uuudown / tartan

Tartan: Evaluating Modern GPU Interconnect via a Multi-GPU Benchmark Suite

####========================================================================

Author: Ang Li, PNNL

Website: http://www.angliphd.com

Created: 03/19/2018 04:12:50 PM, Richland, WA, USA.

####========================================================================

##Introduction:

Tartan is a multi-GPU benchmark suite. It is proposed to evaluate modern GPU interconnect in our IISWC-18 paper "Tartan: Evaluating Modern GPU Interconnect via a Multi-GPU Benchmark Suite". Please see our paper (as included) for more details.

Tartan contains three sub-directories:

microbenchmark The microbenmarking routines to measure the startup latency, sustainable uni-/bi-direction bandwidth, bandwidth with message size, etc. for Peer-to-Peer (P2P) and and Collective Communication (CL) on PCI-e, NVLink-V1, NVLink-V2 and InfiniBand.

scale-up Applications for intra-node scale-up (i.e., single node with multiple GPUs)

scale-out Applications for inter-node scale-out (i.e., GPU-accelerated multi-node system)

##Acknowledge:

The applications in Tartan were modified from their original implementations. It is an joint effort of the community. Here we list the source of the original design.

Microbenchmark:

• Scale-up P2P: NVIDIA, “CUDA SDK Code Samples”, 2015.
• Scale-up Collective: NVIDIA, “NCCL Tests”, http://github.com/NVIDIA/nccl-tests
• Scale-out P2P: “MPI-GPU-BW”, https://github.com/frroberts/MPI-GPU-BW
• Scale-out Collective: NVIDIA, “NCCL Tests”, http://github.com/NVIDIA/nccl-tests

Scale-up:

• ConvNet2: Google, “High-Performance C++/CUDA Implementation of Convolutional Neural Networks Version-2”, https://github.com/akrizhevsky/cuda-convnet2
• Cusimann: A.M.F. Ferreiro, J.A.G. Rodrıguez, J.G.L. Salas, and C.V. Cendon, “CUSIMANN: An optimized simulated annealing software for GPUs”, https://github.com/palmalcheg/cusimann
• GMM: A.D. Pangborn, “Expectation Maximization with a Gaussian Mixture Model using CUDA”, https://github.com/corv/cuda-gmm-multigpu
• Kmeans: NVIDIA, “Kmeans Clustering with Multi-GPU Capabilities”, https://github.com/NVIDIA/kmeans
• MonteCarlo: NVIDIA, “CUDA SDK Code Samples,” 2015.
• Planar: B. Dimitrov, “Multi-GPU Code to Count all PLANAR Langford Sequences”, https://github.com/boris-dimitrov/z4_planar_langford_multigpu
• Trueke: C. Navarro, “Multi-GPU Exchange Monte Carlo for 3D Random Field Ising Model”, https://github.com/crinavar/trueke

Scale-out:

• B2rEqwp: Z. Liu, “Efficient Large-scale Parallel Stencil Computation on Multi-Core and Multi-GPU Accelerated Clusters”, https://github.com/lzhengchun/b2r
• Diffusion: M. A. Diaz, “Multi-GPU (CUDA-MPI) baseline implementation of Heat Equation and the inviscid Burgers’ equation”, https://github.com/wme7/MultiGPU_AdvectionDiffusion
• Lulesh: LLNL, “Livermore Unstructured Lagrangian Explicit Shock Hydrodynamics”, https://codesign.llnl.gov/lulesh.php
• CoMD: NVIDIA, “GPU implementation of classical molecular dynamics proxy application”, https://github.com/NVIDIA/CoMD-CUDA
• Prbench: NVIDIA, “A CUDA implementation of the PageRank Pipeline Benchmark”, https://github.com/NVIDIA/PRBench
• HIT: M.V.Martin, “HIT: a parallel GPGPU code to simulate Homogeneous Isotropic Turbulence”, https://github.com/albertovelam/HIT_MPI
• Mavtec: T. Agarwal, “Multi-GPU Matrix Multiplication using CUDA and MPI”, https://github.com/tejaswiagarwal/multigpumatmul

The Tartan Benchmark Suite and the evaluation research was supported by the U.S. DOE Office of Science, Office of Advanced Scientific Computing Research, under award 66150: "CENATE - Center for Advanced Architecture Evaluation". The Pacific Northwest National Laboratory is operated by Battelle for the U.S. Department of Energy under contract DE-AC05-76RL01830. This research was also supported by the Exascale Computing Project (17-SC-20-SC), a collaborative effort of the U.S. DOE Office of Science and National Nuclear Security Administration. Part of the computing resources is from the Oak Ridge Leadership Computing Facility. The Oak Ridge National Laboratory is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725.

##MIT License

This benchmark suite is modified from individual applications. Please refer to each application's license requirement.

Copyright 2018 Ang Li, PNNL.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

tartan's Issues

Scale-up-nvlink kmeans memcpy and memzero errors

When I try to make the kmeans application in scale-up-nvlink, I get the following error:

centroids.cu(105): error: no instance of overloaded function "kmeans::detail::memcpy" matches the argument list
            argument types are: (thrust::device_vector<int, thrust::device_allocator<int>>, thrust::device_vector<int, thrust::device_allocator<int>>)

centroids.cu(116): error: no instance of function template "kmeans::detail::memzero" matches the argument list
            argument types are: (thrust::device_vector<double, thrust::device_allocator<double>>)

centroids.cu(119): error: no instance of function template "kmeans::detail::memzero" matches the argument list
            argument types are: (thrust::device_vector<int, thrust::device_allocator<int>>)

where is OLCF_SPECTRUM_MPI_ROOT ?

Hello, I want to use the above test program, but I can't find this definition below:
"source $OLCF_SPECTRUM_MPI_ROOT/jsm_pmix/bin/export_smpi_env -gpu "

Hardcoded Paths in Makefiles

At least some of the Makefiles in this project hardcode paths to MPI, NCCL, etc. at the lowest directory level, which means a user from a different system has to change those defaults just to be able to build the project. Example: microbenchmark/scale_out_cl/src/Makefile

There are shared.mk files to centralize that and they are used in SOME of the Makefiles, but not in most of them. Please fix that.