ASpT (ppopp19_AE)

This repository contains materials for artifact evaluation of the paper "Adaptive Sparse Tiling for Sparse Matrix Multiplication". It contains:

1. Source code for ASpT (Adaptive Sparse Tiling)
2. Scripts for installation of ASpT and other compared implementations: MKL, CSB, TACO, cuSPARSE, Merge-SpMM, and BIDMach
3. The entire data set for BIDMach
4. The experimental results at ppopp19_ae_result.xlsx

Datasets

• Run ./download.sh to download all datasets. The minimum required space is 200 GB
• BIDMach uses custom data format. The entire dataset in BIDMach format is included in the repository itself

Note: Downloading and running full dataset may take more than 24 hours. Hence for convenience, we have provided ./download_small.sh which downloads only 6 datasets

Installation and Benchmarking

SpMM & SDDMM KNL:

• Run cd nlibs; make; cd ..
• Run ./compile_KNL.sh to compile all implementations
• Run ./run_KNL.sh to run all implementations (results are available on KNL_result folder)

Notes:

• INTEL_PATH (line 1) and MKL_FLAGS (line 2) in Makefile.in should be set properly before installation
• Make sure that the compiler and library path for Intel compiler and MKL is set properly (modify line 1 in run_KNL.sh )

Software requirements:

1. Intel ICC 18.0.3 (with MKL) 

Hardware requirements:

1. Intel Xeon Phi (AVX 512, 68 cores, 272 threads)
2. Clustering mode: 'All-to-All'
3. Memory mode: 'cache-mode'

SpMM GPU:

• Run ./compile_GPU_SpMM.sh to compile all implementations.
• Run ./run_GPU_SpMM.sh to run all implementations (results are available on GPU_SpMM_result folder)

Notes:

• For some datasets, Merge-SpMM throws errors and hence some results may not be available
• The performance of cuSPARSE in CUDA 9.1.85 is higher than CUDA 8.0.44. Hence for fairness, we used CUDA 9.1.85 for comparing our framework with cuSPARSE. However, for SDDMM, since BIDMach only supports CUDA 8.*, we used CUDA 8.0.44

Software requirements:

1. NVCC 9.1.85
2. GNU 4.8.5 or 4.9.3
3. Cmake 3.11.4
4. Boost 1.58

Hardware requirements:

1. Nvidia Pascal-P100 GPU  
2. Compute Capability 6.0
3. Global Memory >= 16 GB

SDDMM GPU:

• Run ./compile_GPU_SDDMM.sh to compile ASpT implementations
• Installation instruction for BIDMach is provided below
• Run './run_GPU_SDDMM.sh' to run implementations (results are availble on GPU_SDDMM_result folder).

Note:

• The performance of cuSPARSE in CUDA 9.1.85 is higher than CUDA 8.0.44. Hence for fairness, we used CUDA 9.1.85 for comparing our framework with cuSPARSE. However, for SDDMM, since BIDMach only supports CUDA 8.*, we used CUDA 8.0.44

Software requirements:

1. NVCC 8.0.44 (BIDMach only supports CUDA 8.0)
2. GNU 4.8.5 (for installiation), GNU 6.1.0 (for execution)
3. Apache Maven 3.6.0
4. JDK 8

Hardware requirements:

1. Nvidia Pascal-P100 GPU 
2. Compute Capability 6.0
3. Global Memory >= 16 GB

BIDMach installiaton:

0. mkdir build; cd build;
1. wget https://downloads.apache.org/maven/maven-3/3.6.3/binaries/apache-maven-3.6.3-bin.tar.gz
2. tar xf apache-maven-3.6.3-bin.tar.gz
3. export M2_HOME=`readlink -f apache-maven-3.6.3`
4. export M2=$M2_HOME/bin
5. export PATH=$M2:$PATH
5.1. After doing this, mvn -version should properly print the version.     Then go to this site : http://www.oracle.com/technetwork/java/javase/downloads/jdk8-downloads-2133151.html - and download jdk-8u191-linux-x64.tar.gz (182.87 MB) and do the following.
6. tar -xvf <name of downloaded file>
7. export JAVA_HOME= <path to unzipped jdk directory>
8. export PATH="$JAVA_HOME/bin:$PATH"
8.1. After doing this "javac -version" should print javac 1.8X.
9. Running './compile_GPU_SDDMM.sh'