HPC Technique Propels Deep Learning at Scale

By Tiffany Trader

February 21, 2017

Researchers from Baidu’s Silicon Valley AI Lab (SVAIL) have adapted a well-known HPC communication technique to boost the speed and scale of their neural network training and now they are sharing their implementation with the larger deep learning community.

The technique, a modified version of the OpenMPI algorithm “ring all-reduce,” is being used at Baidu to parallelize the training of their speech recognition model, Deep Speech 2, across many GPU nodes. The two pieces of software Baidu is announcing today are the baidu-allreduce C library, as well as a patch for TensorFlow, which allows people who have already modeled in TensorFlow to compile this new version and use it for parallelizing across many devices. The codes are available on GitHub.

Ring all-reduce – all GPUs send data simultaneously

Baidu’s SVAIL team developed the approach about two years ago for their internal deep learning framework, named Gene and Majel (in tribute to the famous Star Trek creator and the actress who voiced the onboard computer interfaces for the series). The technique is commonplace in HPC circles, but underused within artificial intelligence and deep learning, according to Baidu.

Many of the researchers in the SVAIL group had come from the high performance computing space and recognized the competitive edge it offered.

“The algorithm is actually part of OpenMPI, but the OpenMPI implementation is not as fast,” comments Baidu Research Scientist Shubho Sengupta. “So the way we stumbled upon it was we started using OpenMPI for doing training and we realized it was not scaling to the extent that we want it to scale. I started digging through the OpenMPI source, found the algorithm, saw that it’s not very efficient, and reimplemented it.”

The SVAIL researchers wrote their own implementation of the ring algorithm for higher performance and better stability. The key distinction from the OpenMPI version is that the SVAIL implementation avoids extraneous copies between the CPU and GPU.

Explains Sengupta, “Once OpenMPI does the communication of these matrices, if the matrices are in GPU memory, it actually copies to CPU memory to do the reduction part of it – that’s actually quite wasteful. You don’t really need to do a copy, you could just write a small kernel that does the reduction in GPU memory space itself. And this especially helps when you are doing all-reduce within a node and all the GPUs are within a PCI root complex, then it doesn’t do any of the copies actually – it can just do everything in GPU memory space. This very simple idea of eliminating this copy resulted in this speedup in scaling over OpenMPI’s own implementation.”

Employing this algorithm along with SVAIL’s focus on fast networking (InfiniBand) and careful hardware-software codesign has enabled the team to get linear GPU scaling up to 128 GPUs, an achievement that was detailed in their December 2015 paper, “Deep Speech 2: End-to-End Speech Recognition in English and Mandarin.”

With their internal implementation of ring all-reduce, the team achieves between a 2.3-21.4X speedup over OpenMPI (version 1.8.5) depending on the number of GPUs.

Sengupta notes that their implementation is fastest for a small number of GPUs. “At 8 GPUs it’s about 20x faster, then as you increase the number of GPUs, it drops because now you actually have to copy data to the CPU to send across the network. But for the internal framework, we can scale all the way up to 128 GPUs and get linear scaling.”

Comparison of two different all-reduce implementations. All times are in seconds. Performance gain is the ratio of OpenMPI all-reduce time to SVAIL’s all-reduce time. (Source: Deep Speech 2 paper)

Sengupta’s teammate Baidu Research Scientist Andrew Gibiansky says similar benefits can now be seen with TensorFlow: “In terms of the TensorFlow implementation, we get the same linear scaling path past eight. In terms of a comparison with running on a single GPU, it ends up being about 31x faster at 40 GPUs.”

After the Deep Speech 2 paper was published, the SVAIL team began getting requests from the community who wanted to know more about the implementation. Given that the algorithm is pretty tightly coupled to SVAIL’s proprietary deep learning framework, they needed to come up with a different way to release it, so they created two new implementations, one specifically for TensorFlow and one that is more general.

Gibiansky, who led the work on the TensorFlow patch, describes their multi-pronged approach to disseminating the information. “You can read the blog post [for a thorough technical explanation] and figure it out. If you’re using TensorFlow, you can use our modification to train your own models with this. And if you’re a deep learning author, you can look at our C library and integrate that. The goal is really to take this idea we’ve found to be really successful internally and try to start spreading it so that other people can also take advantage of it.”

Sengupta shares an interesting perspective on the opportunities to be mined for deep learning within HPC.

“With MPI – people [in deep learning] think that it is this old technology, that it is not relevant, but I think because of our work we have shown that you can build very fast collectives using MPI and that allows you to do synchronous gradient descent which converges faster, gives you deterministic results and you don’t need to do asynchronous gradient descent with parameter servers which was the dominant way of doing this when we first started,” says Sengupta.

As for the reduced-copy approach propagating back to MPI, Gibiansky notes that if you look at some of the other MPI implementations, they’re slowly moving their collectives to GPU versions. “MVPICH recently introduced an all-gather that doesn’t end up copying to CPU – so OpenMPI will probably get there, it just might take a while. Potentially giving this a little more visibility, we can spur that on.”

“There’s a lot of interest now in collectives and one thing we also realized is the all-reduce operation used in traditional HPC setups, it actually transfers data that’s actually not very large,” Sengupta adds. “What it usually does, when I talk to HPC people, it’s trying to figure out the status of something across a bunch of machines – while in deep learning we are transferring these large matrices – like 2048×2048, essentially 4 million 32-bit floating points. For the traditional HPC community, this is a very atypical input for all-reduce. The traditional HPC community does not actually use all-reduce with really large data sizes. I think with deep learning, more and more people are realizing that collective operations for really large matrices is also very important.”

A detailed explanation of ring all-reduce and Baidu’s GPU implementation is covered in this technical blog post, published today by Baidu Research. A variant of the technique is also used to provide high-performance node-local scaling for PaddlePaddle, the company’s open source deep learning framework.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

Nvidia Debuts Turing Architecture, Focusing on Real-Time Ray Tracing

August 16, 2018

From the SIGGRAPH professional graphics conference in Vancouver this week, Nvidia CEO Jensen Huang unveiled Turing, the company's next-gen GPU platform that introduces new RT Cores to accelerate ray tracing and new Tenso Read more…

By Tiffany Trader

HPC Coding: The Power of L(o)osing Control

August 16, 2018

Exascale roadmaps, exascale projects and exascale lobbyists ask, on-again-off-again, for a fundamental rewrite of major code building blocks. Otherwise, so they claim, codes will not scale up. Naturally, some exascale pr Read more…

By Tobias Weinzierl

STAQ(ing) the Quantum Computing Deck

August 16, 2018

Quantum computers – at least for now – remain noisy. That’s another way of saying unreliable and in diverse ways that often depend on the specific quantum technology used. One idea is to mitigate noisiness and perh Read more…

By John Russell

HPE Extreme Performance Solutions

Introducing the First Integrated System Management Software for HPC Clusters from HPE

How do you manage your complex, growing cluster environments? Answer that big challenge with the new HPC cluster management solution: HPE Performance Cluster Manager. Read more…

IBM Accelerated Insights

Super Problem Solving

You might think that tackling the world’s toughest problems is a job only for superheroes, but at special places such as the Oak Ridge National Laboratory, supercomputers are the real heroes. Read more…

NREL ‘Eagle’ Supercomputer to Advance Energy Tech R&D

August 14, 2018

The U.S. Department of Energy (DOE) National Renewable Energy Laboratory (NREL) has contracted with Hewlett Packard Enterprise (HPE) for a new 8-petaflops (peak) supercomputer that will be used to advance early-stage R&a Read more…

By Tiffany Trader

STAQ(ing) the Quantum Computing Deck

August 16, 2018

Quantum computers – at least for now – remain noisy. That’s another way of saying unreliable and in diverse ways that often depend on the specific quantum Read more…

By John Russell

NREL ‘Eagle’ Supercomputer to Advance Energy Tech R&D

August 14, 2018

The U.S. Department of Energy (DOE) National Renewable Energy Laboratory (NREL) has contracted with Hewlett Packard Enterprise (HPE) for a new 8-petaflops (peak Read more…

By Tiffany Trader

CERN Project Sees Orders-of-Magnitude Speedup with AI Approach

August 14, 2018

An award-winning effort at CERN has demonstrated potential to significantly change how the physics based modeling and simulation communities view machine learni Read more…

By Rob Farber

Intel Announces Cooper Lake, Advances AI Strategy

August 9, 2018

Intel's chief datacenter exec Navin Shenoy kicked off the company's Data-Centric Innovation Summit Wednesday, the day-long program devoted to Intel's datacenter Read more…

By Tiffany Trader

SLATE Update: Making Math Libraries Exascale-ready

August 9, 2018

Practically-speaking, achieving exascale computing requires enabling HPC software to effectively use accelerators – mostly GPUs at present – and that remain Read more…

By John Russell

Summertime in Washington: Some Unexpected Advanced Computing News

August 8, 2018

Summertime in Washington DC is known for its heat and humidity. That is why most people get away to either the mountains or the seashore and things slow down. H Read more…

By Alex R. Larzelere

NSF Invests $15 Million in Quantum STAQ

August 7, 2018

Quantum computing development is in full ascent as global backers aim to transcend the limitations of classical computing by leveraging the magical-seeming prop Read more…

By Tiffany Trader

By the Numbers: Cray Would Like Exascale to Be the Icing on the Cake

August 1, 2018

On its earnings call held for investors yesterday, Cray gave an accounting for its latest quarterly financials, offered future guidance and provided an update o Read more…

By Tiffany Trader

Leading Solution Providers

SC17 Booth Video Tours Playlist

Altair @ SC17


AMD @ SC17


ASRock Rack @ SC17

ASRock Rack



DDN Storage @ SC17

DDN Storage

Huawei @ SC17


IBM @ SC17


IBM Power Systems @ SC17

IBM Power Systems

Intel @ SC17


Lenovo @ SC17


Mellanox Technologies @ SC17

Mellanox Technologies

Microsoft @ SC17


Penguin Computing @ SC17

Penguin Computing

Pure Storage @ SC17

Pure Storage

Supericro @ SC17


Tyan @ SC17


Univa @ SC17


  • arrow
  • Click Here for More Headlines
  • arrow
Do NOT follow this link or you will be banned from the site!
Share This