Background and Introduction
I'm working on researches in the field of Statistics. Modern days, because of the evolving applications with big data, the statistical models have to cater for the problems with more and more covariates included. Many people use R to build and test their model. However, running simulations with a large amount of covariates may take several weeks or months. Therefore, we want to know whether the simple codes written in base R is appropriate for such projects. If not, we want to code in an efficient way so that we can spend more time on the theory or model itself rather than switching between servers and testing the model through intensive but less efficient simulations. Since many people who work in this field do not have enough knowledges in computer science before, the following article will guide you to install the math libraries accordingly and the codes can be served as a basic benchmarking tool to test whether you have done the correct configuration so that when you code your own project using C++ with such libraries, the high efficiency can be achieved.
List of Contents
- Installation of required packages and the benchmarking tool
- Linear algebra operations for different libraries in c++. MKL v.s. OpenBLAS v.s. Eigen
- Linear algebra operations for different languages. c++ (MKL) v.s. Matlab (MKL) v.s. R (LAPACK)
- Linear algebra operations for different architechures. Intel KabyLake v.s. AMD Zen+
- Conclusion and suggestion
Getting Started
Prerequisites
The MKL, OpenBLAS and Eigen libraries should be installed (linux version) in order to run the following benchmark.
OpenBLAS installation
OpenBLAS can be cloned into your local directory by using git.
git clone "https://github.com/xianyi/OpenBLAS.git"
It can be installed by invoking make in the directory which contains Makefile. Make sure gfortran is installed and gcc is updated to the latest stable version. Use make install to copy your include and lib folder your preferred installation directory.
make install PREFIX=your_installation_directory
MKL installation
MKL can be downloaded from intel's official webset for free. After dowloading the whole package, the installation can be done by invoking the program install_GUI. The default installation directory is /home/usr/intel, the include and lib folder can be found in the subfolder /intel64 for 64 bit processors.
Eigen installation
Similar to MKL, Eigen can also be downloaded from the official webset. Eigen can be used driectly without installation. After untar the archive, the include libraries are in the subforder /Eigen. Usually, the dense matrix manipulation libraries can be used by including <Eigen/Dense> as the header file in your C++ code.
Installation of Benchmarking tool
Clone the files into your local directory.
git clone "https://github.com/heilokchow/math_lib_benchmark.git"
Change the value of variables for include and library directories MKL_INCDIR, MKL_LIBDIR, OMP_INCDIR, EIGEN_INCDIR, OPENBLAS_INCDIR and OPENBLAS_LIBDIR in the Makefile according to your installation path of MKL, OpenBLAS and Eigen libraries.
By invoking make MODEL=x, the benchmarking binary file can be created for different libraries.
- MODEL=1: MKL
- MODEL=2: OpenBLAS
- MODEL=3: Eigen
- MODEL=4: Eigen + MKL
- MODEL=5: Eigen + OpenBLAS
Then, the benchmark can be run by ./a.out <n> <nrep>, where n is the dimension of the matrix and nrep is the number of replications. You can use the command export OMP_NUM_THREADS=k before running the program to force MKL and OpenBLAS use k threads. An important note is that if your own program use parrallel programming techniques already, you should use single thread for these math libraries. Also, Eigen can use EIGEN_DONT_PARALLELIZE to disable parallelization.
Benchmarking for different libraries (i5-8250u @ 3.2GHz)
Here, the benchmark is run on i5-8250u @ 3.2GHz. Only single threaded performance is considered. The comparison results is ploted using summary.R and the results are shown below. The y-axis is the real time (not cpu time) spent and x-axis is the dimension of the matrix. Most paper uses the term "GFlops" to measure the performance of different implementations or algorithms. However, in real-life applications from fields like Statistics, the actual time spent is more crucial. Therefore, the elapsed time is applied for benchmarking. The gcc flags are -O2, -mfma. The FMA instruction is more efficient than SSE or AVX2, if the code is not compiled with FMA instrution sets, the Eigen library will have a dramastic performance loss.
Matrix multiplication (DGEMM subroutine)
The DGEMM subroutine from BLAS library calculates the general matrix and matrix product. It is widely used in all kinds of scientific computation and the efficiency of matrix multiplication is of vital importance. Both five methods have similar efficiency.
Solving linear equations (DGESV subroutine)
The DGESV subroutine solves the linear equations which is widly used in mathematical optimization problems. Here, Eigen library performs better than MKL and OpenBLAS. However, the llt() function is used here for Eigen and only positive definite matrix is tested here. If more general martix is evaluated in your application or you want more accuarcy, you may consider functions partialPivLu() or fullPivLu() which will be slower.
Cholesky factorization (DPOTRF subroutine)
The DPOTRF subroutine does the Cholesky factorization. The Cholesky factorization is the essential part for dependent multivariate normal distribution's construction. Both five methods have similar efficiency.
Benchmarking for different languages (i5-8250u @ 3.2GHz)
Here, we still consider the single threaded performance when running benchmarks. The benchmarking codes for R and Matlab are contained in summary.R and matlab_r2019.mlx. Note that the default run on Matlab uses multi-threaded MKL library. To force Matlab to use single thread, parrallel tool box could be installed and one worker thread should be used during the benchmarking run. The x-axis is the dimension of the matrix and the y-axis is t^(1/3) in term of seconds since all subroutines have an O(n^3) computation time. For example, from the figure below, the actual elapsed time for R to perform a 2000 x 2000 matrix multiplication is 1.55^3 which is around 4 seconds.
Matlab and Eigen performs similar and are about 10-50 times faster than R(base) for large matrix related operations. Note that the default BLAS and LAPACK library is used for R. To boost the matrix manipulations, RcppEigen packages or relinking the library to OpenBLAS could be considerred. But both two methods are not easy to be implemented, the previous one still requires coding in c++ and the later one requires a reinstallation of R which is inapplicable on most HPC servers. Therefore, R may not be a suitable choice for big data manipulation if such linear operations is needed.
Benchmarking for different architectures
i5-8250u @ 3.2GHz v.s. r7-2700x @ 4.0GHz
Due to the advantages of Eigen from benchmark results above (efficient and portable), the Eigen library is adopt for benchmarking on different architectures. Here, Intel's KabyLake is compared with AMD's Zen+. Since the actual real application's performance in researches is of vital important, both computers are run on auto boost with different speeds. The single thread performance is recorded below. y-axis is the actual time (in seconds) takes, the lower, the better.
Running at a lower clock, Intel's cpu is 10-25% times faster than AMD's cpu when Eigen library is used. Therefore, Zen or Zen+ architechture is not a good choice for such double precision floating point computation which is common in most researches. If dense matrix operations is an essentail part of your program, Intel's most architectures or Zen2 architecture is more suitable for these tasks.
Conclusion
If only 100 covariates are included in your model, R is still a good choice since the coding process is much simplier than C++ and their are many existing packages for you to use. However, if you works on models with more than 1000 covariates, you may consider other choices. If a controllable parallelization is not needed, Matlab is a good choice. Otherwise, using Eigen library with correct configuration will help you do simulations more efficiently.
