R HPC Benchmark

James R. McCombs

2017-05-23

Introduction

Benchmarking is a way to assess the performance of software on a computing platform and to compare performance between different platforms. The benchmark performance results can also be used to prioritize software performance optimization efforts on emerging High Performance Computing (HPC) systems. Optimization of the R interpreter, its intrinsic functionality, and R packages for specific hardware architectures will be necessary for data analysts to take full advantage of the latest HPC clusters, and to obviate the need to reengineer their analysis workflows. We developed the RHPCBenchmark package to determine the single-node run time performance of compute intensive linear algebra kernels that are common to many data analytics algorithms, and the run time performance of machine learning functionality commonly implemented with linear algebra operations.

The benchmarks are divided into three categories: dense matrix linear algebra kernels, sparse matrix linear algebra kernels, and machine learning functionality. All of the dense linear algebra kernels are implemented around BLAS or LAPACK interfaces. The sparse linear algebra kernels are members of the R Matrix library. The machine learning benchmarks currently only cover variants of K-means functionality for clustering using the cluster package. The dense matrix linear algebra kernels, sparse matrix linear algebra kernels, and machine learning functions that are benchmarked are all part of the R interpreter’s intrinsic functionality or packages included the with the R programming environment standard distributions from CRAN.

The following functions are intended to be called by users of this library to perform benchmarking for each of the three categories of benchmarks:

Each of these benchmark functions is a top-level function which executes microbenchmarks on several R computational kernels and other functionality relevant to each benchmark category. The top-level benchmarks execute multiple performance trials for each microbenchmark in which the run time (synonymous with wall-clock time in this document) is obtained. The benchmarks then compute the average, standard deviation, maximum, and minimum run time for each microbenchmark. The user may also specify one or more warm-up runs to ensure the R programming environment is settled before executing the performance trials. The results of each microbenchmark performance trial are written to a data frame returned by the top-level benchmark functions. The summary statistics for each microbenchmark are written to files in CSV format as the benchmark function progresses through the microbenchmarks, permitting retention of data if the benchmark function is terminated prematurely. If a CSV file already exists in an output directory, the new results will be concatenated to the existing file.

See the object documentation for the top-level benchmark functions and the microbenchmark definition classes listed below for information on how to configure the individual microbenchmarks. If any of the microbenchmarks fails to run in a timely manner or fails due to memory constraints, the matrix sizes and number of performance trials per matrix can be adjusted. Each benchmark function takes as input microbenchmark definition objects that specify data sets and microbenchmarking functions used to time the functionality being benchmarked. The microbenchmark definition classes supported are listed below.

Microbenchmark definition classes:

Each microbenchmark definition class has an allocator function field and a benchmark function field which specify functions for allocating the data for the microbenchmark and the function that performs the timing of functionality being tested. The microbenchmark definition classes also have parameters used to specify dimensions of the data sets to be used with the microbenchmarks. See the object documentation for the microbenchmark definition classes for specifics on the class fields and how they should be populated.

Supported Microbenchmarks

The microbenchmarks supported by each benchmark are described in this section.

Supported Dense Matrix Microbenchmarks

The dense matrix microbenchmarks supported by the dense matrix benchmark are: Cholesky factorization, matrix cross product, matrix determinant, eigendecomposition, linear solve with multiple right hand sides, least squares fit, matrix deformation and transpose, matrix-matrix multiplication, matrix-vector multiplication, QR decomposition, and singular value decomposition.

Each microbenchmark is tested with several matrices of increasing size. The tested matrix dimensions are parameterized by \(N\) with values of \(N\) equal to: 1000, 2000, 4000, 8000, 10000, 15000, and 20000. The lsfit microbenchmark tests matrices of dimension \(2N\)-by-\(N/2\), while the remaining microbenchmarks all test \(N\)-by-\(N\) (square) matrices.

The table below lists the name of each microbenchmark supported along with a brief description of the functionality it tests. The kernel functions are applied to matrix or vector operands, where A and B are input matrices and x is an input vector generated by the corresponding allocator function.

Microbenchmark name Functionality tested Kernel functions
choleksy Cholesky factorization chol(A)
crossprod Matrix cross product crossprod(A)
deformtrans Transpose, reshape, and retranspose matrix B <- t(A); dim(B) <- c(N/2, 2*N); t(B)
determinant Matrix determinant determinant(A)
eigen Eigendecomposition eigen(A, symmetric=FALSE, only.values=FALSE)
lsfit Least squares fit lsfit(A, b, intercept=FALSE)
solve Linear solve solve(A, B)
transpose Matrix transpose t(A)
matmat Matrix-matrix multiplication A %*% B
matvec Matrix-vector multiplication A %*% x
qr QR decomposition qr(A, LAPACK=TRUE)
svd Singular value decomposition svd(A)

For fast performance of the dense matrix kernels, it is crucial to link the R programming environment with optimized BLAS and LAPACK libraries. It is also important to have substantial amounts of memory (16GB minimum) to run most of the microbenchmarks.

The function GetDenseMatrixDefaultMicrobenchmarks defines the default microbenchmarks referenced in the table. It returns a vector of DenseMatrixMicrobenchmark objects specifying each microbenchmark. Microbenchmark parameters that can be specified include the dimensions of the matrices to be performance tested with, number of performance trials per matrix, and the allocator and microbenchmarking functions to be used. By default, the benchmark function RunDenseMatrixBenchmark calls GetDenseMatrixDefaultMicrobenchmarks to specify the microbenchmarks to be executed. See the object documentation for GetDenseMatrixDefaultMicrobenchmarks for more details.

The allocator functions for the dense matrix microbenchmarks take a DenseMatrixMicrobenchmark object specifying the microbenchmark and an integer index indicating which of the above values of \(N\) is to be applied. The allocator must return a list of allocated data objects, including the matrix, for the microbenchmark to operate on. The microbenchmarking functions take the DenseMatrixMicrobenchmark object defining the microbenchmark and the list of data objects returned by the allocator function.

Supported Sparse Matrix Microbenchmarks

The sparse matrix microbenchmarks supported by the sparse matrix benchmark are: matrix-vector multiplication, Cholesky factorization, LU factorization, and QR factorization. Each sparse matrix kernel tested is from the Matrix library included with the R distribution, and each kernel is performance tested with two or three sparse matrices from different application domains. The microbenchmarks, their associated identifiers and brief descriptions of the tested matrices are given in the table below. The table below lists each sparse matrix tested in the sparse matrix benchmark along with some fundamental characteristics of each matrix. The Laplacian matrices were generated for use with the benchmark, the rest of the matrices come from the University of Florida Sparse Matrix Collection. More detailed descriptions of the matrices from the Sparse Matrix Collection can be found on the web site. The R data files containing the sparse matrices can be downloaded in the companion package RHPCBenchmarkData.

Matrix name Number of rows Number of columns Number of non-zeros Symmetry and Structure
laplacian7pt_100 1000000 1000000 6940000 symmetric banded
laplacian7pt_200 8000000 8000000 55760000 symmetric banded
ca2010 710145 710145 3489366 symmetric block
ct20stif 52329 52329 2600295 symmetric block
Andrews 60000 60000 760154 symmetric unstructured
G3_circuit 1585478 1585478 7660826 symmetric block
circuit5M_dc 3523317 3523317 14865409 non-symmetric block/banded
stomach 213360 213360 3021648 non-symmetric banded
torso3 259156 259156 4429042 non-symmetric banded
Maragal 21255 10152 537694 rectangular unstructured
landmark 71952 2704 1146848 rectangular banded

The following table lists the microbenchmarks supported by the sparse matrix benchmark, the matrices each microbenchmark is executed with, and the kernel function that is microbenchmarked. The kernel functions are applied to matrix or vector operands, where A and B are input matrices and x is an input vector generated by the corresponding allocator function.

Microbenchmark name Functionality tested Problem Domain Description Kernel function
matvec_laplacian7pt_100 Sparse matrix-vector multiplication Laplacian 7-point stencil applied to 100x100x100 grid A %*% x
matvec_laplacian7pt_200 Sparse matrix-vector multiplication Laplacian 7-point stencil applied to 200x200x200 grid A %*% x
matvec_ca2010 Sparse matrix-vector multiplication Undirected weighted graph from congressional redistricting A %*% x
cholesky_ct20stif Sparse Cholesky factorization Stiffness matrix from structural Matrix::Cholesky(A)
cholesky_Andrews Sparse Cholesky factorization Eigenvalue problem from computer graphics/vision Matrix::Cholesky(A)
cholesky_G3_circuit Sparse Cholesky factorization Circuit simulation Matrix::Cholesky(A)
lu_circuit5M_dc Sparse LU factorization Large circuit DC analysis Matrix::lu(A)
lu_stomach Sparse LU factorization 3D electro-physical model of duodenum Matrix::lu(A)
lu_torso3 Sparse LU factorization 3D electro-physcial model of torso Matrix::lu(A)
qr_Maragal_6 Sparse QR factorization Rank deficient least squares Matrix::QR(A)
qr_landmark Sparse QR factorization Least squares Matrix::QR(A)

The functions GetSparseMatrixVectorDefaultMicrobenchmarks, GetSparseCholeskyDefaultMicrobenchmarks, GetSparseLuDefaultMicrobenchmarks, and GetSparseQrDefaultMicrobenchmarks define the default microbenchmarks referenced in the table. Each returns a vector of SparseMatrixMicrobenchmark objects specifying each microbenchmark. Microbenchmark parameters that can be specified include the dimensions of the matrices to be performance tested with, the number of performance trials to be conducted per matrix, and the allocator and microbenchmarking functions to be used. By default, the benchmark function RunSparseMatrixBenchmark calls each of the GetSparse* functions to specify the microbenchmarks to be executed. See the object documentation for RunSparseMatrixBenchmarks and the GetSparse* functions for more details.

The allocator functions for the sparse matrix microbenchmarks take a SparseMatrixMicrobenchmark object specifying the microbenchmark and an integer index indicating which matrix is being tested. The integer index is unused by the microbenchmarks specified by the GetSparse* default functions because the sparse matrix microbenchmarks read the test matrices from files as opposed to dynamically generating them. The allocator returns a list of data objects, including the matrix, for the microbenchmark to operate on. The microbenchmarking functions take the SparseMatrixMicrobenchmark object defining the microbenchmark and the list of data objects returned by the allocator function.

Supported Machine Learning Microbenchmarks

Currently, clustering microbenchmarks are the only microbenchmarks supported by the machine learning benchmark. The clustering benchmark generates clusters of normally distributed feature vectors in an \(N\)-dimensional, real-valued feature space where the mean of one cluster is located at the origin and the means of two clusters are each located at positions -1 and 1 of one or more of the axes. See the object documentation for the function GenerateClusterData for more details on how the clusters are generated. Microbenchmarking is performed with two cluster identification functions pam and clara, from the cluster package provided with the R distribution. The pam function implements the partitioning around medoids algorithm which has quadratic time complexity. The clara function is used for large data sets and operates in linear time complexity. The pam function is called with cluster.only=TRUE so that only assignments of feature vectors to clusters is reported. The clara function is called with samples=50, keep.data=FALSE, rngG=TRUE, and pamLike=TRUE. See the object documentation for the RunMachineLearningBenchmark, pam, and clara functions for more details.

Microbenchmark name Functionality tested Kernel function Number of features Number of clusters Number of feature vectors per cluster
pam_cluster_3_7_2500 clustering cluster::pam 3 7 2500
pam_cluster_3_7_5000 clustering cluster::pam 3 7 5000
pam_cluster_3_7_5715 clustering cluster::pam 3 7 5715
pam_cluster_16_33_1213 clustering cluster::pam 16 33 1213
pam_cluster_64_33_1213 clustering cluster::pam 64 33 1213
pam_cluster_16_7_2858 clustering cluster::pam 16 7 2858
pam_cluster_32_7_2858 clustering cluster::pam 32 7 2858
pam_cluster_64_7_5715 clustering cluster::pam 64 7 5715
clara_cluster_64_33_1213 clustering cluster::clara 64 33 1213
clara_cluster_1000_99_1000 clustering cluster::clara 1000 99 1000

The function GetClusteringDefaultMicrobenchmarks defines the default microbenchmarks referenced in the table. The function returns a vector of ClusteringMicrobenchmark objects specifying each microbenchmark. By default, the benchmark function RunMachineLearningBenchmark calls the GetClusteringDefaultMicrobenchmarks function to specify the microbenchmarks given in the table. See the object documentation for RunMachineLearningBenchmarks and the GetClusteringDefaultMicrobenchmarks functions for more details.

The allocator functions for the clustering microbenchmarks take a ClusteringMicrobenchmark object specifying the microbenchmark. The allocator must return a list of allocated data objects, including the matrix of feature vectors, for the microbenchmark to operate on. The microbenchmarking functions take the ClusteringMicrobenchmark object defining the microbenchmark and the list of data objects returned by the allocator function.

Support For Multithreading

There is no explicit support for multithreading by the benchmarking functions in the the R HPC benchmark. However, if the functionality being microbenchmarked is implemented with support for multithreading, and the number of threads can be controlled through the use of environment variables, as is often the case, then the benchmarks can be executed multithreaded. For example, if a benchmark user wishes to benchmark functionality implemented with a package engineered with OpenMP support, they can set the OMP_NUM_THREADS environment variable before running the benchmarks. If a microbenchmark tests kernel functions implemented with a multithreaded BLAS or LAPACK library, then the environment variable specific to that library for controlling the number of threads should be set. For example, the MKL_NUM_THREADS environment variable should be set when the dense matrix benchmarks are tested using an instance of R that is linked to the parallel Intel Math Kernel Library which implements multithreaded BLAS and LAPACK functionality.

Even though the benchmark functions do not control the number of threads to be utilized, the benchmarks must still report the number of threads used in the CSV files and data frames generated for reporting results. Thus, the environment variable specifying the number of threads must be retrievable in a way that is portable regardless of which multithreaded library the R programming environment is linked with. To accomplish this, the environment variable R_BENCH_NUM_THREADS_VARIABLE must be set by the benchmark user to specify the name of the particular environment variable specifying the number of threads to be used. So if R_BENCH_NUM_THREADS_VARIABLE is set to MKL_NUM_THREADS, and MKL_NUM_THREADS is set to 16, then the number of threads reported in the timing results will be 16.

For consistent behavior, the user should set the environment variable for the number of threads before the R programming environment is initialized. This is because the multithreading libraries with which the R environment may be linked will not necessarily register changes to the number of threads after initialization.

Examples

Examples are given in this section to show how to run each benchmark. Before running the examples, certain environment variables need to be set to specify the number of threads for parallel processing; see the previous section for how to properly do this.

Dense Matrix Microbenchmark Examples

This example runs all of the default dense matrix microbenchmarks, saves the summary statistics for each microbenchmark in the directory DenseMatrixResults, and saves the data frame returned from the dense matrix benchmark to a file named allResultsFrame.RData.

# Create the directory where the CSV files and data frames will be saved.
resultsDirectory <- "./DenseMatrixResults"
dir.create(resultsDirectory)
# Set the run identifier to be appended to the names of the CSV files generated
# so that they can be identified with this run.
runIdentifier <- "all"
# Run all of the default microbenchmarks
allResults <- RunDenseMatrixBenchmark(runIdentifier, resultsDirectory)
save(allResults, file="./DenseMatrixResults/allResultsFrame.RData")

This example runs all but the matrix transpose microbenchmarks, which tend to run very slowly, and saves the results to the same directory as in the previous example.

# Get the default microbenchmark definition objects and deactivate
# the unwanted matrix transpose microbenchmarks.
myDenseMatrixMicrobenchmarks <- GetDenseMatrixDefaultMicrobenchmarks()
myDenseMatrixMicrobenchmarks[["deformtrans"]]$active <- FALSE
myDenseMatrixMicrobenchmarks[["transpose"]]$active <- FALSE
# Set an appropriate run identifier
runIdentifier <- "no_transpose"
exTransposeResults <- RunDenseMatrixBenchmark(runIdentifier, resultsDirectory, myDenseMatrixMicrobenchmarks)
save(exTransposeResults, file="./DenseMatrixResults/exTransposeResultsFrame.RData")

This example runs only the matrix-matrix microbenchmark, and it modifies the default matrix dimensions to test only a few small matrices.

# Get the default microbenchmark definition objects and select only
# the matrix-matrix multiplication microbenchmark to run.
myMatmatMicrobenchmark <- list()
myMatmatMicrobenchmark[["matmat"]] <- GetDenseMatrixDefaultMicrobenchmarks()[["matmat"]]
# Set the matrix dimensions to be tested to 2000 and 3000.  The
# number of trials and number of warmup trials per matrix must also be
# updated.
myMatmatMicrobenchmark[["matmat"]]$dimensionParameters <- as.integer(c(2000, 3000))
myMatmatMicrobenchmark[["matmat"]]$numberOfTrials <- as.integer(c(3, 3))
myMatmatMicrobenchmark[["matmat"]]$numberOfWarmupTrials <- as.integer(c(1, 1))
# Set an appropriate run identifier
runIdentifier <- "matmat"
matmatResults <- RunDenseMatrixBenchmark(runIdentifier, resultsDirectory, microbenchmarks=myMatmatMicrobenchmark)
save(matmatResults, file="./DenseMatrixResults/matmatResultsFrame.RData")

Sparse Matrix Microbenchmark Examples

This example runs all of the default sparse matrix microbenchmarks, saves the summary statistics for each microbenchmark in the directory SparseMatrixResults, and saves the data frame returned from the dense matrix benchmark to a file named allResultsFrame.RData.

# Create the directory where the CSV files and data frames will be saved.
resultsDirectory <- "./SparseMatrixResults"
dir.create(resultsDirectory)
# Set the run identifier to be appended to the names of the CSV files generated
# so that they can be identified with this run.
runIdentifier <- "all"
# Run all of the default microbenchmarks
allResults <- RunSparseMatrixBenchmark(runIdentifier, resultsDirectory)
save(allResults, file="./SparseMatrixResults/allResultsFrame.RData")

This example runs only the Cholesky factorization microbenchmarks and saves the results to the same directory as in the previous example.

# Set an appropriate run identifier
runIdentifier <- "cholesky_only"
# Run only the sparse Cholesky factorization microbenchmarks
choleskyResults <- RunSparseMatrixBenchmark(runIdentifier, resultsDirectory,
   matrixVectorMicrobenchmarks=NULL, luMicrobenchmarks=NULL, qrMicrobenchmarks=NULL)
save(choleskyResults, file="./SparseMatrixResults/choleskyResultsFrame.RData")

Machine Learning Microbenchmark Examples

This example runs all of the default machine learning microbenchmarks, saves the summary statistics for each microbenchmark in the directory MachineLearningResults, and saves the data frame returned from the dense matrix benchmark to a file named allResultsFrame.RData.

# Create the directory where the CSV files and data frames will be saved.
resultsDirectory <- "./MachineLearningResults"
dir.create(resultsDirectory)
# Set the run identifier to be appended to the names of the CSV files generated
# so that they can be identified with this run.
runIdentifier <- "all"
# Run all of the default microbenchmarks
allResults <- RunMachineLearningMatrixBenchmark(runIdentifier, resultsDirectory)
save(allResults, file="./MachineLearningResults/allResultsFrame.RData")

This example shows how to specify a new clustering microbenchmark and run it.

# Create a new clustering microbenchmark that tests the `clara` method from the
# `cluster` package using a data set with 250 features, 50 clusters, and
# 2000 normally distributed feature vectors per cluster.
claraMicrobenchmark <- list()
claraMicrobenchmark[["clara_cluster_250_50_2000"]] <- methods::new(
      "ClusteringMicrobenchmark",
      active = TRUE,
      benchmarkName = "clara_cluster_250_50_2000",
      benchmarkDescription = "Clustering of 100000 250-dimensional feature vectors into 50 clusters using clara function",
      numberOfFeatures = as.integer(250),
      numberOfClusters = as.integer(50),
      numberOfFeatureVectorsPerCluster = as.integer(2000),
      dataObjectName = NA_character_,
      numberOfTrials = as.integer(3),
      numberOfWarmupTrials = as.integer(1),
      allocatorFunction = ClusteringAllocator,
      benchmarkFunction = ClaraClusteringBenchmark
   )
# Set an appropriate run identifier
runIdentifier <- "clara"
# Run the clara microbenchmark
claraResults <- RunMachineLearningBenchmark(runIdentifier, resultsDirectory,
   clusteringMicrobenchmarks=claraMicrobenchmark)
save(claraResults, file="./MachineLearningResults/claraResultsFrame.RData")