Scholarly article on topic 'Big Data Analytics Performance for Large Out-of-Core Matrix Solvers on Advanced Hybrid Architectures'

Big Data Analytics Performance for Large Out-of-Core Matrix Solvers on Advanced Hybrid Architectures Academic research paper on "Computer and information sciences"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"Matrix Multiplication" / "Out-Of-Core Matrices" / "Hybrid Architectures" / "Phi Co Processors" / "Tesla GPU"}

Abstract of research paper on Computer and information sciences, author of scientific article — Raghavendra Shruti Rao, Milton Halem, John Dorband

Abstract This paper examines the performance of advanced computer architectures for large Out-Of-Core matrices to assess the optimal Big Data system configurations., The performance evaluation is based on a large dense Lower-Upper Matrix Decomposition (LUD) employing a highly tuned, I/O managed, slab based LUD software package developed by the Lockheed Martin Corporation. We present extensive benchmark studies conducted with this package on UMBC's Bluegrit and Bluewave clusters, and NASA-GFSC's Discover cluster systems. Our results show the speedup for a single node achieved by Phi Co-Processors relative to the host CPU SandyBridge processor is about a 1.5X improvement, which is an even smaller relative performance gain compared with the studies by F. Masci where he obtains a 2-2.5x performance. Surprisingly, the Westmere with the Tesla GPU scales comparably with the Sandy Bridge and the Phi Co-Processor up to 12 processes and then fails to continue to scale. The performances across 20 CPU nodes of SandyBridge obtains a uniform speedup of 0.5X over Westmere for problem sizes of 10K, 20K and 40K unknowns. With an Infiniband DDR, the performance of Nehalem processors is comparable to Westmere without the interconnect.

Academic research paper on topic "Big Data Analytics Performance for Large Out-of-Core Matrix Solvers on Advanced Hybrid Architectures"

Procedia Computer Science

CrossMark Volume 51,2015, Pages 2774-2778 ICCS 2015 International Conference On Computational Science

Big Data Analytics Performance for Large Out-Of-Core Matrix Solvers on Advanced Hybrid Architectures

Raghavendra Shruti Rao, Dr. Milton Halem and Dr. John Dorband

University of Maryland, Baltimore County (UMBC), Maryland, U.S.A rrao1@umbc.edu, halem@umbc.edu, dorband@umbc.edu

Abstract

This paper examines the performance of advanced computer architectures for large Out-Of-Core matrices to assess the optimal Big Data system configurations., The performance evaluation is based on a large dense Lower-Upper Matrix Decomposition (LUD) employing a highly tuned, I/O managed, slab based LUD software package developed by the Lockheed Martin Corporation. We present extensive benchmark studies conducted with this package on UMBC's Bluegrit and Bluewave clusters, and NASA-GFSC's Discover cluster systems.

Our results show the speedup for a single node achieved by Phi Co-Processors relative to the host CPU SandyBridge processor is about a 1.5X improvement, which is an even smaller relative performance gain compared with the studies by F.Masci where he obtains a 2-2.5x performance. Surprisingly, the Westmere with the Tesla GPU scales comparably with the Sandy Bridge and the Phi Co-Processor up to 12 processes and then fails to continue to scale. The performances across 20 CPU nodes of SandyBridge obtains a uniform speedup of 0.5X over Westmere for problem sizes of 10K, 20K and 40K unknowns. With an Infiniband DDR, the performance of Nehalem processors is comparable to Westmere without the interconnect.

Keywords: Matrix Multiplication, Out-Of-Core Matrices, Hybrid Architectures, Phi Co Processors, Tesla GPU

1 Introduction

Matrices form the basis of Big Data Analytics involving stock control, cost, revenue and profit calculations. However, when these dense matrices grow so large that they cannot fit in the main memory of computer clusters, one must either change the mathematical algorithm or use secondary memory (e.g., Disk). As a result, new methods for managing the I/O flow of these data are being developed which, along with advances in hybrid computing architectures, offer promising capabilities for addressing the challenges of Big Data analytics.

2774 Selection and peer-review under responsibility of the Scientific Programme Committee of ICCS 2015

(gi The Authors. Published by Elsevier B.V.

doi:10.1016/j.procs.2015.05.432

Performance evaluation is of primary interest when computer systems are designed, selected, or improved (Ferrari, 1972). We provide performance evaluations on various HPC platforms available to the NSF sponsored Center for Hybrid Multicore Productivity Research at UMBC as well as those at NASA Centre for Climate Simulation at the Goddard Space Flight Center (GFSC). We employ a highly tuned, I/O managed, and slab based LU Decomposition software package developed by the Lockheed Martin Corporation for systems employing MICs and GPUs. We perform our evaluations using the Lower-Upper (LU) factorization or commonly referred to as the decomposition version of Gauss elimination, the most popular method for solving large dense systems of equations.

2 Related Work

In (Dongarra J. a., 2013), they present several fundamental dense linear algebra (DLA) algorithms for multicore with Xeon Phi Coprocessors and give an overview of MAGMA MIC (MAGMA MIC 1.2 Beta, 2014). F. Masci performs a preliminary benchmarking of a single Xeon Phi card in (Masci, 2013). Two of the three tests used for benchmarking dealt with multiplication of two large matrices of size 3000 * 3000. One implementation used explicit looping of matrix elements where loops were multithreaded using OpenMP pragmas and ran in "native-MIC", "MIC-offload", and "Host-exclusive" mode. The other implementation used highly (Phi-optimized) Math Kernel Library (MKL) to perform the matrix calculations and OpenMP to assist with multithreading. His results show that with Xeon Phi gained a speedup factor of 2-2.5 compared with an Intel Xeon E5 processor with 16 2.6 GHz cores. He opines this speed up isn't worth the cost, if the application spends most of the time in moving data off/onto disks and/or in/out of RAM. Nonetheless, he concurs some scientists may benefit from the Xeon Phi by running compute intensive customized code, e.g., N-body simulations.

3 System Architecture Overview

We performed our evaluations on three different systems. First, we chose the Linux cluster, discover, at the NASA Center for Computational Sciences (NCCS) with an IBM iDataplex with 480 Intel Many Integrated Core (Phi) co-processors. The host processor consists of dual 2.6 GHz Intel Xeon Sandy Bridge processor, each having 2 oct-core processors per node. The Interconnect is the Quad data rate Infiniband (QDR). Discover also has a set of nodes configured with Graphical Processing Units (GPUs). We ran our test on a single Tesla GPU (M2070) with 1 PCIe x16 Gen2 system interface hosted on 2 Hex-core 2.8 GHz Intel Xeon Westmere Processor. The Interconnect used in this SCU is the Dual Data Rate Infiniband (DDR).

Second, the Bluegrit cluster hosted at University of Maryland, Baltimore County (UMBC) is a distributed memory consisting of 47 Power PC nodes, one Power 7 node, 13 Intel Nehalem blades and 7 Intel Westmere blades. We used the following nodes: A single Intel(R) Xeon(R) CPU E5504 (Nehalem) running at 2.00GHz, a single Intel(R) Xeon(R) CPU X5670 (Westmere) running at 2.93GHzand a IBM blade cluster with dual 6-core Intel Xeon X5670 (Westmere) blades with 24 GB of system memory and with an attached NVIDIA Tesla M2070 GPU, which has 448 cores and 6GB graphics memory.

And finally, the Linux based distributed-memory cluster Bluewave, also hosted at UMBC, consists of 160 IBM iDataplex compute nodes, each having 2 quad core Intel(R) Xeon(R) Nehalem X5560 running at 2.80GHz. Infiniband DDR is used as the interconnect.

3 Benchmark Application

Slab Lower-Upper Decomposition (SlabLUD), the code used for this study, is a research code maintained by the Lockheed Martin Corporation (LMCO) Centralized Super Computer Facility (CSCF). It is written in Fortran95, has separate binaries as part of the benchmark: matrix_fill, matrix decomp, and matrix_solve.

The binary matrix_fill writes a matrix filled with random values to disk, using the I/O library function calls. The binary matrix_decomp factors the LHS Matrix A using LU Decomposition and stores it in while the binary matrix_solve checks for the correctness of solution. During the solve phase, the solution vectors overwrite the RHS, that is Matrix B. When data is too big to fit in the main memory, the I/O library is used to store matrices and right hand sides (RHS) on disk in slab format and utilizes a triple buffer in order to improve out-of-core performance. The triple buffer system overlaps disk I/O with computations, while two slabs are being operated and the next slab is being transferred from disk.

A GPU version of SlabLUD code has been developed by T.Blattner (Blattner, 2013). He extends the existing I/O algorithm to include the buffering of data from the compute unit memory to the accelerator memory (through the PCI Express) allowing the algorithm to seamlessly utilize accelerator technology and its available GPU library support.

4 Performance Results

Average Compute Time Per Processor (s), 1 Node, 10K * 10K Matrix

• \ \

——

I Nehaiam -Biuewave I Westmere -Bluegrit

I Nehalem - Bluegrit

I Phi-Auto Offload -Discover I Westmere -Discover I Bluegrit GPU-Tesla M2070 I Discover GPU -Tesla M2070

8 10 12

MPI Processes

Figure 5.1: Average Compute Time per Processor on one Node, on a 10K * 10K matrix

Figure 5.1 above shows average compute time per processor for 10K unknowns for a slab size of 100 MB on various architectures. The dimensions of right hand side matrix (B) is same as that of a slab, typically 100 MB or 327 columns for 10,000 unknowns. The two left dots on the top left corner indicate the time taken by the GPUs. Although both the GPU computing modules were the same, the Tesla M2070 on Bluegrit took 448 seconds as against Discover's 764 seconds. This Bluegrit GPU increased performance may be due to the fact that Bluegrit's IBM/GPU blade is connected directly to the Westmere CPU blade. Nehalem processor on Bluewave is 30% faster than that of Bluegrit. Further, the time taken on NASA's Westmere on is 4 times faster than on Bluegrit. This is made possible by the presence of Infiniband DDR as an interconnect in Bluewave and Discover systems. Xeon C5 2670 Sandy Bridge Processor and the Phi Coprocessors deliver 2.5x and 3x-5x speedup respectively, over Westmere. However, it was noted that the automatic offload does not take place, unless the slab size is increased to 4000 columns or 1220 MB. A speedup of 1.5X is observed when Phi Coprocessors are compared to the host SandyBridge, which is slightly less than the speedup observed by Masci in (Masci, 2013). The Phi Coprocessor outperforming Fermi Architecture GPU in terms of computing power indicates that the MKL library is far more optimized and well suited for computationally intensive problems.

Scalability Analysis, 10K Matrix, 1 Node

■ Nehalam -Bluewave

■ Westmere -Bluegrit

■ Nehalem -Bluegrit

■ Sandy Bridge -Discover

■ Phi-Auto Offload -Discover

■ Westmere -Discover

■ Ideal

0 2 4 6 8 10 12 14

Figure 5.2: Scalability analysis of different processor on one Node, using a 10K * 10K matrix

When the results are analyzed for their scalability as the number of processes (Fig 5.2), it can be observed that the Phi Coprocessors and SandyBridge scale very close to the ideal behavior. The 2 Nehalem (bottom 2 lines) display very similar performance, but with increased scalability on Bluewave. The same observation can be made about Westmere (2 curves at the center), where the scalability factor is 0.6-0.75x times more on Discover.

Figure 5.3 below shows the breakdown of the performance of 20 Westmere nodes for different problem sizes. We see that for 10K, the total wall time continues to scale until 12 processes. Additionally, for 20K, we observe scaling taking place until 32 processes and up to 64 processes for 40K unknowns. Thus, we can infer that for a problem size of 80K unknowns utilizing 20 nodes, the peak performance will be observed 128 processes.

Westmere - Total Wall Time on 20 Nodes with 10K, 20K and 40K unknowns

| 10K-Westmere -Discover I 20K-Westmere -Discover I 40 K-Westmere -Discover

Sandy Bridge - Total Wall Time on 20 Nodes with 10K, 20K and 40K unknowns

I 10K-Sandy Bridge -Discover I 20K-Sandy Bridge -Discover I 40K- Sandy Bridge -Discover

25 50 75 100 125 150 175 200 225 250

25 50 75 100 125 150 175 200 225 250

Figure 5.3 Figure 5.4

Similarly, we can observe in figure 5.4 that the Sandy Bridge continues to improve by a factor of 2 up to 8 processes and then performance degrades after 64 processes. For 80K, we expect to see the peak performance with 128 processes on 20 Nodes.

Figure 5.5 below depicts total wall clock time taken by Sandy Bridge and Westmere on Discover, to decompose a matrix of 10K, 20K and 40K unknowns on 20 nodes. We find that for 10K unknowns, the peak performance for both Sandy Bridge and Westmere is observed with 12 processes with a relative speed up of 0.5X for the former. When the problem size is increased to 20K, the peak for the 20 node processors is observed at 32 processes, again with a speedup of 0.5X. Further, the same speedup is observed when the compare the peak performance for 40K unknowns (64 processes). Thus, we can conclude that Sandy Bridge has a uniform speedup of 0.5X over Westmere for the problem sizes defined above.

Total Wall Time on 20 Nodes for Sandy Bridge and Westmere

110K-Sandy Bridge -Discover I ЮК -Westmere

- Discover I 20K-Sandy

Bridge -Discover I 20K -Westmere

- Discover 140K- Sandy

Bridge - Disco.. I 40K-Westmere

- Discover

12 16 32 64 128 256

MPI Processes

Figure 5.5: Total Wall Time SandyBridge and Westmere on 20 Nodes, for 10K, 20K and 40K unknowns

6 Conclusions

In this paper, we presented the performance evaluation of the I/O managed Slab LUD code on Nehalem, Westmere, Sandy Bridge and NVIDIA Tesla M2070 GPU and the PHI processors. The results show that the performance impact is most strongly correlated with the memory interconnect.

• Westmere scales slightly better than Sandy Bridge after about 8 processes for matrices that are larger than 10K* 10K but is 50% slower than Sandy Bridge.

• Bluegrit GPU performance with Westmere blade 2x faster than same Discover GPU with Westmere despite lack of Infiniband.

• On Multiple nodes, Sandy Bridge has a uniform speedup of 0.5X over Westmere for problem sizes of 10K, 20K and 40K unknowns. Performance on Sandy Bridge, with 40K unknown using 64 processes, is better than with 20K unknowns using 1 or 2 processes.

• Phi processors with Sandy Bridge offer a 3-fold performance increase over Westmere but do not auto off load unless compiler indicates it will improve performance.

• Our results show better performance for Sandy Bridge relative to the Phi Processor than reported by Masci (Masci, 2013). The speedup achieved by Phi over host Sandy Bridge is 1.5X compared with his 2X-2.5X speedup. We agree with Masci that Sandy Bridge processors more cost efficient than adding Phi processors.

7 Acknowledgements

This work was supported by Centre for Hybrid Multicore Productivity and Research (CHMPR) grant NSF- 0934364. The authors would also to thank NASA Centre for Climate Simulation for access to the Discover system and Lockheed Martin Corporation for providing the Slab LUD software.

8 References

Blattner, T. (2013). A Hybrid CPU/GPU Pipeline Workflow System. ProQuest Dissertations and Theses,80.

Dongarra, J. et. al. (2013). Portable HPC Programming on Intel Many-Integrated-Core Hardware with MAGMA Port to Xeon Phi.

Masci, F. (2013). Benchmarking the Intel Xeon Phi Coprocessor.

MAGMA MIC 1.2 Beta. (2014). Retrieved from http://icl.cs.utk.edu/magma/software/