the kernel and devices
 Usage: blktrace -d [device(s)]
 Then: blkparse -st [commontracefilepart]
 Package: RHEL: blktrace SLES: blktrace
 Shows
– Events like merging, request creation, I/O submission, I/O completion, ...
– Timestamps and disk offsets for each event
– Associated task and executing CPU
– Application and CPU summaries
 Hints
– Filter masks allow lower overhead if only specific events are of interest
– Has an integrated client/server mode to stream data away
• Avoids extra disk I/O on a system with disk I/O issues

4
Blktrace – when is it useful
 Often its easy to identify that I/O is slow, but

→ Where?
→ Because of what?
 Blocktrace allows to
– Analyze Disk I/O characteristics like sizes and offsets
• Maybe your I/O is split in a layer below
– Analyze the timing with details about all involved Linux layers
• Often useful to decide if HW or SW causes stalls
– Summaries per CPU / application can identify imbalances

4
Blktrace - events
Common:
A -- remap For stacked devices, incoming i/o is remapped to device below it in the i/o stack. The remap action details what exactly is being remapped to
what.
Q -- queued This notes intent to queue i/o at the given location. No real requests exists yet.
G -- get request To send any type of request to a block device, a struct request container must be allocated first.
I -- inserted A request is being sent to the i/o scheduler for addition to the internal queue and later service by the driver. The request is fully formed at this
time.
D -- issued A request that previously resided on the block layer queue or in the i/o scheduler has been sent to the driver.
C -- complete A previously issued request has been completed. The output will detail the sector and size of that request, as well as the success or failure
of it.
Plugging & Merges:

P -- plug When i/o is queued to a previously empty block device queue, Linux will plug the queue in anticipation of future I/Os being added before this
data is needed.
U -- unplug Some request data already queued in the device, start sending requests to the driver. This may happen automatically if a timeout period has
passed (see next entry) or if a number of requests have been added to the queue.
Recent kernels associate the queue with the submitting task and unplug also on a context switch.
T -- unplug due to timer If nobody requests the i/o that was queued after plugging the queue, Linux will automatically unplug it after a defined period has
passed.
M -- back merge A previously inserted request exists that ends on the boundary of where this i/o begins, so the i/o scheduler can merge them together.
F -- front merge Same as the back merge, except this i/o ends where a previously inserted requests starts.
Special:
B -- bounced The data pages attached to this bio are not reachable by the hardware and must be bounced to a lower memory location. This causes a big
slowdown in i/o performance, since the data must be copied to/from kernel buffers. Usually this can be fixed with using better hardware -- either a better
i/o controller, or a platform with an IOMMU.
S -- sleep No available request structures were available, so the issuer has to wait for one to be freed.
X -- split On raid or device mapper setups, an incoming i/o may straddle a device or internal zone and needs to be chopped up into smaller pieces for
service. This may indicate a performance problem due to a bad setup of that raid/dm device, but may also just be part of normal boundary conditions. dm
is notably bad at this and will clone lots of i/o.

4
Block device layer – events (simplified)
App / A / X Q Merge with an

existing request
N Y
G mergeable M/F
Need to Generate a
new request I P
Unplug on upper limit (stream) or

Plug queue and wait a bit if U
Time reached / submitting task ctx switch
following requests can be merged
Dispatch from block device

D
Add device driver info like dasdstat and layer to device driver
scsi sysfs statistics to fill this gap
and gain a full circle latency insight
Time from Dispatch to Complete

C

4
blktrace
 Example Case
– The snippet shows a lot of 4k requests (8x512 byte sectors)
• We expected the I/O to be 32k
– Each one is dispatched separately (no merges)
• This caused unnecessary overhead and slow I/O
Maj/Min CPU Seqnr sec.nsec pid Action RWBS sect + size map source / task
94,4 27 21 0.059363692 18994 A R 20472832 + 8 < (94,5) 20472640
94,4 27 22 0.059364630 18994 Q R 20472832 + 8 [qemukvm]
94,4 27 23 0.059365286 18994 G R 20472832 + 8 [qemukvm]
94,4 27 24 0.059365598 18994 I R 20472832 + 8 ( 312) [qemukvm]
94,4 27 25 0.059366255 18994 D R 20472832 + 8 ( 657) [qemukvm]
94,4 27 26 0.059370223 18994 A R 20472840 + 8 < (94,5) 20472648
94,4 27 27 0.059370442 18994 Q R 20472840 + 8 [qemukvm]
94,4 27 28 0.059370880 18994 G R 20472840 + 8 [qemukvm]
94,4 27 29 0.059371067 18994 I R 20472840 + 8 ( 187) [qemukvm]
94,4 27 30 0.059371473 18994 D R 20472840 + 8 ( 406) [qemukvm]

4
blktrace
 Example Case
– Analysis turned out that the I/O was from the swap code
• Same offsets were written by kswapd
– A recent code change there disabled the ability to merge I/O
– The summary below shows the difference after a fix
Total initially
Reads Queued: 560,888, 2,243MiB Writes Queued: 226,242, 904,968KiB
Read Dispatches: 544,701, 2,243MiB Write Dispatches: 159,318, 904,968KiB
Reads Requeued: 0 Writes Requeued: 0
Reads Completed: 544,716, 2,243MiB Writes Completed: 159,321, 904,980KiB
Read Merges: 16,187, 64,748KiB Write Merges: 61,744, 246,976KiB
IO unplugs: 149,614 Timer unplugs: 2,940
Total after Fix
Reads Queued: 734,315, 2,937MiB Writes Queued: 300,188, 1,200MiB
Read Dispatches: 214,972, 2,937MiB Write Dispatches: 215,176, 1,200MiB
Reads Requeued: 0 Writes Requeued: 0
Reads Completed: 214,971, 2,937MiB Writes Completed: 215,177, 1,200MiB
Read Merges: 519,343, 2,077MiB Write Merges: 73,325, 293,300KiB
IO unplugs: 337,130 Timer unplugs: 11,184

4
ziomon
 Characteristics: in depth zfcp based I/O analysis
 Objective: Analyze your FCP based I/O
 Usage: “ziomon” → “ziorep*”
 Package: RHEL: s390utils(-ziomon) SLES: s390-tools
ziomon Tools ziorep_config
Collect data Generate reports

using ziomon using ziorep_* ziorep_traffic
ziorep_utilization
Target
system Data Data
.csv
.log .agg .cfg .config
 Be aware that ziomon can be memory greedy if you have very memory constrained systems
 The has many extra functions please check out the live virtual class of Stefan Raspl
– PDF: http://www.vm.ibm.com/education/lvc/LVC0425.pdf
– Replay: http://ibmstg.adobeconnect.com/p7zvdjz0yye/
4
TCPDump
 Characteristics: dumps network traffic to console/file
 Objective: analyze packets of applications manually
 Usage: “tcpdump ...”
 Package: RHEL: tcpdump SLES: tcpdump
tcpdump host pserver1
tcpdump: verbose output suppressed, use v or vv for full protocol decode
listening on eth0, linktype EN10MB (Ethernet), capture size 65535 bytes
13:30:00.326581 IP pserver1.boeblingen.de.ibm.com.38620 > p10lp35.boeblingen.de.ibm.com.ssh: Flags [.], ack 3142, win
102, options [nop,nop,TS val 972996696 ecr 346994], length 0
13:30:00.338239 IP p10lp35.boeblingen.de.ibm.com.ssh > pserver1.boeblingen.de.ibm.com.38620: Flags [P.], seq 3142:3222,
ack 2262, win 2790, options [nop,nop,TS val 346996 ecr 972996696], length 80
13:30:00.375491 IP pserver1.boeblingen.de.ibm.com.38620 > p10lp35.boeblingen.de.ibm.com.ssh: Flags [.], ack 3222, win
102, options [nop,nop,TS val 972996709 ecr 346996], length 0
[...]
^C
31 packets captured
31 packets received by filter
0 packets dropped by kernel
 Not all devices support dumping packets in older distribution releases

– Also often no promiscuous mode
 Check flags or even content if your expectations are met
 -w flag exports captured unparsed data to a file for later analysis in libpcap format
– Also supported by wireshark
 Usually you have to know what you want to look for
4
Java Performance in general
 “Too” many choices

– There are many Java performance tools out there
 Be aware of common Java myths often clouding perception
 Differences
– Profiling a JVM might hide the Java methods
– Memory allocation of the JVM isn't the allocation of the Application

4
Java - Health Center
 Characteristics: Lightweight Java Virtual Machine Overview

 Objective: Find out where memory is leaked, sub-optimally cached, ...
 Usage: IBM Support Assistant (Eclipse)
 Package: RHEL: n/a SLES: n/a WWW: ibm.com/developerworks/java/jdk/tools/healthcenter
Java Agents integrated V5SR10+, V6SR3+, usually no target install required
 Shows
– Memory usage
– Method Profiling
– I/O Statistics
– Class loading
– Locking
 Hints
– Low overhead, therefore even suitable for monitoring
– Agent activation -Xhealthcenter:port=12345
– Can trigger dumps or verbosegc for in-depth memory analysis

4
Health Center - example
 The method profiling view will show the name of the methods and their time
consumption as percentage
139 Linux-Performance-know it all series © 2014 IBM Corporation
Java - Garbage Collection and Memory Visualizer
 Characteristics: in-depth Garbage Collection analysis

 Objective: Analyze JVM memory management
 Usage: IBM Support Assistant (Eclipse)
 Package: RHEL: n/a SLES: n/a WWW: ibm.com/developerworks/java/jdk/tools/gcmv
reads common verbosegc output, so usually no target install required
 Shows
– Memory usage
– Garbage Collection activities
– Pauses
– Memory Leaks by stale references
 Hints
– GCMV can also compare output of two runs
– Activate verbose logs verbose:gc Xverbosegclog:<log_file>
4
Garbage Collection and Memory Visualizer
 The GCMV reports statistics about garbage collections

 Most important values / indicators are:
– Proportion of time spent in gc pauses (should be less than 5%)
– For gencon: global collections << minor collections
Java - Jinsight
 Characteristics: zoomable call stack

 Objective: Analyze method call frequency and duration
 Usage: jinsight_trace tracemethods <yourProgram> <yourProgramArgs>
 Package: RHEL: n/a SLES: n/a WWW: IBM alphaworks
 Shows
– Call Stack and time
 Hints
– Significant slowdown, not applicable to production systems
– No more maintained, but so far still working

4
Jinsight Execution View
Threads
 The execution view groups threads in columns, select one to zoom in

 function calls can be seen as vertical bar (colors represent the call depth)
 Time passes from top to bottom

Jinsight Execution View, continued

Method Call Stack
Execution Time
 Many horizontal stages mean deep call stacks
 Long vertical areas mean long method execution
 Rectangles full of horizontal lines can be an issue

End of Part IV

4
Agenda


4
Cachestat
 Characteristics: Simple per page views of caching
 Objective: Detect what parts of a file are in page cache
 Usage: Write – or search for example code
 Package: n/a (pure code around the mincore system call)
 Shows
– How much of a file is in cache
 Hints
– We are now going from unsupported to non existent packages
– Still the insight can be so useful, it is good to know

4
Cachestat usage
./cachestat v ../Music/mysong.flac
pages in cache: 445/12626 (3.5%) [filesize=50501.0K, pagesize=4K]
cache map:
0: |x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|
[...]
416: |x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x|x| | | | |
448: | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
480: | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
[...]
12576: | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
12608: | | | | | | | | | | | | | | | | | |x|
 Here I show how much of a file is in cache while playing a song

– You'll see readahead here
– You'll also see the last block is almost always read in this case

4
smem
 Characteristics: Memory usage details per process/mapping

 Objective: Where is userspace memory really used
 Usage: smem -tk -c "pid user command swap vss uss pss rss”
 smem -m -tk -c "map count pids swap vss uss rss pss avgrss avgpss"
 Package: RHEL: n/a SLES: n/a WWW http://www.selenic.com/smem/

 Shows
– Pid, user, Command or Mapping, Count, Pid
– Memory usage in categories vss, uss, rss, pss and swap
 Hints
– Has visual output (pie charts) and filtering options as well
– No support for huge pages or transparent huge pages (kernel interface missing)

4
smem – process overview

smem tk c "pid user command swap vss uss pss rss”
PID User Command Swap VSS USS PSS RSS
1860 root /sbin/agetty s sclp_line0 0 2.1M 92.0K 143.0K 656.0K
1861 root /sbin/agetty s ttysclp0 11 0 2.1M 92.0K 143.0K 656.0K
493 root /usr/sbin/atd f 0 2.5M 172.0K 235.0K 912.0K
1882 root /sbin/udevd 0 2.8M 128.0K 267.0K 764.0K
1843 root /usr/sbin/crond n 0 3.4M 628.0K 693.0K 1.4M
514 root /bin/dbusdaemon system 0 3.2M 700.0K 771.0K 1.5M
524 root /sbin/rsyslogd n c 5 0 219.7M 992.0K 1.1M 1.9M
2171 root ./hhhptest 0 5.7G 1.0M 1.2M 3.2M
1906 root bash 0 103.8M 1.4M 1.5M 2.1M
2196 root ./hhhptest 0 6.2G 2.0M 2.2M 3.9M
1884 root sshd: root@pts/0 0 13.4M 1.4M 2.4M 4.2M
1 root /sbin/init 0 5.8M 2.9M 3.0M 3.9M
2203 root /usr/bin/python /usr/bin/sm 0 109.5M 6.1M 6.2M 6.9M
 How much of a process is:

– Swap - Swapped out
– VSS - Virtually allocated
– USS - Really unique
– RSS - Resident
– PSS - Resident accounting a proportional part of shared memory

4
smem – mappings overview

smem m tk c "map count pids swap vss uss rss pss avgrss avgpss"
Map Count PIDs Swap VSS USS RSS PSS AVGRSS AVGPSS
[stack:531] 1 1 0 8.0M 0 0 0 0 0
[vdso] 25 25 0 200.0K 0 132.0K 0 5.0K 0
/dev/zero 2 1 0 2.5M 4.0K 4.0K 4.0K 4.0K 4.0K
/usr/lib64/sasl2/libsasldb.so.2.0.23 2 1 0 28.0K 4.0K 4.0K 4.0K 4.0K 4.0K
/bin/dbusdaemon 3 1 0 404.0K 324.0K 324.0K 324.0K 324.0K 324.0K
/usr/sbin/sshd 6 2 0 1.2M 248.0K 728.0K 488.0K 364.0K 244.0K
/bin/systemd 2 1 0 768.0K 564.0K 564.0K 564.0K 564.0K 564.0K
/bin/bash 2 1 0 1.0M 792.0K 792.0K 792.0K 792.0K 792.0K
[stack] 25 25 0 4.1M 908.0K 976.0K 918.0K 39.0K 36.0K
/lib64/libc2.14.1.so 75 25 0 40.8M 440.0K 9.3M 1.2M 382.0K 48.0K
/lib64/libcrypto.so.1.0.0j 8 4 0 7.0M 572.0K 2.0M 1.3M 501.0K 321.0K
[heap] 16 16 0 8.3M 6.4M 6.9M 6.6M 444.0K 422.0K
<anonymous> 241 25 0 55.7G 20.6M 36.2M 22.3M 1.4M 913.0K
 How much of a mapping is:

– Swap - Swapped out
– VSS - Virtually allocated
– USS - Really unique
– RSS - Resident
– PSS - Resident accounting a proportional part of shared memory
– Averages as there can be multiple mappers

4
smem - visualizations
 There are many options to visualize memory distribution

 But avoid using that for monitoring as the proc/<pid>/smaps interface is an
expensive one
4
Valgrind
 Characteristics: in-depth memory analysis

 Objective: Find out where memory is leaked, sub-optimally cached, ...
 Usage: valgrind [program]
 Package: RHEL: valgrind SLES: valgrind
 Shows
– Memory leaks
– Cache profiling
– Heap profiling
 Hints
– Runs on binaries, therefore easy to use
– Debug Info not required but makes output more useful

4
Valgrind Overview
 Technology is based on a JIT (Just-in-Time Compiler)

 Intermediate language allows debugging instrumentation
– Binary is translated and augmented by valgrind instrumentation
– New binary is then executed
Binary
000000008000062c <main>: valgrind
kernel
stmg %r9,%r15,72(%r15)
lay %r15,-80160(%r15)
lhi %r12,0
lhi %r10,10000
la %r9,160(%r15)
lgr %r13,%r9
lgr %r11,%r9
System call interface

lghi %r2,1
brasl %r14,8000044c <malloc@plt>
lgfr %r1,%r12
translation
Replace
ahi %r12,1
stg %r2,0(%r11)
into IR
sllg %r1,%r1,3
aghi %r11,8
pfd
brct
2,96(%r1,%r9)
%r10,8000064c <main+0x20> some of New System call
lay %r12,80160(%r15)
The library
lg %r2,0(%r13)
aghi %r13,8
brasl %r14,8000048c <free@plt>
binary
cgrjne %r12,%r13,8000067e <main+0x52>
lhi %r13,0
calls by wrapper
instrumentation
lhi %r12,10000 xxx
lgfr %r2,%r13
Using a
ahi %r13,1
brasl %r14,800005c0 <stacker>
brct %r12,8000069c <main+0x70>
lg %r4,80272(%r15)
lmg
br
%r9,%r15,80232(%r15)
%r4 preload
library translation
libraries To machine code

4
Valgrind – sample output of “memcheck”
# valgrind buggy_program
==2799== Memcheck, a memory error detector
==2799== Copyright (C) 2002-2010, and GNU GPL'd, by Julian Seward et al.
==2799== Using Valgrind-3.6.1 and LibVEX; rerun with -h for copyright info
==2799== Command: buggy_program
==2799==
==2799== HEAP SUMMARY:
==2799== in use at exit: 200 bytes in 2 blocks
==2799== total heap usage: 2 allocs, 0 frees, 200 bytes allocated
==2799==
==2799== LEAK SUMMARY:
==2799== definitely lost: 100 bytes in 1 blocks
==2799== indirectly lost: 0 bytes in 0 blocks
==2799== possibly lost: 0 bytes in 0 blocks
==2799== still reachable: 100 bytes in 1 blocks
==2799== suppressed: 0 bytes in 0 blocks
==2799== Rerun with --leak-check=full to see details of leaked memory
[...]
 Important parameters:
– --leak-check=full
– --track-origins=yes
4
Valgrind - Tools
 Several tools
– Memcheck (default): detects memory and data flow problems
– Cachegrind: cache profiling
– Massif: heap profiling
– Helgrind: thread debugging
– DRD: thread debugging
– None: no debugging (for valgrind JIT testing)
– Callgrind: codeflow and profiling
 Tool can be selected with –tool=xxx

 System z support since version 3.7 (SLES-11-SP2)
 Backports into 3.6 (SLES-10-SP4, RHEL6-U1)

4
Valgrind - Good to know
 No need to recompile, but

– Better results with debug info
– Gcc option -O0 might result in more findings(the compiler might hide some
errors)
– Gcc option -fno-builtin might result in more findings
 --trace-children=yes will also debug child processes

 Setuid programs might cause trouble
– Valgrind is the process container (→ no setuid)
– Possible solution: remove setuid and start as the right user, check
documentation for other ways
 The program will be slower

– 5-30 times slower for memcheck

4
IRQ Statistics
 Characteristics: Low overhead IRQ information

 Objective: Condensed overview of IRQ activity
 Usage: cat /proc/interrupts and cat /proc/softirqs
 Shows
– Which interrupts happen on which cpu
– Where softirqs and tasklets take place
 Hints
– Recent Versions (SLES11-SP2) much more useful due to better naming
– If interrupts are unintentionally unbalanced
– If the amount of interrupts matches I/O
• This can point to non-working IRQ avoidance
– Essentially this is the data source for “mpstat -I” which we had before
4
IRQ Statistics
 Example
– Network focused on CPU zero (in this case unwanted)
– Scheduler covered most of that avoiding idle CPU 1-3
– But caused a lot migrations, IPI's and cache misses
CPU0 CPU1 CPU2 CPU3
EXT: 21179 24235 22217 22959
I/O: 1542959 340076 356381 325691
CLK: 15995 16718 15806 16531 [EXT] Clock Comparator
EXC: 255 325 332 227 [EXT] External Call
EMS: 4923 7129 6068 6201 [EXT] Emergency Signal
TMR: 0 0 0 0 [EXT] CPU Timer
TAL: 0 0 0 0 [EXT] Timing Alert
PFL: 0 0 0 0 [EXT] Pseudo Page Fault
DSD: 0 0 0 0 [EXT] DASD Diag
VRT: 0 0 0 0 [EXT] Virtio
SCP: 6 63 11 0 [EXT] Service Call
IUC: 0 0 0 0 [EXT] IUCV
CPM: 0 0 0 0 [EXT] CPU Measurement
CIO: 163 310 269 213 [I/O] Common I/O Layer Interrupt
QAI: 1 541 773 338 857 354 728 324 110 [I/O] QDIO Adapter Interrupt
DAS: 1023 909 1384 1368 [I/O] DASD
[…] 3215, 3270, Tape, Unit Record Devices, LCS, CLAW, CTC, AP Bus, Machine Check

4
IRQ Statistics II
 Also softirqs can be tracked which can be useful to
– check if tasklets execute as intended
– See if network, scheduling and I/O behave as expected
CPU0 CPU1 CPU2 CPU3
HI: 498 1522 1268 1339
TIMER: 5640 914 664 643
NET_TX: 15 16 52 32
NET_RX: 18 34 87 45
BLOCK: 0 0 0 0
BLOCK_IOPOLL: 0 0 0 0
TASKLET: 13 10 44 20
SCHED: 8055 702 403 445
HRTIMER: 0 0 0 0
RCU: 5028 2906 2794 2564

4
Wireshark
 Characteristics: Analyzes captured network traffic
 Objective: In depth analysis of handshakes, missing replies, protocols, ...
 Usage: Dump in libpcap or pcap-ng format (tcpdump, dumpcap)
then analyze on remote system via “wireshark”
 Package: RHEL: wireshark SLES: wireshark
 No “direct” invocation on System z usually

– e.g. on RH6 there is not even a wireshark binary
 Scrolling huge files on Remote X isn't fun anyway
– Capturing tools are available
 Custom columns and profiles are important to visualize what you want to look for
 For more details you might start at
– The share sessions of Mathias Burkhard
https://share.confex.com/share/121/webprogram/Session13282.html
– Official documentation http://www.wireshark.org/docs/wsug_html/
4
Wireshark example
 1. Dump via “tcpdump w” or wiresharks “dumpcap”
 2. analyze on remote system via “wireshark”
tcpdump host pserver1 w traceme
tcpdump: listening on eth0, linktype EN10MB (Ethernet), capture size 65535 bytes
^C40 packets captured
40 packets received by filter
0 packets dropped by kernel
[scp to my system]
wireshark traceme

4
Tracepoints (Events)
 Characteristics: Complex interface, but a vast source of information

 Objective: In kernel latency and activity insights
 Usage: Access debugfs mount point /tracing
 Shows
– Timestamp and activity name
– Tracepoints can provide event specific context data
– Infrastructure adds extra common context data like cpu, preempts depth, ...
 Hints
– Very powerful and customizable, there are hundreds of tracepoints
• Some tracepoints have tools to be accessed “perf sched”, “blktrace” both base on them
• Others need custom postprocessing
– There are much more things you can handle with tracepoints check out
Kernel Documentation/trace/tracepoint-analysis.txt (via perf stat)
Kernel Documentation/trace/events.txt (custom access)
4
Tracepoints – example I/III
 Here we use custom access since there was tool

– We searched for 1.2ms extra latency
• Target is it lost in HW, Userspace, Kernel or all of them
– Workload was a simple 1 connection 1 byte←→1 byte load
– Call “perf list” for a list of currently supported tracepoints
– We used the following tracepoints

Abbreviation Tracepoint Meaning
R netif_receive_skb low level receive
P napi_poll napi work related to receive
Q net_dev_queue enqueue in the stack
S net_dev_xmit low level send

4
Tracepoints – example II/III

– (Simplified) Script
• # full versions tunes buffer sizes, checks files, ...
echo latencyformat > /sys/kernel/debug/tracing/trace_options # enable tracing type
echo net:* >> /sys/kernel/debug/tracing/set_event # select specific events
echo napi:* >> /sys/kernel/debug/tracing/set_event # “
echo "name == ${dev}" > /sys/kernel/debug/tracing/events/net/filter # set filters
echo "dev_name == ${dev}" > /sys/kernel/debug/tracing/events/napi/filter # “
cat /sys/kernel/debug/tracing/trace >> ${output} # synchronous
echo !*:* > /sys/kernel/debug/tracing/set_event # disable tracing
– Output
# _=> CPU#
# / _=> irqsoff
# | / _=> needresched
# || / _=> hardirq/softirq
# ||| / _=> preemptdepth
# |||| / delay
# cmd pid ||||| time | caller
# \ / ||||| \ | /
<...>24116 0..s. 486183281us+: net_dev_xmit: dev=eth5 skbaddr=0000000075b7e3e8 len=67 rc=0
<idle>0 0..s. 486183303us+: netif_receive_skb: dev=eth5 skbaddr=000000007ecc6e00 len=53
<idle>0 0.Ns. 486183306us+: napi_poll: napi poll on napi struct 000000007d2479a8 fordevice eth
<...>24116 0..s. 486183311us+: net_dev_queue: dev=eth5 skbaddr=0000000075b7e3e8 len=67
<...>24116 0..s. 486183317us+: net_dev_xmit: dev=eth5 skbaddr=0000000075b7e3e8 len=67 rc=0

4
Tracepoints – example III/III
 Example postprocessed
SUM COUNT AVERAGE MIN MAX STDDEV
P2Q: 8478724 1572635 5.39 4 2140 7.41
Q2S: 12188675 1572638 7.65 3 71 4.89
S2R: 38562294 1572636 24.42 1 2158 9.08
R2P: 4197486 1572633 2.57 1 43 2.39
SUM: 63427179 1572635 40.03
SUM COUNT AVERAGE MIN MAX STDDEV
P2Q: 7191885 1300897 5.53 4 171 1.31
Q2S: 10622270 1300897 8.17 3 71 5.99
S2R: 32078550 1300898 24.66 2 286 5.88
R2P: 3707814 1300897 2.85 1 265 2.59
SUM: 53600519 1300897 41.20
– Confirmed that ~all of the 1.2 ms were lost inside Linux (not in the fabric)
– Confirmed that it was not at/between specific function tracepoints
• Eventually it was an interrupt locality issue causing bad caching

4
Systemtap
 Characteristics: tool to “tap” into the kernel for analysis
 Objective: analyze in kernel values or behavior that otherwise would be
inaccessible or require a modification/recompile cycle
 Usage (mini example): “stap v e 'probe vfs.read {printf(“read
performed\n”); exit()}'”
 Package: RHEL: systemtap + systemtap-runtime SLES: systemtap
 Also requires kernel debuginfo and source/devel packages
 Procedural and C-like language based on two main constructs

– Probes – “catching events”
• On functions, syscalls or single statements via file:linenumber
– Functions – “what to do”
• Supports local and global variables
• Program flow statements if, loops, …
 Tapsets provide pre written probe libraries
 Fore more check out “Using SystemTap on Linux on System z” from Mike O'Reilly
https://share.confex.com/share/118/webprogram/Handout/Session10452/atlanta.pdf
4
There would be even more tools to cover ...
 Further tools - (no slides yet)

– Collectl – full system monitoring
– Ftrace – kernel function tracing
– Lttng – complex latency tracing infrastructure (packages start to appear
in Fedora 19)
– Nicstat, ktap, stap, ...

4
End of Part V

4
Questions
 Further information is available at

–Linux on System z – Tuning hints and tips
http://www.ibm.com/developerworks/linux/linux390/perf/index.html
– Live Virtual Classes for z/VM and Linux
http://www.vm.ibm.com/education/lvc/
Christian Ehrhardt Research & Development

Linux on System z Schönaicher Strasse 220
Performance Evaluation 71032 Böblingen, Germany
ehrhardt@de.ibm.com
September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

4

Understanding Linux On System Z Performance Tools

Uploaded by

Copyright:

Available Formats

Understanding Linux On System Z Performance Tools

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Understanding Linux On System Z Performance Tools

Uploaded by

Copyright:

Available Formats

Christian Ehrhardt

Linux Performance know-it-all series

How to Surprise by being a

Christian Ehrhardt, IBM R&D Germany,

© 2013 IBM Corporation

2 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Basic Intermediate Advanced Master Elite

4 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

 Categories qualify what the core was used for

 Accounting unit is Jiffy, reports usually as percentage

 Imagine a laptop is a CPU core

 People using that laptop are different Programs, that could be

*Roles are defined by clothes and equipment

6 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

 If a userspace application is running it is accounted as USR

7 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – SYS, (H)IRQ, SIRQ

8 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – SYS, (H)IRQ, SIRQ

This is the most simple case of “doing nothing”

10 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Waiting for I/O

CPU is Idle Application

Again “doing nothing”, but no more that simple

Waiting for I/O

CPU is Idle, but No one waiting for I/O

 If no one is waiting for I/O, (asynchronous)

 The CPU is doing something, just not for you

13 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Nobody cares about

14 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

16 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

17 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

18 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

19 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

20 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

21 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

22 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

23 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

24 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz

25 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

Utilization – Steal Quiz summary

 So in our metaphor we have multiple people of the same privilege level

27 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

 Well, that is easy a single program means no context switches

28 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

29 September 3, 201 Linux-Performance-know it all series © 2014 IBM Corporation

CPU busy, but primarily for overhead

Add steal time if you want real trouble