Mug21 DL ML v3

Follow us on
https://twitter.com/mvapich
Boosting Performance of Deep Learning, Machine

Learning, and Dask with MVAPICH2
Tutorial at MUG ’21
by
Arpan Jain Aamir Shafi Quentin Anthony

The Ohio State University The Ohio State University The Ohio State University
jain.575@osu.edu shafi.16@osu.edu anthony.301@osu.edu
http://u.osu.edu/jain.575 https://cse.osu.edu/people/shafi.16 https://www.linkedin.com/in/quentin-anthony
Outline
• Introduction
• Deep Neural Network Training and Essential Concepts
• Parallelization Strategies for Distributed DNN Training
• Machine Learning
• Data Science using Dask
• Conclusion
Network Based Computing Laboratory MUG ‘21 2

What is Machine Learning and Deep Learning?
• Machine Learning (ML)
– “the study of computer algorithms to improve
automatically through experience and use of data”
• Deep Learning (DL) – a subset of ML
– Uses Deep Neural Networks (DNNs)
– Perhaps, the most revolutionary subset!
• Based on learning data representation
• DNN Examples: Convolutional Neural Networks, Recurrent
Neural Networks, Hybrid Networks
• Data Scientist or Developer Perspective for using
DNNs
1. Identify DL as solution to a problem
2. Determine Data Set
3. Select Deep Learning Algorithm to Use Courtesy: https://hackernoon.com/difference-between-artificial-intelligence-machine-learning-and-
4. Use a large data set to train an algorithm deep-learning-1pcv3zeg, https://blog.dataiku.com/ai-vs.-machine-learning-vs.-deep-learning,
https://en.wikipedia.org/wiki/Machine_learning

History: Milestones in the Development of ML/DL
Multi-layered
XOR
Problem Perceptron
(Backpropagation) AlexNet
ADALINE WGAN
Perceptron ResNet
Golden Dark Age DBN
Electronic Transformers
Age (“AI Winter”)
Brain
1943 1957 1960 1969 1986 2006 2012 2015 2017
1800 1900 …. 1940 1950 1960 1970 1980 1990 2000 2010 2020
1805 1901 1936 1954 1965 1967 1979 1985 1995 2000 2014 2017
Bayesian Deep
Linear K-Means
Turing Machine Network Forest
Regression
KNN SVM XGBoost
Evolutionary
PCA Decision Trees Random Forest CatBoost
Algorithms
A. Legendre – J. Gauss K. Pearson A. Turing S. McCulloch – W. Pitts F. Rosenblatt B. Widrow – M. Hoff M. Minsky – S. Papert J. Pearl D. Rumelhart – G. Hinton – R. Wiliams V. Vapnik– C. Cortes A. Krizhevsky A. Ng Y. LeCun Y. Bengio

Deep Learning meets Super Computers Accelerator/CP Family
Performance Share
• Computation requirement is increasing exponentially

– CPUs still dominates HPC arena and can be used for Deep Learning
www.top500.org
Courtesy: https://openai.com/blog/ai-and-compute/
Outline
• Introduction
• Conclusion

So what is a Deep Neural Network?
• Example of a 3-layer Deep Neural Network (DNN) – (input layer is not counted)
Courtesy: http://cs231n.github.io/neural-networks-1/

DNN Training: Forward Pass
Input Hidden Hidden Output

Layer Layer Layer Layer

Forward Pass
W1
X
W2


Forward Pass
W3
W1
W4
X
W5
W2
W6


Forward Pass
W3
W1 W7
W4
X
W5
W2 W8
W6


Forward Pass
W3
W1 W7
W4
X Pred
W5
W2 W8
Error = Loss(Pred,Output)
W6


DNN Training: Backward Pass
Forward Pass
E7
E8
Backward Pass


Forward Pass
E3
E7
E4
E5
E8
E6
Backward Pass


Forward Pass
E3
E1 E7
E4
E5
E2 E8
E6
Backward Pass


DNN Training

Essential Concepts: Activation function and Back-propagation
• Back-propagation involves
complicated mathematics.
– Luckily, most DL Frameworks give
you a one line implementation --
model.backward()
• What are Activation functions? Courtesy: https://www.jeremyjordan.me/neural-networks-training/
– RELU (a Max fn.) is the most common activation fn.

– Sigmoid, tanh, etc. are also used
Essential Concepts: Learning Rate (α)
Courtesy: https://www.jeremyjordan.me/nn-learning-rate/

Essential Concepts: Batch Size
• Batched Gradient Descent N
– Batch Size = N
• Stochastic Gradient Descent

– Batch Size = 1
• Mini-batch Gradient Descent
– Somewhere in the middle
– Common:
• Batch Size = 64, 128, 256, etc.
• Finding the optimal batch

Batch Size One full pass over N is called an epoch of training
size will yield the fastest
learning.
Courtesy: https://www.jeremyjordan.me/gradient-descent/

Key Phases of Deep Learning
• Training is compute intensive
– Many passes over data
– Can take days to weeks
– Model adjustment is done
• Inference
– Single pass over the data
– Should take seconds
– No model adjustment Courtesy: https://devblogs.nvidia.com/
• Challenge: How to make “Training” faster?

– Need Parallel and Distributed Training…

Outline
• Introduction
• Conclusion

Parallelization Strategies
• Some parallelization strategies..
– Data Parallelism
– Model Parallelism
– Hybrid Parallelism
Model Parallelism
Hybrid (Model and Data) Parallelism

Data Parallelism
Courtesy: http://engineering.skymind.io/distributed-deep-learning-part-1-an-introduction-to-distributed-training-of-neural-networks
Outline
• Introduction
– Model and Hybrid Parallelism
• Conclusion

Need for Data Parallelism
Let’s revisit Mini-Batch Gradient Descent
Drawback: If the dataset has 1 million images, then it will take forever to run the model on such a
big dataset
Solution: Can we use multiple machines to speedup the training of Deep learning models? (i.e.
Utilize Supercomputers to Parallelize)
Need for Communication in Data Parallelism
Y Y
N N
Y Y
Y Y
N N
Y Y
Y Y
Machine 1 Y Machine 4 Y
Y Y
N N
Y Y
Y Y
N N
Y Y
Y Y
Y
Machine 2 Y
Machine 5
Y
N Problem: Train a single model on whole dataset,
Y
Y not 5 models on different sets of dataset
N
Y
Y
Machine 3 Y

Data Parallelism
1 7 2 12 19 9
Machine 1 2 2 4
5 1 3
12 12 12
21 5 11
1 3 2 12 19 9
1 2 3
Machine 2 5 1 2
12 12 12
21 5 11
2 1 3 MPI 12 19 9
Machine 3 5 5 2
5 1 1 AllReduce
12 12 12
21 5 11
5 7 1 12 19 9
Machine 4 2 1 2
4 1 3
12 12 12
21 5 11
3 1 1 12 19 9
Machine 5 2 2 1
2 1 2
12 12 12
21 5 11
Gradients Reduced
Gradients
Allreduce Collective Communication Pattern
• Element-wise Sum data from all processes and sends to all processes
int MPI_Allreduce (const void *sendbuf, void * recvbuf, int count, MPI_Datatype datatype,
MPI_Op operation, MPI_Comm comm)
Input-only Parameters Sendbuf (Before)

Parameter Description
T1 T2 T3 T4
sendbuf Starting address of send buffer
1 1 1 1
recvbuf Starting address of recv buffer 2 2 2 2
3 3 3 3
type Data type of buffer elements
4 4 4 4
count Number of elements in the buffers
operation Reduction operation to be performed (e.g. sum) Recvbuf (After)
comm Communicator handle
T1 T2 T3 T4
Input/Output Parameters 4 4 4 4
8 8 8 8
Parameter Description 12 12
12 12
recvbuf Starting address of receive buffer 16 16 16 16

Data Parallelism and MPI Collectives
• Step1: Data Propagation
– Distribute the Data among GPUs
• Step2: Forward Backward Pass
– Perform forward pass and
calculate the prediction
– Calculate Error by comparing
prediction with actual output
– Perform backward pass and
calculate gradients
• Step3: Gradient Aggregation
– Call MPI_Allreduce to reduce the
local gradients
– Update parameters locally using
global gradients

MVAPICH2 (MPI)-driven Infrastructure for ML/DL Training
ML/DL Applications ML/DL Applications
TensorFlow PyTorch MXNet PyTorch
Horovod Torch.distributed DeepSpeed
MVAPICH2 or MVAPICH2-X MVAPICH2-GDR for MVAPICH2 or MVAPICH2-X MVAPICH2-GDR for

for CPU Training GPU Training for CPU Training GPU Training
More details available from: http://hidl.cse.ohio-state.edu

MVAPICH2 2.3.6
• Released on 05/11/2021
• Major Features and Enhancements
– Support collective offload using Mellanox's SHARP for Reduce and Bcast
• Enhanced tuning framework for Reduce and Bcast using SHARP
– Enhanced performance for UD-Hybrid code
– Add multi-rail support for UD-Hybrid code
– Enhanced performance for shared-memory collectives
– Enhanced job-startup performance for flux job launcher
– Add support in mpirun_rsh to use srun daemons to launch jobs
– Add support in mpirun_rsh to specify processes per node using '-ppn' option
– Use PMI2 by default when SLURM is selected as process manager
– Add support to use aligned memory allocations for multi-threaded applications
– Architecture detection and enhanced point-to-point tuning for Oracle BM.HPC2 cloud shape
– Enhanced collective tuning for Frontera@TACC and Expanse@SDSC
– Add support for GCC compiler v11
– Add support for Intel IFX compiler
– Update hwloc v1 code to v1.11.14 & hwloc v2 code to v2.4.2

MVAPICH2-GDR 2.3.6
• Released on 08/12/2021
• Major Features and Enhancements
– Added several new MPI_T PVARs
– Based on MVAPICH2 2.3.6
– Added support for CUDA 11.3
– Added support for 'on-the-fly' compression of point-to-point messages used for
– Added support for ROCm 4.1
GPU-to-GPU communication
– Enhanced output for runtime variable MV2_SHOW_ENV_INFO
• Applicable to NVIDIA GPUs
– Added NCCL communication substrate for various MPI collectives – Tested with Horovod and common DL Frameworks
• TensorFlow, PyTorch, and MXNet
• Support for hybrid communication protocols using NCCL-based, CUDA-based, and IB verbs-
based primitives – Tested with MPI4Dask 0.2
• MPI_Allreduce, MPI_Reduce, MPI_Allgather, MPI_Allgatherv, MPI_Alltoall, MPI_Alltoallv, • MPI4Dask is a custom Dask Distributed package with MPI support
MPI_Scatter, MPI_Scatterv, MPI_Gather, MPI_Gatherv, and MPI_Bcast
– Tested with MPI4cuML 0.1
– Full support for NVIDIA DGX, NVIDIA DGX-2 V-100, and NVIDIA DGX-2 A-100 systems • MPI4cuML is a custom cuML package with MPI support
• Enhanced architecture detection, process placement and HCA selection
• Enhanced intra-node and inter-node point-to-point tuning
• Enhanced collective tuning
– Introduced architecture detection, point-to-point tuning and collective tuning for

ThetaGPU @ANL
– Enhanced point-to-point and collective tuning for NVIDIA GPUs on Frontera @TACC,
Lassen @LLNL, and Sierra @LLNL
– Enhanced point-to-point and collective tuning for Mi50 and Mi60 AMD GPUs on
Corona @LLNL

Install Horovod with MVAPICH2-X and MVAPICH2-GDR
Command to install Horovod with MVAPICH2-X
$ HOROVOD_WITH_MPI=1 pip install --no-cache-dir horovod==0.22.1
Command to install Horovod with MVAPICH2-GDR
$ HOROVOD_GPU_ALLREDUCE=MPI HOROVOD_CUDA_HOME=/opt/cuda/10.1 HOROVOD_WITH_MPI=1 pip

install --no-cache-dir horovod

Run PyTorch on a single GPU
+ python pytorch_synthetic_benchmark.py --batch-size 64 --num-iters=5

.
.
Model: resnet50 V100
Batch size: 64
Number of GPUs: 1
Running warmup...
Running benchmark...
Iter #0: 333.9 img/sec per GPU
Img/sec per GPU: 334.0 +-0.2
-----------------------------------------
Total img/sec on 1 GPU(s): 334.0 +-0.2
-----------------------------------------

Run PyTorch on two nodes with 1 GPU/node (using MVAPICH2-
GDR)
+ mpirun_rsh -np 2 gpu11 gpu12 MV2_USE_CUDA=1 MV2_CPU_BINDING_POLICY=hybrid
MV2_HYBRID_BINDING_POLICY=spread MV2_USE_RDMA_CM=0
MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so LD_PRELOAD=mv2-gdr/lib/libmpi.so
python pytorch_synthetic_benchmark.py --batch-size 64 --num-iters=5
.
.
Model: resnet50 V100
Batch size: 64
Number of GPUs: 2
Running warmup...
Running benchmark...
Img/sec per GPU: 316.4 +-2.2
-----------------------------------------
Total img/sec on 2 GPU(s): 632.8 +-4.3
-----------------------------------------
~1.89X on
2 GPUs

Scalable TensorFlow using Horovod and MVAPICH2-GDR
• ResNet-50 Training using TensorFlow benchmark on 1 DGX-2 node (16 Volta GPUs)
Platform: Nvidia DGX-2 system, CUDA 9.2

7000 100
9% higher 90
6000
80
Scaling Efficiency (%)

5000
Image per second
70
60
4000
50
3000 40
2000 30
20
1000
10
0 0
1 2 4 8 16 1 2 4 8 16
Number of GPUs Number of GPUs
NCCL-2.6 MVAPICH2-GDR-2.3.3 NCCL-2.6 MVAPICH2-GDR-2.3.3
C.-H. Chu, P. Kousha, A. Awan, K. S. Khorassani, H. Subramoni and D. K. Panda, "NV-Group: Link-Efficient Reductions for Distributed Deep Learning on Modern Dense GPU
Systems, " ICS-2020.
Distributed TensorFlow on ORNL Summit (1,536 GPUs)
• ResNet-50 Training using
TensorFlow benchmark on 450
Thousands
SUMMIT -- 1536 Volta 400 ImageNet-1k has 1.2 million images
GPUs! 350
MVAPICH2-GDR reaching ~0.42 million
Image per second

300
• 1,281,167 (1.2 mil.) images 250 images per second for ImageNet-1k!
200
150
• Time/epoch = 3 seconds
100
50
• Total Time (90 epochs) 0
= 3 x 90 = 270 seconds = 4.5 1 2 4 6 12 24 48 96 192 384 768 1536
Number of GPUs
minutes!
NCCL-2.6 MVAPICH2-GDR 2.3.4
*We observed issues for NCCL2 beyond 384 GPUs
Platform: The Summit Supercomputer (#2 on Top500.org) – 6 NVIDIA Volta GPUs per node connected with NVLink, CUDA 10.1
Distributed TensorFlow on TACC Frontera (2048 CPU nodes)
• Scaled TensorFlow to 2048 nodes on
Frontera using MVAPICH2 and IntelMPI
262144
65536
16384
Images per sec

• MVAPICH2 and IntelMPI give similar 4096
performance for DNN training 1024
256
64
• Report a peak of 260,000 images/sec on 16
4
2048 nodes 1
1 2 4 8 16 32 64 128 256 512 1024 2048
Nodes
• On 2048 nodes, ResNet-50 can be trained
MVAPICH2-X Ideal
in 7 minutes!
A. Jain, A. A. Awan, H. Subramoni, DK Panda, “Scaling TensorFlow, PyTorch, and MXNet using MVAPICH2 for High-Performance Deep
Learning on Frontera”, DLS ’19 (SC ’19 Workshop).

Outline
• Introduction
– Model and Hybrid Parallelism
• Conclusion

HyPar-Flow: Hybrid Parallelism for TensorFlow
• Why Hybrid parallelism? Memory Consumption (Extrapolated)
– Data Parallel training has 512

Only possible with Model Parallelism!
limits! → 256
• We propose HyPar-Flow 128
64
– An easy to use Hybrid
Memory (GB)
CPU Broadwell (128 GB)
32
parallel training framework Volta GPU
16
• Hybrid = Data + Model CPU Skylake
8 (192 GB)
– Supports Keras models and 4 Pascal GPU
exploits TF 2.0 Eager 2
Execution 1
0 100 200 300 400 500 600 700 800
– Exploits MPI for Point-to-
Input Image Size (Width X Height)
point and Collectives ResNet-1k ResNet-5k
Benchmarking large-models lead to better insights and ability to develop new approaches!
A. A. Awan, A. Jain, Q. Anthony, H. Subramoni, and DK Panda, “HyPar-Flow: Exploiting MPI and Keras for Hybrid
Parallel Training of TensorFlow models”, ISC ’20, https://arxiv.org/pdf/1911.05146.pdf

Model/Hybrid Parallelism and MPI Collectives
• HyPar-Flow is practical (easy-to-use) and high-performance (uses MPI)
– Based on Keras models and exploits TF 2.0 Eager Execution
– Leverages MPI Pt-to-pt. and Collectives for communication
DNN with Identity (Residual) Mappings (Input)
hf.fit(model, num_partitions,
= Max Pooling
num_replicas, strategy)
HyPar-Flow
E.g. hf.fit(model, 3, 2, hybrid)
Allreduce Allreduce Allreduce

comm comm comm recv (D)
send(D) D
recv (∇) send(∇)
∇
Replica-0
send(D)
D recv(D) send(D)
D
recv(D)
recv(∇)
∇ send(∇) recv(∇)
∇
send(∇)
Replica-1
Image Conv3 Conv4 Output
Conv2
Conv1 FC
Partition-0 Partition-1 Partition-2
A. A. Awan, A. Jain, Q. Anthony, H. Subramoni, and DK Panda, “HyPar-Flow:

Exploiting MPI and Keras for Hybrid Parallel Training of TensorFlow models”, ISC ’20,
https://arxiv.org/pdf/1911.05146.pdf
HyPar-Flow at Scale (512 nodes on TACC Frontera)
• ResNet-1001 with variable batch
481x speedup on 512 Intel Xeon Cascade Lake nodes (TACC Frontera)
size 5000
4500
• Approach:
Img/sec (higher is better)

4000
– 48 model-partitions for 56 cores 3500
3000
– 512 model-replicas for 512 nodes 2500
– Total cores: 48 x 512 = 24,576 2000
1500
• Speedup 1000
– 253X on 256 nodes 500
0
– 481X on 512 nodes 1 2 4 8 16 32 64 128 256 512
Nodes (=Replicas)
• Scaling Efficiency
– 98% up to 256 nodes 56 Partitions (1 node) Ideal
– 93.9% for 512 nodes A. A. Awan, A. Jain, Q. Anthony, H. Subramoni, and DK Panda, “HyPar-Flow: Exploiting MPI and Keras for Hybrid
Parallel Training of TensorFlow models”, ISC ‘20, https://arxiv.org/pdf/1911.05146.pdf

GEMS: GPU Enabled Memory Aware Model Parallelism Systems
Why do we need Memory aware
designs?
– Data and Model Parallel
training has limitation!
– Maximum Batch Size

depends on the memory.
– Basic Model Parallelism

suffers from
underutilization of memory
and compute → Memory requirement increases with the increase in image size!
A. Jain, A. Awan, A. Aljuhani, J. Hashmi, Q. Anthony, H. Subramoni, D. Panda, R. Machiraju, A. Parwani, “GEMS: GPU Enabled Me mory Aware Model Parallelism System for
Distributed DNN”, SC ’20.

Exploiting Model Parallelism in AI-Driven Digital Pathology
• Pathology whole slide image (WSI)
– Each WSI = 100,000 x 100,000 pixels
– Can not fit in a single GPU memory
– Tiles are extracted to make training possible
• Two main problems with tiles
– Restricted tile size because of GPU memory limitation
– Smaller tiles loose structural information
• Can we use Model Parallelism to train on larger tiles to get
better accuracy and diagnosis?
• Reduced training time significantly
– 7.25 hours (1 node, 4 GPUs) -> 27 mins (32 nodes, 128 GPUs)
Courtesy: https://blog.kitware.com/digital-slide-
archive-large-image-and-histomicstk-open-source-
informatics-tools-for-management-visualization-and-
A. Jain, A. Awan, A. Aljuhani, J. Hashmi, Q. Anthony, H. Subramoni, D. Panda, R. Machiraju, A. Parwani, “GEMS: analysis-of-digital-histopathology-data/
GPU Enabled Memory Aware Model Parallelism System for Distributed DNN”, SC ’20.

SUPER: SUb-Graph Parallelism for TransformERs
Sub-Graph Parallelism
– Exploits inherent parallelism in
modern DNN architectures
– Improves the Performance of
multi-branch DNN architectures →
Simple example of a multi-branch DNN architecture
– Can be used to accelerate the

training of state-of-the-art
Transformer models
– Provides better than Data-

Parallelism for in-core models
4-way Sub-Graph Parallelism combined with Data-Parallelism (D&SP)
A. Jain, T. moon, T. Benson, H. Subramoni, S. Jacobs, D. Panda, B. Essen, “SUPER: SUb -Graph Parallelism for TransformERs”, IPDPS ’21

Accelerating Transformers using SUPER
• We propose sub-graph parallelism
Up to 3.05X speedup over Data Parallel designs (LLNL Lassen)
integrated with data parallelism to
accelerate the training of DP 2-way D&SP
4-way D&SP 8-way D&SP
Time per mini-batch (secs)

Transformers. 1.6
1.4
• Approach 1.2
1
– Data and Sub-Graph Parallelism (D&SP) 0.8
0.6
• #-way D&SP (#: number of sub-graphs)
0.4
0.2
• Setup 0
4 8 16 32 64 128 256 512 1024
– T5-Large-Mod on WMT Dataset #GPUs
– 1024 NVIDIA V100 GPUs
• Speedup
– Up to 3.05X over Data Parallelism (DP)
A. Jain, T. moon, T. Benson, H. Subramoni, S. Jacobs, D. Panda, B. Essen, “SUPER: SUb-Graph Parallelism for TransformERs”, IPDPS ’21

Outline
• Introduction
• Conclusion

Parallelizing K-Means Clustering
Proc 0
Proc 1
Domain Decomposition – Divide data objects amongst P processors Recompute centroids after MPI_Allreduce
Courtesy: https://github.com/tmscarla/k-means-parallel http://users.eecs.northwestern.edu/~wkliao/Kmeans/

Support for Parallel and Distributed Execution
• Scikit-learn:
– Supports execution via Joblib (https://joblib.readthedocs.io/en/latest/)
– Joblib supports multi-threaded and multi-process execution (on multiple
nodes)
• XGBoost:
– Multiple ways to run on cluster of nodes:
• Dask (http://dask.org)
• Ray (https://ray.io/)
• AWS YARN
• Apache Spark (https://spark.apache.org/) using XGBoost4J-Spark
• cuML:
– Execution is supported on multiple nodes using Dask (http://dask.org) and
NVIDIA’s NCCL
The cuML Library: Accelerating ML on GPUs
• The NVIDIA RAPIDS project aims to build end-to-end data science
analytic pipelines on GPUs
• An important component is the cuML library:
– GPU-accelerated ML library
– GPU-counterpart of Scikit-learn
– Supports the execution of ML workloads on Multi-Node Multi-GPUs (MNMG)
systems
• Most existing ML libraries, including Scikit-learn and Apache Spark’s
MLlib, only support CPU execution of ML algorithms
– Conventional wisdom has been that only DNNs are a good match for GPUs
because of high computational requirements
Main components of the cuML library
Python
• Main components Scikit-learn-like interfaces
– Python layer
Algorithms Linear Logistic
• Provides a Scikit-learn like interface regression regression
K-nearest Random
• Hides the complexities of the CUDA/C/C++ layer neighbors K-means
forest
– Primitives and cuML algorithms built on top of CUDA Primitives Element-wise Matrix
operations multiplication
• ML Algorithms Eigen standard distance/matrix
Decomposition deviation calculations
• Primitives
– Reusable building blocks for building machine learning Dask Python
algorithms
Cython
– Common for different machine learning algorithms
CUDA/C/C++
– Used to build different machine learning algorithms
UCX cuML Algorithms
– Communication Support in cuML:
cuML Primitives
• Point-to-point communication: Dask NCCL
CUDA Libraries
• Collective communication: NVIDIA Collective
Communications Library (NCCL) CUDA

Parallel and Distributed Training in cuML
• Distributed training
– Using data parallelism, the fit() function is
executed on all workers containing a partition of
the training dataset
• Trained Model:
– The trained parameters are brought to a single
worker using MPI_Reduce()
• Inference Stage:
– The trained parameters are broadcasted to all
workers with prediction partitions
Courtesy: https://arxiv.org/pdf/2002.04803.pdf
Accelerating cuML with MVAPICH2-GDR
• Utilize MVAPICH2-GDR (with mpi4py) as communication backend during the
training phase (the fit() function) in the Multi-node Multi-GPU (MNMG) setting
over cluster of GPUs
Dask Python
• Communication primitives:
– Allreduce UCX-Py Cython mpi4py
– Reduce CUDA/C/C++
– Broadcast
cuML Algorithms
• Exploit optimized collectives UCX
cuML Primitives
MVAPICH2-
NCCL
GDR
CUDA Libraries
CUDA

MPI4cuML Release
• MPI4cuML 0.1 was released in Feb ‘21 adding support for MPI to cuML:
– Can be downloaded from: http://hidl.cse.ohio-state.edu
• Features:
– Based on cuML 0.15
– MVAPICH2 support for C++ and Python APIs
• Included use of cuML C++ CUDA-Aware MPI example for KMeans clustering
• Enabled cuML handles to use MVAPICH2-GDR backend for Python cuML applications
– KMeans, PCA, tSVD, RF, LinearModels
• Added switch between available communication backends (MVAPICH2 and NCCL)
– Built on top of mpi4py over the MVAPICH2-GDR library
– Tested with
• Mellanox InfiniBand adapters (FDR and HDR)
• Various x86-based multi-core platforms (AMD and Intel)
• NVIDIA V100 and P100 GPUs

MPI4cuML Installation
• MPI4cuML is available to download from: http://hidl.cse.ohio-state.edu/

– The userguide is available at: http://hidl.cse.ohio-state.edu/download/hidl/mpi4cuml/mpi4cuml-userguide.pdf
• Setup Instructions
– Installation Pre-requisites:
• Install the MVAPICH2-GDR Library
• Install the mpi4py Library
– Install MPI4cuML
• Running GPU-based cuML Applications

– Writing the host file
– KMeans on real and synthetic data

Setup Instructions: Pre-requisites
Install the MVAPICH2-GDR Library
Install the mpi4py Library

$ git clone https://github.com/mpi4py/mpi4py.git
$ cd mpi4py
$ edit mpi.cfg file
# MVAPICH2
# ------------
[MVAPICH2]
mpi_dir = /path/to/MVAPICH2-GDR/install/directory
mpicc = %(mpi_dir)s/bin/mpicc
mpicxx = %(mpi_dir)s/bin/mpicxx
include_dirs = %(mpi_dir)s/include
library_dirs = %(mpi_dir)s/lib64
runtime_library_dirs = %(library_dirs)s
$ python setup.py build --mpi=MVAPICH2-GDR

$ pip install .

Setup Instructions: Install MPI4cuML
Install MPI4cuML
$ wget http://hidl.cse.ohio-state.edu/download/hidl/cuml/mpi4cuml-0.1.tar.gz
$ tar -xzvf mpi4cuml-0.1.tar.gz
$ cd mpi4cuml-0.1/
$ conda env update -n mpi4cuml --file=conda/environments/cuml_dev_cuda10.2.yml
$ export LIBRARY_PATH=/path/to/miniconda3/envs/mpi4cuml/lib:$LIBRARY_PATH
$ export LD_LIBRARY_PATH=/path/to/miniconda3/envs/mpi4cuml/lib:$LIBRARY_PATH
$ ./build.sh cpp-mgtests
$ conda list
$ conda list | grep cuml
Build KMeans
$ cmake .. -DCUML_LIBRARY_DIR=/path/to/directory/with/libcuml.so
-DCUML_INCLUDE_DIR=/path/to/directory/with/kmeans/kmeans_c.h
$ make
$ LD_PRELOAD=$MPILIB/lib64/libmpi.so:$CUML_HOME/cpp/build/libcuml++.so

Running MPI4cuML Applications
Writing the host file
$ cd mpi4cuml/cpp/examples/mg_kmeans
$ vim hosts
.. write name of the compute nodes ..
Run KMeans with synthetic data
$ MPILIB/bin/mpirun_rsh -export-all -np 4 -hostfile hosts MV2_USE_CUDA=1

MV2_USE_GDRCOPY=1 MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so
./kmeans_mg_example
Run KMeans with real dataset

.. Download data from https://www.kaggle.com/c/homesite-quote-conversion/data ..
$ unzip all.zip
$ ./prepare_input.py [train_file=train.csv] [test_file=test.csv] [output=output.txt]
$ LD_PRELOAD=$MV2_HOME/lib/libmpi.so:$CUML_HOME/cpp/build/libcuml++.so
$ MV2_HOME/bin/mpirun_rsh --export-all -np 4 -hostfile=hosts MV2_USE_CUDA=1
MV2_USE_GDRCOPY=1 MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so
./kmeans_mg_example -num_rows 260753 -num_cols 298 -input output.txt
Run K-Means at C++ Layer on one node with 1
GPU/node (using MVAPICH2-GDR)
$ sbatch -N 1 run_cuml_kmeans_single.sh
+ time /opt/tutorials/dl-tutorial-21/mv2-gdr/bin/mpirun_rsh -np 1 gpu04 MV2_USE_CUDA=1 MV2_SUPPORT_DL=1 /opt/tutorials/dl-tutorial-
21/cuML/mpi4cuml-0.1/cpp/examples/mg_kmeans/kmeans_mg_example -num_rows 260753 -num_cols 298 -input data.txt -k 100 -
max_iterations=3000
Running on 1 GPU(s)
Reading input with 260753 rows and 298 columns from ../data.txt.
Run KMeans with k=100, max_iterations=3000
num_pts inertia
0 2466 1.096572e+11
1 2638 1.245608e+11
2 2739 1.362806e+11
3 2659 1.267454e+11
…
98 2343 1.048909e+11
99 2687 1.310442e+11
Global inertia = 1.246384e+13
real 0m31.455s
user 0m0.019s
sys 0m0.008s

Run K-Means at C++ Layer on two nodes with
1 GPU/node (using MVAPICH2-GDR)
$ sbatch -N 1 run_cuml_kmeans_single.sh
+ time /opt/tutorials/dl-tutorial-21/mv2-gdr/bin/mpirun_rsh -np 2 gpu05 gpu04 MV2_USE_CUDA=1 MV2_SUPPORT_DL=1 /opt/tutorials/dl-tutorial-
21/cuML/mpi4cuml-0.1/cpp/examples/mg_kmeans/kmeans_mg_example -num_rows 260753 -num_cols 298 -input data.txt -k 100 -
max_iterations=3000
Running on 2 GPU(s)
Reading input with 260753 rows and 298 columns from ../data.txt.
Run KMeans with k=100, max_iterations=6000
num_pts inertia
0 2466 1.096572e+11
1 2638 1.245608e+11
2 2739 1.362806e+11
3 2659 1.267454e+11
…
98 2343 1.048909e+11
99 2687 1.310442e+11
Global inertia = 1.246384e+13
real 0m33.078s
user 0m0.035s
sys 0m0.017s

2000
K-Means 2.0 2000
Linear Regression 1.6
NCCL NCCL
1.8 1.4
MVAPICH2-GDR 1.6 MVAPICH2-GDR
Training Time (s)
Training Time (s)

1500 1500 1.2
Speedup 1.4 Speedup
1.0
Speedup
Speedup
1.2
1000 1.0 1000 0.8
0.8 0.6
0.6
500 500 0.4
0.4
0.2 0.2
0 0.0 0 0.0
1 2 4 8 16 32 1 2 4 8 16 32
Nearest Neighbors Truncated SVD

2500 2.0
NCCL 1500 1.6
1.8 NCCL
2000 MVAPICH2-GDR 1.6 1.4
Training Time (s)
MVAPICH2-GDR
Speedup
Training Time (s)

1.4 Speedup 1.2
Speedup
1500 1.2 1000
Speedup
1.0
1.0
0.8
1000 0.8
0.6
0.6 500
0.4
500 0.4
0.2 0.2
0 0.0 0 0.0
1 2 4 8 16 32 1 2 4 8 16 32
M. Ghazimirsaeed , Q. Anthony , A. Shafi , H. Subramoni , and D. K. Panda, Accelerating GPU-based Machine Learning in Python
using MPI Library: A Case Study with MVAPICH2-GDR, MLHPC Workshop, Nov 2020
Outline
• Introduction
• Conclusion

Introduction to Dask
• Dask is a popular task-based distributed computing framework:
– Scales Python applications from laptops to high-end systems
– Builds a task-graph that is executed lazily on parallel hardware
– Natively extends popular data processing libraries like numPy, Pandas
• Dask Distributed library supports parallel and distributed execution:
– Built using the asyncio package that allows execution of asynchronous/non-blocking/concurrent
operations called coroutines:
• These are defined using async and invoked using await
– Dask Distributed library currently has two communication backends:

• TCP: Tornado-based
• UCX: Built using a Cython wrapper called UCX-Py
• MPI4Dask: MVAPICH2-based MPI communication device
• Other Data Science frameworks include Apache Spark and Ray

Dask Distributed Execution Model
Client
Cluster
Scheduler
Worker Worker Worker

MPI4Dask: MPI-based Communication Backend for
Dask
• Dask currently has two communication backends:
• TCP: Tornado-based
• UCX: Built using a GPU-aware Cython wrapper called UCX-Py
• MPI4Dask is an MVAPICH2-based communication backend for Dask:

– First MPI-based communication device for Dask
– Optimizes CPU and GPU communication in the Dask framework on modern HPC clusters

Dask Architecture
Dask
Dask Bag Dask Array Dask DataFrame Delayed Future
Task Graph
Dask-MPI Dask-CUDA Dask-Jobqueue
Distributed
Scheduler Worker Client
tcp.py ucx.py MPI4Dask

Comm Layer
UCX-Py mpi4py
(Cython wrappers)
TCP UCX MVAPICH2-GDR
Laptops/
High Performance Computing Hardware
Desktops
MPI4Dask: Bootstrapping and Dynamic Connectivity
• Several ways to start Dask programs:
Client
– Manual 1
– Utility classes: Cluster

Scheduler
• LocalCUDACluster, SLURMCluster, SGECluster, PBCCluster, and others 0
• MPI4Dask uses the Dask-MPI to bootstrap execution of Dask

2 3 4
programs Worker Worker Worker
• Dynamic connectivity is established using the asyncio
package in MPI4Dask:
– Scheduler and workers listen for incoming connections by calling
0 1
asyncio.start_server()
– Workers and client connect using asyncio.open_connection()
2 3
MPI.COMM_WORLD

MPI4Dask: Point-to-point Communication
Coroutines
• Implements communication coroutines for point-to-point MPI functions:
– Using mpi4py (Cython wrappers) and MVAPICH2-GDR
• mpi4py provides two flavors of point-to-point communication functions:
– Send()/Recv() – for exchanging data in buffers (faster and used in MPI4Dask)
– send()/recv() – for communicating Python objects (pickle/unpickle)
– GPU buffers implement the __cuda_array_interface__
• Implemented chunking mechanism for large messages
• The send and receive communication coroutines are as follows:

MPI4Dask Release
• MPI4Dask 0.2 was released in Mar ‘21 adding support for MPI to Dask:
– Can be downloaded from: http://hibd.cse.ohio-state.edu
• Features:
– Based on Dask Distributed 2021.01.0
– Compliant with user-level Dask APIs and packages
– Support for MPI-based communication in Dask for cluster of GPUs
– Implements point-to-point communication co-routines
– Efficient chunking mechanism implemented for large messages
– (NEW) Built on top of mpi4py over the MVAPICH2, MVAPICH2-X, and MVAPICH2-GDR libraries
– (NEW) Support for MPI-based communication for CPU-based Dask applications
– Supports starting execution of Dask programs using Dask-MPI
– Tested with
• (NEW) CPU-based Dask applications using numPy and Pandas data frames
• (NEW) GPU-based Dask applications using cuPy and cuDF
• Mellanox InfiniBand adapters (FDR and EDR)
• Various multi-core platforms
• NVIDIA V100 and Quadro RTX 5000 GPUs
MPI4Dask Installation
• MPI4Dask is available to download from: http://hibd.cse.ohio-state.edu/
– The userguide is available at: http://hibd.cse.ohio-state.edu/static/media/hibd/dask/mpi4dask-0.2-userguide.pdf
• Section 3: Setup Instructions
– 3.1 Installation Pre-requisites:
• 3.1.1 Install Miniconda
• 3.1.2 Modules/Libraries for GPU-based Dask Applications:
• 3.1.3 Modules/Libraries for CPU-based Dask Applications:
• 3.1.4 Install the MVAPICH2 Library (MVAPICH2-X, MVAPICH2, or MVAPICH2-GDR)
• 3.1.5 Install the mpi4py Library
• 3.1.6 Install Dask-MPI package
– 3.2 Install MPI4Dask
• Section 4. Running GPU-based Dask Applications
– 4.1 Writing the host file
– 4.2 Sum of cuPy Array and its Transpose
– 4.3 cuDF Merge
• Section 5. Running GPU-based Dask Applications
– 5.1 Writing the host file
– 5.2 Sum of numPy Array and its Transpose
– 5.3 Sum of Pandas Data Frame
– 5.4 SVD
Installation Pre-requisites
Modules/Libraries for GPU-based Dask Applications:
$ conda install -c conda-forge -c rapidsai -c nvidia automake make libtool pkg-config libhwloc psutil python=3.8 setuptools cython cudatoolkit=10.2 cupy
dask-cudf dask==2021.1.1 distributed numpy rmm
Modules/Libraries for CPU-based Dask Applications:

$ conda install -c conda-forge -c rapidsai -c nvidia automake make libtool pkg-config libhwloc psutil python=3.8 setuptools cython dask==2021.1.1
distributed=2021.1.1 numpy
Install the MVAPICH2 Library (MVAPICH2-X, MVAPICH2, or MVAPICH2-GDR)
Install the mpi4py Library

$ git clone https://github.com/mpi4py/mpi4py.git
$ edit mpi.cfg file
[MVAPICH2]
mpi_dir = /path/to/MVAPICH2-GDR/install/directory
mpicc = %(mpi_dir)s/bin/mpicc
mpicxx = %(mpi_dir)s/bin/mpicxx
include_dirs = %(mpi_dir)s/include
library_dirs = %(mpi_dir)s/lib64
runtime_library_dirs = %(library_dirs)s
$ python setup.py build --mpi=MVAPICH2-GDR; $ pip install .

Install Dask-MPI and MPI4Dask
Install Dask-MPI package
$ git clone https://github.com/dask/dask-mpi.git
$ cd dask-mpi
$ python setup.py build
$ pip install .
Install MPI4Dask
$ wget http://hibd.cse.ohio-state.edu/download/hibd/dask/mpi4dask-0.2.tar.gz
$ tar -xzvf mpi4dask-0.2.tar.gz
$ cd mpi4dask-0.2/distributed
$ python setup.py build
$ pip install .
$ conda list
$ conda list | grep distributed
$ conda list |grep dask

Running GPU and CPU Dask Applications
Sum of cuPy Array and its Transpose [GPU]
$ cd mpi4dask-0.2/dask-apps/gpu
$ LD_PRELOAD=$MPILIB/lib64/libmpi.so $MPILIB/bin/mpirun_rsh -export-all -np 4 -hostfile hosts MV2_USE_CUDA=1
MV2_USE_GDRCOPY=1 MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so MV2_CPU_BINDING_LEVEL=SOCKET
MV2_CPU_BINDING_POLICY=SCATTER python cupy_sum_mpi.py
cuDF Merge [GPU]

$ cd mpi4dask-0.2/dask-apps/gpu
$ LD_PRELOAD=$MPILIB/lib/libmpi.so $MPILIB/bin/mpirun_rsh -export-all -np 4 -hostfile hosts MV2_USE_CUDA=1
MV2_USE_GDRCOPY=1 MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so MV2_CPU_BINDING_LEVEL=SOCKET
MV2_CPU_BINDING_POLICY=SCATTER python cudf_merge_mpi.py --type gpu --protocol mpi --runs 20 --chunk-size 1_000_000_00
Sum of numPy Array and its Transpose [CPU]

$ cd mpi4dask-0.2/dask-apps/cpu
$ LD_PRELOAD=$MPILIB/lib/libmpi.so $MPILIB/bin/mpirun_rsh -export-all -np 4 -hostfile hosts MV2_CPU_BINDING_LEVEL=SOCKET
MV2_CPU_BINDING_POLICY=SCATTER python numpy_sum_mpi.py
Sum of Pandas DataFrame [CPU]

$ cd mpi4dask-0.2/dask-apps/cpu
$ LD_PRELOAD=$MPILIB/lib/libmpi.so $MPILIB/bin/mpirun_rsh -export-all -np 4 -hostfile hosts MV2_CPU_BINDING_LEVEL=SOCKET
MV2_CPU_BINDING_POLICY=SCATTER python dask-cudf_sum_mpi.py
Sum of cuPy Array and its Transpose (with TCP device)
$ srun -N 4 --reservation=dltutorial run_cupy_sum_tcp.sh
Sample Output:
+ /opt/tutorials/dl-tutorial-21/dask/mvapich2/install/bin/mpirun_rsh -np 4 -hostfile
/tmp/shafi.16/hosts LD_PRELOAD=/opt/tutorials/dl-tutorial-
21/dask/mvapich2/install/lib/libmpi.so MV2_USE_CUDA=1 MV2_CPU_BINDING_LEVEL=SOCKET
MV2_CPU_BINDING_POLICY=SCATTER MV2_USE_GDRCOPY=1
MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so
MV2_SUPPRESS_JOB_STARTUP_PERFORMANCE_WARNING=1 /opt/tutorials/dl-tutorial-
21/dask/miniconda3/envs/mpi4dask_tutorial/bin/python /opt/tutorials/dl-tutorial-
21/labs/lab1/cupy_sum_tcp.py
<Client: 'tcp://10.3.1.1:46566' processes=2 threads=56, memory=269.79 GB>
Time for iteration 0 : 15.613967895507812
Median Wall-clock Time: 13.07 s
+ set +x

Sum of cuPy Array and its Transpose (with MPI device)
$ srun -N 4 --reservation=dltutorial run_cupy_sum_mpi.sh
Sample Output:
+ /opt/tutorials/dl-tutorial-21/dask/mvapich2/install/bin/mpirun_rsh -np 4 -hostfile
/tmp/shafi.16/hosts LD_PRELOAD=/opt/tutorials/dl-tutorial-
21/dask/mvapich2/install/lib/libmpi.so MV2_USE_CUDA=1 MV2_CPU_BINDING_LEVEL=SOCKET
MV2_CPU_BINDING_POLICY=SCATTER MV2_USE_GDRCOPY=1
MV2_GPUDIRECT_GDRCOPY_LIB=/opt/gdrcopy2.0/lib64/libgdrapi.so
MV2_SUPPRESS_JOB_STARTUP_PERFORMANCE_WARNING=1 /opt/tutorials/dl-tutorial-
21/dask/miniconda3/envs/mpi4dask_tutorial/bin/python /opt/tutorials/dl-tutorial-
21/labs/lab1/cupy_sum_mpi.py
<Client: 'mpi://10.3.1.1:31549' processes=2 threads=56, memory=269.79 GB>
Time for iteration 1 : 5.288544416427612 Configurable options in the script (cupy_sum_mpi.py)
Time for iteration 3 : 4.088623523712158 DASK_PROTOCOL = 'mpi’ # Options are ['mpi', 'tcp']
Time for iteration 4 : 4.115798234939575 GPUS_PER_NODE = 1 # number of GPUs in the system
Median Wall-clock Time: 4.12 s RUNS = 5 # repetitions for the benchmark
+ set +x DASK_INTERFACE = 'ib0' # interface to use for communication
THREADS_PER_NODE = 28 # number of threads per node.
MVAPICH2-GDR: 3.17x faster

Benchmark #1: Sum of cuPy Array and its Transpose
(RI2) 3.47x better on average 6.92x better on average
10 2.0
IPoIB UCX MPI4Dask IPoIB UCX MPI4Dask
9 1.8
8 1.6
Communication Time (s)

Total Execution Time (s)
7 1.4
6 1.2
5 1.0
4 0.8
3 0.6
2 0.4
1 0.2
0 0.0
2 3 4 5 6 2 3 4 5 6
Number of Dask Workers Number of Dask Workers
A. Shafi , J. Hashmi , H. Subramoni , and D. K. Panda, Efficient MPI-based MPI4Dask 0.2 release
Communication for GPU-Accelerated Dask Applications, CCGrid ’21,
https://arxiv.org/abs/2101.08878 (http://hibd.cse.ohio-state.edu)
Benchmark #1: Sum of cuPy Array and its Transpose
(TACC Frontera GPU Subsystem) 1.71x better on average
A. Shafi , J. Hashmi , H. Subramoni , and D. K. Panda, Efficient MPI-based MPI4Dask 0.2 release
Communication for GPU-Accelerated Dask Applications, CCGrid ‘21
Benchmark #2: cuDF Merge (TACC Frontera GPU
Subsystem)
2.91x better on average 2.90x better on average
A. Shafi , J. Hashmi , H. Subramoni , and D. K. Panda, Efficient MPI-based

MPI4Dask 0.2 release
Communication for GPU-Accelerated Dask Applications, CCGrid ‘21
Outline
• Introduction
• Conclusion

Convergence of ML/DL, Data Science, and HPC
• Is Machine Learning/Deep Learning and Data Science an HPC Problem?
– Distributed Model/DNN Training is definitely an HPC problem
– Inference – not yet an HPC problem
– Support for Data Science frameworks on HPC systems is improving (yet lagging)
• Why HPC can help?
– Decades of research for communication models and performance optimizations
– MPI, PGAS, and other communication runtimes can help for “data-parallel” training
• Some of the needs for ML/DL and Data Science frameworks are an exact match
– Compute intensive problem
• Some needs are new for distributed/parallel communication runtimes
– Large Message Communication
– CUDA-Aware Communication

Exploiting GPUs and Unified Communication Layer
for Data Science Frameworks
• The support for GPUs has been added to Dask as part of the NVIDIA RAPIDS
project
• The RAPIDS Accelerator for Apache Spark is currently adding support for GPU
processing for the Apache Spark Framework
– SQL Plugin for executing SQL queries entirely on GPUs
– Shuffle Plugin based on the UCX communication library (GPU-to-GPU)
– https://nvidia.github.io/spark-rapids/
• Opportunity to utilize GPU-to-GPU communication provided by MPI libraries like
MVAPICH2-GDR
• Is it possible to have an MPI-based communication layer for Data Science
frameworks?

Conclusion
• Exponential growth in Machine Learning/Deep Learning/Data Science
frameworks
• Provided an overview of issues, challenges, and opportunities for
designing efficient communication runtimes
– Efficient, scalable, and hierarchical designs are crucial for ML/DL/Data Science frameworks
– Co-design of communication runtimes and ML/DL/Data Science frameworks will be essential
• Presented use-cases to demonstrate the complex interaction between DL/ML/Dask
middleware with the underling HPC technologies and middleware
• Need collaborative efforts to achieve the full potential

Funding Acknowledgments
Funding Support by
Equipment Support by

Acknowledgments to all the Heroes (Past/Current Students and Staffs)
Current Students (Graduate) Current Research Scientists Current Software Engineers
– N. Alnaasan (Ph.D.) – K. S. Khorassani (Ph.D.) – A. H. Tu (Ph.D.)
– A. Shafi – B. Seeds
– Q. Anthony (Ph.D.) – P. Kousha (Ph.D.) – S. Xu (Ph.D.)
– H. Subramoni – N. Shineman
– C.-C. Chun (Ph.D.) – B. Michalowicz (Ph.D.) – Q. Zhou (Ph.D.)
– K. Al Attar (M.S.)
Current Students (Undergrads)
– N. Contini (Ph.D.) – B. Ramesh (Ph.D.)
– N. Sarkauskas (Ph.D.) – M. Lieber
– A. Jain (Ph.D.) – K. K. Suresh (Ph.D.)
– L. Xu
Past Students
– A. Awan (Ph.D.) – T. Gangadharappa (M.S.) – P. Lai (M.S.) – D. Shankar (Ph.D.)
– A. Augustine (M.S.) – K. Gopalakrishnan (M.S.) – J. Liu (Ph.D.) – G. Santhanaraman (Ph.D.) Past Research Scientists
– P. Balaji (Ph.D.) – J. Hashmi (Ph.D.) – M. Luo (Ph.D.) – N. Sarkauskas (B.S.) – K. Hamidouche
– M. Bayatpour (Ph.D.) – W. Huang (Ph.D.) – A. Mamidala (Ph.D.) – N. Senthil Kumar (M.S.) – S. Sur
– R. Biswas (M.S.) – W. Jiang (M.S.) – G. Marsh (M.S.) – A. Singh (Ph.D.) – X. Lu
– S. Bhagvat (M.S.) – J. Jose (Ph.D.) – V. Meshram (M.S.) – J. Sridhar (M.S.)
Past Senior Research Associate
– A. Bhat (M.S.) – M. Kedia (M.S.) – A. Moody (M.S.) – S. Srivastava (M.S.)
– S. Sur (Ph.D.)
– J. Hashmi
– D. Buntinas (Ph.D.) – S. Kini (M.S.) – S. Naravula (Ph.D.)
– L. Chai (Ph.D.) – M. Koop (Ph.D.) – R. Noronha (Ph.D.) – H. Subramoni (Ph.D.) Past Programmers
– B. Chandrasekharan (M.S.) – K. Kulkarni (M.S.) – X. Ouyang (Ph.D.) – K. Vaidyanathan (Ph.D.) – A. Reifsteck
– S. Chakraborthy (Ph.D.) – R. Kumar (M.S.) – S. Pai (M.S.) – A. Vishnu (Ph.D.) – D. Bureddy
– N. Dandapanthula (M.S.) – S. Krishnamoorthy (M.S.) – S. Potluri (Ph.D.) – J. Wu (Ph.D.) – J. Perkins
– V. Dhanraj (M.S.) – K. Kandalla (Ph.D.) – K. Raj (M.S.) – W. Yu (Ph.D.)
Past Research Specialist
– C.-H. Chu (Ph.D.) – M. Li (Ph.D.) – R. Rajachandrasekar (Ph.D.) – J. Zhang (Ph.D.)
– M. Arnold
Past Post-Docs
– D. Banerjee – H.-W. Jin – E. Mancini – A. Ruhela – J. Smith
– X. Besseron – J. Lin – K. Manian – J. Vienne
– M. S. Ghazimeersaeed – M. Luo – S. Marcarelli – H. Wang

Thank You!
{jain.575,shafi.16, anthony.301}@osu.edu
Follow us on
https://twitter.com/mvapich
Network-Based Computing Laboratory

http://nowlab.cse.ohio-state.edu/
The MVAPICH2 Project The High-Performance Deep Learning Project

http://mvapich.cse.ohio-state.edu/ http://hidl.cse.ohio-state.edu/

Mug21 DL ML v3

Uploaded by

Copyright:

Available Formats

Mug21 DL ML v3

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

Mug21 DL ML v3

Uploaded by

Copyright:

Available Formats

Follow us on

Boosting Performance of Deep Learning, Machine

Arpan Jain Aamir Shafi Quentin Anthony

Network Based Computing Laboratory MUG ‘21 2

Network Based Computing Laboratory MUG ‘21 3

Network Based Computing Laboratory MUG ‘21 4

• Computation requirement is increasing exponentially

Network Based Computing Laboratory MUG ‘21 6

Network Based Computing Laboratory MUG ‘21 7

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 8

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 9

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 10

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 11

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 12

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 13

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 14

Input Hidden Hidden Output

Network Based Computing Laboratory MUG ‘21 15

Network Based Computing Laboratory MUG ‘21 16

• What are Activation functions? Courtesy: https://www.jeremyjordan.me/neural-networks-training/

– RELU (a Max fn.) is the most common activation fn.

Network Based Computing Laboratory MUG ‘21 18

• Stochastic Gradient Descent

• Finding the optimal batch

Network Based Computing Laboratory MUG ‘21 19

• Challenge: How to make “Training” faster?

Network Based Computing Laboratory MUG ‘21 20

Network Based Computing Laboratory MUG ‘21 21

Hybrid (Model and Data) Parallelism

Network Based Computing Laboratory MUG ‘21 23

Network Based Computing Laboratory MUG ‘21 25

Input-only Parameters Sendbuf (Before)

Network Based Computing Laboratory MUG ‘21 27

Network Based Computing Laboratory MUG ‘21 28

ML/DL Applications ML/DL Applications

TensorFlow PyTorch MXNet PyTorch

Horovod Torch.distributed DeepSpeed

MVAPICH2 or MVAPICH2-X MVAPICH2-GDR for MVAPICH2 or MVAPICH2-X MVAPICH2-GDR for

More details available from: http://hidl.cse.ohio-state.edu

Network Based Computing Laboratory MUG ‘21 30

– Introduced architecture detection, point-to-point tuning and collective tuning for

Network Based Computing Laboratory MUG ‘21 31

Command to install Horovod with MVAPICH2-X

$ HOROVOD_WITH_MPI=1 pip install --no-cache-dir horovod==0.22.1

Command to install Horovod with MVAPICH2-GDR

$ HOROVOD_GPU_ALLREDUCE=MPI HOROVOD_CUDA_HOME=/opt/cuda/10.1 HOROVOD_WITH_MPI=1 pip

Network Based Computing Laboratory MUG ‘21 32

+ python pytorch_synthetic_benchmark.py --batch-size 64 --num-iters=5

Network Based Computing Laboratory MUG ‘21 33

Network Based Computing Laboratory MUG ‘21 34

Platform: Nvidia DGX-2 system, CUDA 9.2

Scaling Efficiency (%)