How to Achieve High-Performance, Scalable and Distributed DNN Training on Modern HPC Systems?

How to Achieve High-Performance, Scalable and
Distributed Training on Modern HPC Systems?
Dhabaleswar K. (DK) Panda
The Ohio State University
E-mail: panda@cse.ohio-state.edu
http://www.cse.ohio-state.edu/~panda
Talk at HPCAC-AI Stanford Conference (April ’20)
by
Follow us on
https://twitter.com/mvapich

HPCAC-AI-Stanford (April ‘20) 2Network Based Computing Laboratory
Understanding the Deep Learning Resurgence
Adopted from: http://www.deeplearningbook.org/contents/intro.html
• Deep Learning (DL) is a sub-set of
Machine Learning (ML)
– Perhaps, the most revolutionary
subset!
– Feature extraction vs. hand-crafted
features
• Deep Learning
– A renewed interest and a lot of hype!
– Key success: Deep Neural Networks
(DNNs)
– Everything was there since the late 80s
except the “computability of DNNs”
AI
Machine
Learning
Deep
Learning
Examples:
MLPs, DNNs,
Examples:
Logistic
Regression

• Modern and efficient hardware enabled
– Computability of DNNs – impossible in
the past!
– GPUs – at the core of DNN training
– CPUs – catching up fast
• Availability of Datasets
– MNIST, CIFAR10, ImageNet, and more…
• Excellent Accuracy for many application
areas
– Vision, Machine Translation, and several
others...
Deep Learning in the Many-core Era
Courtesy: A. Canziani et al., “An Analysis of Deep Neural Network Models for Practical Applications”, CoRR, 2016.
0
2000
4000
6000
8000
10000
2 GTX 580 DGX-2
MinutestoTrain
AlexNet
~500X in 5 years

Big Data
(Hadoop, Spark,
HBase,
Memcached,
etc.)
Deep Learning
(Caffe, TensorFlow, BigDL,
etc.)
HPC
(MPI, RDMA,
Lustre, etc.)
Increasing Usage of HPC, Big Data and Deep Learning
Convergence of HPC, Big
Data, and Deep Learning!
Increasing Need to Run these
applications on the Cloud!!

Drivers of Modern HPC Cluster Architectures
• Multi-core/many-core technologies
• Remote Direct Memory Access (RDMA)-enabled networking (InfiniBand and RoCE)
• Solid State Drives (SSDs), Non-Volatile Random-Access Memory (NVRAM), NVMe-SSD
• Accelerators (NVIDIA GPGPUs and Intel Xeon Phi)
• Available on HPC Clouds, e.g., Amazon EC2, NSF Chameleon, Microsoft Azure, etc.
Accelerators / Coprocessors
high compute density, high
performance/watt
>1 TFlop DP on a chip
High Performance Interconnects -
InfiniBand
<1usec latency, 200Gbps Bandwidth>Multi-core Processors SSD, NVMe-SSD, NVRAM
K - ComputerSunway TaihuLightSummit Sierra

• Deep Learning has two major tasks
1. Training of the Deep Neural Network
2. Inference (or deployment) that uses a trained DNN
• DNN Training
– Training is a compute/communication intensive process – can take days to
weeks
– Faster training is necessary!
• Faster training can be achieved by
– Using Newer and Faster Hardware – But, there is a limit!
– Can we use more GPUs or nodes?
• The need for Parallel and Distributed Training
Key Phases of Deep Learning

• Scale-up: Intra-node Communication
– Many improvements like:
• NVIDIA cuDNN, cuBLAS, NCCL, etc.
• CUDA 9 Co-operative Groups
• Scale-out: Inter-node Communication
– DL Frameworks – most are optimized for
single-node only
– Distributed (Parallel) Training is an emerging
trend
• OSU-Caffe – MPI-based
• Microsoft CNTK – MPI/NCCL2
• Google TensorFlow – gRPC-based/MPI/NCCL2
• Facebook Caffe2 – Hybrid (NCCL2/Gloo/MPI)
Scale-up and Scale-out
Scale-upPerformance
Scale-out Performance
cuDNN
gRPC
Hadoop
MPI
MKL-DNN
Desired
NCCL2

Holistic Evaluation is Important!!
• My framework is faster than
your framework!
• This needs to be understood
in a holistic way.
• Performance depends on
the entire execution
environment (the full stack)
• Isolated view of
performance is not helpful
A. A. Awan, H. Subramoni, and Dhabaleswar K. Panda. “An In-depth Performance Characterization of CPU- and GPU-based DNN Training on
Modern Architectures”, In Proceedings of the Machine Learning on HPC Environments (MLHPC'17). ACM, New York, NY, USA, Article 8.
MKL/
MKL-DNN

How to efficiently scale-out a
Deep Learning (DL) framework and take
advantage of heterogeneous
High Performance Computing (HPC)
resources?
Broad Challenge: Exploiting HPC for Deep Learning

1. What are the fundamental
issues in designing DL
frameworks?
– Memory Requirements
– Computation
Requirements
– Communication Overhead
2. Why do we need to support
distributed training?
– To overcome the limits of
single-node training
– To better utilize hundreds
of existing HPC Clusters
Research Challenges to Exploit HPC Technologies
InfiniBand GPUCPU
CNTK
Gradient
Aggregation
Model Propagation
Forward
Backward
Deep Learning and Machine Learning Frameworks
Caffe/
OSU-Caffe
Caffe2 TensorFlow MXNet
Communication Runtimes to support
Distributed Training
HPC Platforms
Major Computation and Communication Phases in DL Frameworks
1
2

3. What are the new design challenges
brought forward by DL frameworks for
Communication runtimes?
– Large Message Collective
Communication and Reductions
– GPU Buffers (CUDA-Awareness)
4. Can a Co-design approach help in
achieving Scale-up and Scale-out efficiently?
– Co-Design the support at Runtime
level and Exploit it at the DL
Framework level
– What performance benefits can
be observed?
– What needs to be fixed at the
communication runtime layer?
Research Challenges to Exploit HPC Technologies (Cont’d)
CUDA-
Awareness
InfiniBand GPUCPU
Large-message
Collectives
CNTK
Point-to-
Point
Operations
Gradient
Aggregation
Model Propagation
Forward
Backward
Deep Learning and Machine Learning Frameworks
Caffe/
OSU-Caffe
Caffe2 TensorFlow MXNet
Communication Runtimes (MPI/NCCL/Gloo/MLSL)
HPC Platforms
Major Computation and Communication Phases in DL Frameworks
3
4 Co-Design
Opportunities

• MPI-driven Deep Learning
– CPU-based Deep Learning
– GPU-based Deep Learning
• Co-designing Deep Learning Stacks with High-Performance MPI
• Out-of-core DNN training
• Exploiting Hybrid (Data and Model) Parallelism
• Use-Case: AI-Driven Digital Pathology
Multiple Approaches taken up by OSU

Data Parallel Deep Learning and MPI Collectives
MPI_Bcast (GPU 0)
packed_comm_buff
L1
L2
..
Ln
F
L1
L2
..
Ln
L1
L2
..
Ln
L1
L2
..
Ln
Params
GPU0
Params
GPU1
Params
GPU2
Params
GPU3
Gradients
1. Data
Propagation
2. Forward
Backward
Pass
3. Gradient
Aggregatio
n
B F B F B F B
packed_red
uce_buff
packed_red
uce_buff
packed_red
uce_buff
packed_red
uce_buff
ApplyUpdates
MPI_Reduce (GPU 0)
Loop {}
• Major MPI Collectives
involved in Designing
distributed frameworks
• MPI_Bcast – required for
DNN parameter exchange
• MPI_Reduce – needed for
gradient accumulation
from multiple solvers
• MPI_Allreduce – use just
one Allreduce instead of
Reduce and Broadcast
A. A. Awan, K. Hamidouche, J. M. Hashmi, and D. K. Panda, S-Caffe: Co-designing MPI Runtimes and Caffe for Scalable Deep Learning on Modern GPU
Clusters. In Proceedings of the 22nd ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP '17)

Overview of the MVAPICH2 Project
• High Performance open-source MPI Library for InfiniBand, Omni-Path, Ethernet/iWARP, and RDMA over Converged Ethernet (RoCE)
– MVAPICH (MPI-1), MVAPICH2 (MPI-2.2 and MPI-3.1), Started in 2001, First version available in 2002
– MVAPICH2-X (MPI + PGAS), Available since 2011
– Support for GPGPUs (MVAPICH2-GDR) and MIC (MVAPICH2-MIC), Available since 2014
– Support for Virtualization (MVAPICH2-Virt), Available since 2015
– Support for Energy-Awareness (MVAPICH2-EA), Available since 2015
– Support for InfiniBand Network Analysis and Monitoring (OSU INAM) since 2015
– Used by more than 3,075 organizations in 89 countries
– More than 732,000 (> 0.7 million) downloads from the OSU site directly
– Empowering many TOP500 clusters (November ‘19 ranking)
• 3rd ranked 10,649,640-core cluster (Sunway TaihuLight) at NSC, Wuxi, China
• 5th, 448, 448 cores (Frontera) at TACC
• 8th, 391,680 cores (ABCI) in Japan
• 14th, 570,020 cores (Nurion) in South Korea and many others
– Available with software stacks of many vendors and Linux Distros (RedHat, SuSE, OpenHPC, and Spack)
– http://mvapich.cse.ohio-state.edu
• Empowering Top500 systems for over a decade Partner in the 5th ranked TACC Frontera System

0
100000
200000
300000
400000
500000
600000
700000
800000
Sep-04
Mar-05
Sep-05
Mar-06
Sep-06
Mar-07
Sep-07
Mar-08
Sep-08
Mar-09
Sep-09
Mar-10
Sep-10
Mar-11
Sep-11
Mar-12
Sep-12
Mar-13
Sep-13
Mar-14
Sep-14
Mar-15
Sep-15
Mar-16
Sep-16
Mar-17
Sep-17
Mar-18
Sep-18
Mar-19
Sep-19
NumberofDownloads
Timeline
MV0.9.4
MV20.9.0
MV20.9.8
MV21.0
MV1.0
MV21.0.3
MV1.1
MV21.4
MV21.5
MV21.6
MV21.7
MV21.8
MV21.9
MV2-GDR2.0b
MV2-MIC2.0
MV2-GDR2.3.2
MV2-X2.3rc3
MV2Virt2.2
MV22.3.3
OSUINAM0.9.5
MV2-Azure2.3.2
MV2-AWS2.3
MVAPICH2 Release Timeline and Downloads

Architecture of MVAPICH2 Software Family (HPC and DL)
High Performance Parallel Programming Models
Message Passing Interface
(MPI)
PGAS
(UPC, OpenSHMEM, CAF, UPC++)
Hybrid --- MPI + X
(MPI + PGAS + OpenMP/Cilk)
High Performance and Scalable Communication Runtime
Diverse APIs and Mechanisms
Point-to-
point
Primitives
Collectives
Algorithms
Energy-
Awareness
Remote
Memory
Access
I/O and
File Systems
Fault
Tolerance
Virtualization
Active
Messages
Job Startup
Introspection
& Analysis
Support for Modern Networking Technology
(InfiniBand, iWARP, RoCE, Omni-Path, Elastic Fabric Adapter)
Support for Modern Multi-/Many-core Architectures
(Intel-Xeon, OpenPOWER, Xeon-Phi, ARM, NVIDIA GPGPU)
Transport Protocols Modern Features
RC SRD UD DC UMR ODP
SR-
IOV
Multi
Rail
Transport Mechanisms
Shared
Memory
CMA IVSHMEM
Modern Features
Optane* NVLink CAPI*
* Upcoming
XPMEM

• gRPC
– Officially available and supported
– Open-source – can be enhanced by others
– Accelerated gRPC (add RDMA to gRPC)
• gRPC+X
– Use gRPC for bootstrap and rendezvous
– Actual communication is in “X”
– X MPI, Verbs, GPUDirect RDMA (GDR), etc.
• No-gRPC
– Baidu – the first one to use MPI Collectives for TF
– Horovod – Use NCCL, or MPI, or any other future library (e.g. IBM DDL support recently added)
Data Parallel Training with TensorFlow
*Awan et al., “Scalable Distributed DNN Training using TensorFlow and CUDA-Aware MPI: Characterization, Designs, and Performance Evaluation ”, CCGrid ’19.

MVAPICH2 (MPI)-driven Infrastructure for ML/DL Training
MVAPICH2-X for
CPU-Based Training
MVAPICH2-GDR for
GPU-Based Training
Horovod
TensorFlow PyTorch MXNet
ML/DL Applications

MVAPICH2 Software Family (CPU-Based Deep Learning)
High-Performance Parallel Programming Libraries
MVAPICH2 Support for InfiniBand, Omni-Path, Ethernet/iWARP, and RoCE
MVAPICH2-X Advanced MPI features, OSU INAM, PGAS (OpenSHMEM, UPC, UPC++, and CAF), and
MPI+PGAS programming models with unified communication runtime
MVAPICH2-GDR Optimized MPI for clusters with NVIDIA GPUs and for GPU-enabled Deep Learning
Applications
MVAPICH2-Virt High-performance and scalable MPI for hypervisor and container based HPC cloud
MVAPICH2-EA Energy aware and High-performance MPI
MVAPICH2-MIC Optimized MPI for clusters with Intel KNC
Microbenchmarks
OMB Microbenchmarks suite to evaluate MPI and PGAS (OpenSHMEM, UPC, and UPC++)
libraries for CPUs and GPUs
Tools
OSU INAM Network monitoring, profiling, and analysis for clusters with MPI and scheduler
integration
OEMT Utility to measure the energy consumption of MPI applications

Performance of CNTK with MVAPICH2-X on CPU-based Deep Learning
0
200
400
600
800
28 56 112 224
ExecutionTime(s)
No. of Processes
Intel MPI
MVAPICH2
MVAPICH2-XPMEM
CNTK AlexNet Training
(B.S=default, iteration=50, ppn=28)
20%
9%
• CPU-based training of AlexNet neural
network using ImageNet ILSVRC2012
dataset
• Advanced XPMEM-based designs show
up to 20% benefits over Intel MPI (IMPI)
for CNTK DNN training using All_Reduce
• The proposed designs show good
scalability with increasing system size
Designing Efficient Shared Address Space Reduction Collectives for Multi-/Many-cores, J. Hashmi, S. Chakraborty, M. Bayatpour, H.
Subramoni, and DK Panda, 32nd IEEE International Parallel & Distributed Processing Symposium (IPDPS '18), May 2018
Available since MVAPICH2-X 2.3rc1 release

Distributed TensorFlow on TACC Frontera (2,048 CPU nodes)
• Scaled TensorFlow to 2048 nodes on
Frontera using MVAPICH2 and IntelMPI
• MVAPICH2 and IntelMPI give similar
performance for DNN training
• Report a peak of 260,000 images/sec on
2,048 nodes
• On 2048 nodes, ResNet-50 can be trained
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).

Scaling DL Frameworks using MVAPICH2-X on TACC Frontera
• On single node, TensorFlow
(TF) is 8% better than
MXNet
• TF (tf_cnn_benchmark) is
2.13x better than PyTorch
• TensorFlow is 1.7x better
than MXNet
• TF (Keras) gives better
performance compared to
PyTorch and MXNet.
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).

MVAPICH2 Software Family (GPU-Based Deep Learning)
High-Performance Parallel Programming Libraries
MVAPICH2 Support for InfiniBand, Omni-Path, Ethernet/iWARP, and RoCE
MVAPICH2-X Advanced MPI features, OSU INAM, PGAS (OpenSHMEM, UPC, UPC++, and CAF), and
MPI+PGAS programming models with unified communication runtime
MVAPICH2-GDR Optimized MPI for clusters with NVIDIA GPUs and for GPU-enabled Deep Learning
Applications
MVAPICH2-Virt High-performance and scalable MPI for hypervisor and container based HPC cloud
MVAPICH2-EA Energy aware and High-performance MPI
MVAPICH2-MIC Optimized MPI for clusters with Intel KNC
Microbenchmarks
OMB Microbenchmarks suite to evaluate MPI and PGAS (OpenSHMEM, UPC, and UPC++)
libraries for CPUs and GPUs
Tools
OSU INAM Network monitoring, profiling, and analysis for clusters with MPI and scheduler
integration
OEMT Utility to measure the energy consumption of MPI applications

At Sender:
At Receiver:
MPI_Recv(r_devbuf, size, …);
inside
MVAPICH2
• Standard MPI interfaces used for unified data movement
• Takes advantage of Unified Virtual Addressing (>= CUDA 4.0)
• Overlaps data movement from GPU with RDMA transfers
High Performance and High Productivity
MPI_Send(s_devbuf, size, …);
GPU-Aware (CUDA-Aware) MPI Library: MVAPICH2-GPU

CUDA-Aware MPI: MVAPICH2-GDR 1.8-2.3.3 Releases
• Support for MPI communication from NVIDIA GPU device memory
• High performance RDMA-based inter-node point-to-point communication
(GPU-GPU, GPU-Host and Host-GPU)
• High performance intra-node point-to-point communication for multi-GPU
adapters/node (GPU-GPU, GPU-Host and Host-GPU)
• Taking advantage of CUDA IPC (available since CUDA 4.1) in intra-node
communication for multiple GPU adapters/node
• Optimized and tuned collectives for GPU device buffers
• MPI datatype support for point-to-point and collective communication from
GPU device buffers
• Unified memory

0
2000
4000
6000
1
2
4
8
16
32
64
128
256
512
1K
2K
4K
Bandwidth(MB/s)
Message Size (Bytes)
GPU-GPU Inter-node Bi-Bandwidth
MV2-(NO-GDR) MV2-GDR-2.3
0
1000
2000
3000
4000
1
2
4
8
16
32
64
128
256
512
1K
2K
4K
Bandwidth(MB/s)
GPU-GPU Inter-node Bandwidth
MV2-(NO-GDR) MV2-GDR-2.3
0
10
20
30 0
1
2
4
8
16
32
64
128
256
512
1K
2K
4K
8K
Latency(us)
GPU-GPU Inter-node Latency
MV2-(NO-GDR) MV2-GDR 2.3
MVAPICH2-GDR-2.3
Intel Haswell (E5-2687W @ 3.10 GHz) node - 20 cores
NVIDIA Volta V100 GPU
Mellanox Connect-X4 EDR HCA
CUDA 9.0
Mellanox OFED 4.0 with GPU-Direct-RDMA
10x
9x
Optimized MVAPICH2-GDR Design
1.85us
11X

MVAPICH2-GDR vs. NCCL2 – Allreduce Operation (DGX-2)
• Optimized designs in MVAPICH2-GDR offer better/comparable performance for most cases
• MPI_Allreduce (MVAPICH2-GDR) vs. ncclAllreduce (NCCL2) on 1 DGX-2 node (16 Volta GPUs)
1
10
100
1000
10000
Latency(us)
MVAPICH2-GDR-2.3.3 NCCL-2.5
~2.5X better
Platform: Nvidia DGX-2 system (16 Nvidia Volta GPUs connected with NVSwitch), CUDA 9.2
0
5
10
15
20
25
30
35
40
45
50
8
16
32
64
128
256
512
1K
2K
4K
8K
16K
32K
64K
128K
Latency(us)
MVAPICH2-GDR-2.3.3 NCCL-2.5
~4.7X better
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, Accepted to be Presented.

MVAPICH2-GDR: Enhanced MPI_Allreduce at Scale
• Optimized designs in MVAPICH2-GDR offer better performance for most cases
• MPI_Allreduce (MVAPICH2-GDR) vs. ncclAllreduce (NCCL2) up to 1,536 GPUs
0
1
2
3
4
5
6
32M 64M 128M 256M
Bandwidth(GB/s)
Bandwidth on 1,536 GPUs
MVAPICH2-GDR-2.3.3 NCCL 2.5
1.7X better
0
50
100
150
200
250
300
350
400
450
4 16 64 256 1K 4K 16K
Latency(us)
Latency on 1,536 GPUs
MVAPICH2-GDR-2.3.3 NCCL 2.5
1.6X better
Platform: Dual-socket IBM POWER9 CPU, 6 NVIDIA Volta V100 GPUs, and 2-port InfiniBand EDR Interconnect
0
2
4
6
8
10
24 48 96 192 384 768 1536
Bandwidth(GB/s)
Number of GPUs
128MB Message
SpectrumMPI 10.3 OpenMPI 4.0.1
NCCL 2.5 MVAPICH2-GDR-2.3.3
1.7X better
Systems, " ICS-2020. Accepted to be Presented.

Scalable TensorFlow using Horovod and MVAPICH2-GDR
• ResNet-50 Training using TensorFlow benchmark on 1 DGX-2 node (16 Volta GPUs)
0
1000
2000
3000
4000
5000
6000
7000
1 2 4 8 16
Imagepersecond
Number of GPUs
NCCL-2.5 MVAPICH2-GDR-2.3.3
9% higher
Platform: Nvidia DGX-2 system, CUDA 9.2
0
20
40
60
80
100
1 2 4 8 16
ScalingEfficiency(%)
Number of GPUs
NCCL-2.5 MVAPICH2-GDR-2.3.3
Scaling Efficiency =
Actual throughput
Ideal throughput at scale
× 100%
Systems, " ICS-2020. Accepted to be Presented.

Distributed TensorFlow on ORNL Summit (1,536 GPUs)
• ResNet-50 Training using
TensorFlow benchmark on
SUMMIT -- 1536 Volta
GPUs!
• 1,281,167 (1.2 mil.) images
• Time/epoch = 3.6 seconds
• Total Time (90 epochs)
= 3.6 x 90 = 332 seconds =
5.5 minutes!
0
50
100
150
200
250
300
350
400
1 2 4 6 12 24 48 96 192 384 768 1536
Imagepersecond
Thousands
Number of GPUs
NCCL-2.5 MVAPICH2-GDR 2.3.3
Platform: The Summit Supercomputer (#1 on Top500.org) – 6 NVIDIA Volta GPUs per node connected with NVLink, CUDA 9.2
*We observed errors for NCCL2 beyond 96 GPUs
MVAPICH2-GDR reaching ~0.35 million
images per second for ImageNet-1k!
ImageNet-1k has 1.2 million images

• LLNL/Lassen  POWER9 CPU
with 4 V100 GPUs/node
• PyTorch is becoming a very
important DL framework
• Scaling PyTorch models with
Horovod is simple
• MVAPICH2-GDR provides
better performance and
scalability compared to NCCL2
Scaling PyTorch on LLNL/Lassen using MVAPICH2-GDR
0
5000
10000
15000
20000
25000
30000
35000
40000
45000
1 2 4 8 16 32 64 128
Images/sec(higherisbetter)
No. of GPUs
MVAPICH2-GDR NCCL2
Platform: The Lassen Supercomputer (#10 on Top500.org) – 4 NVIDIA Volta GPUs per node connected with NVLink, CUDA 10.1

• RTX 5000 are NVIDIA’s GPUs targeted
for data centers
• Different from GTX series
– Supports GPUDirect RDMA (GDR)
– Supports GDRCOPY
• MVAPICH2-GDR offers good
performance and reasonable scaling
• Scaling is not as good as Lassen because
– Nodes are connected with IB FDR
– No NVLink between GPUs (only PCIe)
Early Exploration of MXNet using MVAPICH2-GDR on TACC
Frontera RTX GPU Nodes
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
1 2 4 8 16 32 64
Images/second(higherisbetter)
No. of GPUs
MVAPICH2-GDR
Platform: TACC Frontera-RTX – 4 NVIDIA RTX 5000 GPUs per node, CUDA 10.1

• TACC Longhorn is like the
LLNL/Lassen system (POWER9
with 4 V100 GPUs/node)
• Some restrictions to keep in
mind
– TensorFlow 2.1 on POWER9
only built with CUDA 10.2
• Early results are promising
– Good scaling up to 256 GPUs
Early Exploration of TensorFlow 2.1 on TACC/Longhorn using MVAPICH2-GDR
0
10000
20000
30000
40000
50000
60000
70000
1 2 4 8 16 32 64 128 256
Images/sec(higherisbetter) No. of GPUs
MVAPICH2-GDR
Platform: The Longhorn System at TACC – 4 NVIDIA Volta GPUs per node connected with NVLink, CUDA 10.2

• MPI-driven Deep Learning
– CPU-based Deep Learning
– GPU-based Deep Learning
• Co-designing Deep Learning Stacks with High-Performance MPI
• Out-of-core DNN training
• Exploiting Hybrid (Data and Model) Parallelism
• Use-Case: AI-Driven Digital Pathology
Multiple Approaches taken up by OSU

• What are the fundamental bottlenecks in the existing DL frameworks?
• How to design a Scalable and Distributed Framework?
– Achieve both Scale-up and Scale-out efficiently
• What are the new requirements and expectations for Communication Runtimes?
• Can a Co-design approach help in achieving better training performance?
– Can a DL framework like Caffe be co-designed with an MPI runtime?
• What is the impact of the co-design?
– What performance benefits can be observed? At what levels?
S-Caffe: Co-designing HPC and DL
A. A. Awan, K. Hamidouche, J.M. Hashmi, DK Panda, “S-Caffe: Co-designing MPI Runtimes
and Caffe for Scalable Deep Learning on Modern GPU Clusters”, ACM PPoPP ’17

• Caffe : A flexible and layered Deep Learning framework.
• Benefits and Weaknesses
– Multi-GPU Training within a single node
– Performance degradation for GPUs across different
sockets
– Limited Scale-out
• OSU-Caffe: MPI-based Parallel Training
– Enable Scale-up (within a node) and Scale-out (across
multi-GPU nodes)
– Scale-out on 64 GPUs for training CIFAR-10 network on
CIFAR-10 dataset
– Scale-out on 128 GPUs for training GoogLeNet network on
ImageNet dataset
OSU-Caffe: Scalable Deep Learning
0
50
100
150
200
250
8 16 32 64 128
TrainingTime(seconds)
No. of GPUs
GoogLeNet (ImageNet) on 128 GPUs
Caffe OSU-Caffe (1024) OSU-Caffe (2048)
Invalid use case
OSU-Caffe publicly available from
http://hidl.cse.ohio-state.edu/

Out-of-core Training via CUDA Unified Memory
• Large DNNs cannot be trained on GPUs due to memory limitation!
– ResNet-50 -- current frameworks only allow small batch sizes
– Next-generation models like Neural Machine Translation
(NMT), AmoebaNet, etc.
• Ridiculously large consists of billions of parameters
– Can we design Out-of-core DNN training support using new
software features in CUDA 8/9 and hardware mechanisms in
Pascal/Volta GPUs?
• General intuition is that managed allocations “will be” slow!
– The proposed OC-Caffe (Out-of-Core Caffe) exploits the
potential of CUDA Unified Memory.
• OC-Caffe-Opt: up to 80% better than Intel-optimized CPU Caffe for
ResNet-50 training on the Volta V100 GPU with CUDA9 and CUDNN7
A. Awan et al., OC-DNN: Exploiting Advanced Unified Memory Capabilities in CUDA 9 and Volta GPUs for Out-of-Core DNN Training, HiPC ’18

• Data-Parallelism– only for models that fit the memory
• Out-of-core models
– Deeper model  Better accuracy but more memory required!
• Model parallelism can work for out-of-core models!
• Key Challenges
– Model Partitioning is difficult for application
programmers
– Finding the right partition (grain) size is hard
– cut at which layer and why?
– Developing a practical system for model-parallelism
• Redesign DL Framework or create additional layers?
• Existing Communication middleware or extensions needed?
HyPar-Flow: Hybrid and Model Parallelism for CPUs
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 (Accepted to be
presented), https://arxiv.org/pdf/1911.05146.pdf

• ResNet-1001 with variable batch size
• Approach:
– 48 model-partitions for 56 cores
– 512 model-replicas for 512 nodes
– Total cores: 48 x 512 = 24,576
• Speedup
– 253X on 256 nodes
– 481X on 512 nodes
• Scaling Efficiency
– 98% up to 256 nodes
– 93.9% for 512 nodes
Out-of-core Training with HyPar-Flow (512 nodes on TACC Frontera)
481x speedup on 512 Intel Xeon Skylake nodes (TACC Frontera)
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 (Accepted to be presented), https://arxiv.org/pdf/1911.05146.pdf

• The field of Pathology (like many other medical disciplines) is moving Digital
• Traditionally, a pathologist reviews a slide and carries out diagnosis based on prior
knowledge and training
• Experience matters
• Can Deep Learning to be used to train thousands of slides and
– Let the computer carry out the diagnosis
– Narrow down the diagnosis and help a pathologist to make a final decision
• Significant benefits in
– Reducing diagnosis time
– Making pathologists productive
– Reducing health care cost
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)
Exploiting Model Parallelism in AI-Driven Digital Pathology
Courtesy: https://blog.kitware.com/digital-slide-
archive-large-image-and-histomicstk-open-source-
informatics-tools-for-management-visualization-and-
analysis-of-digital-histopathology-data/

• Scalable distributed training is getting important
• Requires high-performance middleware designs while exploiting modern
interconnects
• Provided a set of solutions to achieve scalable distributed training
– MPI (MVAPICH2)-driven solution with Horovod for TensorFlow, PyTorch and
MXNet
– Optimized collectives for CPU-based training
– CUDA-aware MPI with optimized collectives for GPU-based training
– Out-of-core training and Hybrid Parallelism
• Will continue to enable the DL community to achieve scalability and
high-performance for their distributed training
Conclusions

• Supported through X-ScaleSolutions (http://x-scalesolutions.com)
• Benefits:
– Help and guidance with installation of the library
– Platform-specific optimizations and tuning
– Timely support for operational issues encountered with the library
– Web portal interface to submit issues and tracking their progress
– Advanced debugging techniques
– Application-specific optimizations and tuning
– Obtaining guidelines on best practices
– Periodic information on major fixes and updates
– Information on major releases
– Help with upgrading to the latest release
– Flexible Service Level Agreements
• Support being provided to Lawrence Livermore National Laboratory (LLNL) and KISTI, Korea
Commercial Support for MVAPICH2, HiBD, and HiDL Libraries

• High-performance solution for distributed training for your complex
AI problems
• Features:
– Integrated package with TensorFlow, PyTorch, MXNet, Horovod, and
MVAPICH2 MPI libraries
– Targeted for both CPU-based and GPU-based Deep Learning Training
– Integrated profiling support across the stacks
– Support for x86 and OpenPOWER platforms
– Support for InfiniBand, RoCE and NVLink Interconnects
– Out-of-the-box optimal performance
– One-click deployment and execution
• Send an e-mail to contactus@x-scalesolutions.com for free trial!!
X-ScaleAI Product

Funding Acknowledgments
Funding Support by
Equipment Support by

Personnel Acknowledgments
Current Students (Graduate)
– A. Awan (Ph.D.)
– M. Bayatpour (Ph.D.)
– C.-H. Chu (Ph.D.)
– J. Hashmi (Ph.D.)
– A. Jain (Ph.D.)
– K. S. Kandadi (Ph.D.)
Past Students
– A. Augustine (M.S.)
– P. Balaji (Ph.D.)
– R. Biswas (M.S.)
– S. Bhagvat (M.S.)
– A. Bhat (M.S.)
– D. Buntinas (Ph.D.)
– L. Chai (Ph.D.)
– B. Chandrasekharan (M.S.)
– S. Chakraborthy (Ph.D.)
– N. Dandapanthula (M.S.)
– V. Dhanraj (M.S.)
– R. Rajachandrasekar (Ph.D.)
– D. Shankar (Ph.D.)
– G. Santhanaraman (Ph.D.)
– A. Singh (Ph.D.)
– J. Sridhar (M.S.)
– S. Sur (Ph.D.)
– H. Subramoni (Ph.D.)
– K. Vaidyanathan (Ph.D.)
– A. Vishnu (Ph.D.)
– J. Wu (Ph.D.)
– W. Yu (Ph.D.)
– J. Zhang (Ph.D.)
Past Research Scientist
– K. Hamidouche
– S. Sur
– X. Lu
Past Post-Docs
– D. Banerjee
– X. Besseron
– H.-W. Jin
– T. Gangadharappa (M.S.)
– K. Gopalakrishnan (M.S.)
– W. Huang (Ph.D.)
– W. Jiang (M.S.)
– J. Jose (Ph.D.)
– S. Kini (M.S.)
– M. Koop (Ph.D.)
– K. Kulkarni (M.S.)
– R. Kumar (M.S.)
– S. Krishnamoorthy (M.S.)
– K. Kandalla (Ph.D.)
– M. Li (Ph.D.)
– P. Lai (M.S.)
– J. Liu (Ph.D.)
– M. Luo (Ph.D.)
– A. Mamidala (Ph.D.)
– G. Marsh (M.S.)
– V. Meshram (M.S.)
– A. Moody (M.S.)
– S. Naravula (Ph.D.)
– R. Noronha (Ph.D.)
– X. Ouyang (Ph.D.)
– S. Pai (M.S.)
– S. Potluri (Ph.D.)
– Kamal Raj (M.S.)
– K. S. Khorassani (Ph.D.)
– P. Kousha (Ph.D.)
– A. Quentin (Ph.D.)
– B. Ramesh (Ph.D.)
– S. Xu (Ph.D.)
– J. Lin
– M. Luo
– E. Mancini
Past Programmers
– D. Bureddy
– J. Perkins
Current Research Specialist
– J. Smith
– S. Marcarelli
– A. Ruhela
– J. Vienne
Current Post-doc
– M. S. Ghazimeersaeed
– K. Manian
Current Students (Undergraduate)
– V. Gangal (B.S.)
– N. Sarkauskas (B.S.)
Past Research Specialist
– M. Arnold
Current Research Scientist
– A. Shafi
– H. Subramoni
– Q. Zhou (Ph.D.)
– H. Wang

Thank You!
Network-Based Computing Laboratory
http://nowlab.cse.ohio-state.edu/
panda@cse.ohio-state.edu
The High-Performance MPI/PGAS Project
http://mvapich.cse.ohio-state.edu/
The High-Performance Deep Learning Project
http://hidl.cse.ohio-state.edu/
The High-Performance Big Data Project
http://hibd.cse.ohio-state.edu/

How to Achieve High-Performance, Scalable and Distributed DNN Training on Modern HPC Systems?

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to How to Achieve High-Performance, Scalable and Distributed DNN Training on Modern HPC Systems?

Similar to How to Achieve High-Performance, Scalable and Distributed DNN Training on Modern HPC Systems? (20)

More from inside-BigData.com

More from inside-BigData.com (20)

Recently uploaded

Recently uploaded (20)

How to Achieve High-Performance, Scalable and Distributed DNN Training on Modern HPC Systems?