ScienceCloud2010

Size: px

Start display at page:

Download "ScienceCloud2010"

下元
4 years ago
Views:

1 Case Study for Running HPC Applications in Public Clouds OPEN RESEARCH, INC. QIMING HE* NASA-GODDARD SPACE FLIGHT CENTER (GSFC) SHUJIA ZHOU, BEN KOBLER, DAN DUFFY, TOM MCGLYNN

2 Motivation: HPC-in-the-cloud HPC: MPI-based parallel computing programs on distributed-memory systems. Public clouds: Virtualized computing resources Pay-as-you-go pricing model. Technically feasible? Deploy and execute any-size app in any cloud? Satisfactory performance vs. in-house HPC systems? HPC-friendly clouds? HPC-friendly application type? Economically feasible? Cloud servers costs vs. local cost of person, facilities, energy,

3 Previous Work Ed. Walker, Benchmarking Amazon EC2 for highperformance scientific computing, NPB Benchmark on EC2 vs. NCSA 5-20% performance gap for NPB-OMP. 10%-1000% performance gap for NPB-MPI. Jeff. Napper, et. al., Can cloud computing reach the top500?, LINPACK/HPL Benchmark on EC2 1) not scalable as problem and cluster size increase. 2) not economical in terms of FLOPS/$. Why EC2 only?

4 Our Approach Benchmarks & Apps Public Clouds NPB EC2 LINPACK/ HPL GoGrid Conclusions & Insights? Full-size Application IBM

NPB (from NASA) Benchmarks and Applications Benchmark Name derived from Description BT Block Tridiagonal Solve a synthetic system of nonlinear PDEs using algorithms block tridiagonal.

of linear equations EP Embarrassingly Parallel Generate independent Gaussian random variates using the Marsaglia polar method FT Fast Fourier Transform Solve a three-dimensional partial differential

5 NPB (from NASA) Benchmarks and Applications Benchmark Name derived from Description BT Block Tridiagonal Solve a synthetic system of nonlinear PDEs using algorithms block tridiagonal. CG Conjugate Gradient Estimate the largest eigenvalue of a sparse symmetric positive-definite matrix using the inverse iteration with the conjugate gradient method as a subroutine for solving systems of linear equations EP Embarrassingly Parallel Generate independent Gaussian random variates using the Marsaglia polar method FT Fast Fourier Transform Solve a three-dimensional partial differential equation (PDE) using the fast Fourier transform (FFT) IS Integer Sort Sort small integers using the bucket sort LU Lower-Upper symmetric Gauss-Seidel Solve a synthetic system of nonlinear PDEs using LU-SSOR algorithm MG MultiGrid Approximate the solution to a three-dimensional discrete Poisson equation using the V-cycle multigrid method SP Scalar Pentadiagonal Solve a synthetic system of nonlinear PDEs using scalar pentadiagonal algorithm LINAPCK/HPL (used by TOP500) CSFV (NASA weather prediction app based on Cubed- Sphere Finite-Volume Core)

6 Public Clouds Hardware Specifications EC2 (Xeon GoGrid(Xeon IBM(Nehalem Now 8 cores

7 NPB-OMP NPB OpenMP Benchmark (No MPI over network) To evaluate virtualization overhead per node

8 Cloud Networking Latency and Throughput (MPI Messages) IBM (100Mbps): consistently slow EC2 (1000Mbps?): inconsistently fast, multi-hop IP packets GoGrid (1000Mbps): consistently fast But all three are orders of magnitude slower than NCSA (with InfiniBand)

9 NPB-MPI Re-run Ed Walker s TestCase in three clouds Better network, better performance (in line with earlier researchers) GoGrid could be better if it has 8-core servers.

10 HPL Re-run Jeff Naper s TesCases in three clouds (from Naper s paper)

11 CSFV Full-size app (with 110K line of Fortran/MPI code). Compile- and run-time dependencies, e.g., NETCDF. Over-subscription helps (a little). Best-case scenario: only underperforms NASA system by 20%

12 Other Observations Programming paradigm matters MPI, OpenMP, MPI+OpenMP Embarrassingly Parallel (EP) app works the best. EP represents a wide-spectrum of apps. Commercial vs. Open source software in the cloud No performance difference observed. Cloud economics FLOPS/$, FLOPS/watt,

13 Future Work Benchmarks with larger size Re-do GoGrid with 8-core server More public clouds Terremark s vcloud Express (vmware-based) Penguin Computing s HPC-as-a-Service IBM upgrades to (at least) 1GbE network soon? Private Clouds NASA Nebula DOE Magellan Q&A

Microsoft PowerPoint - Aqua-Sim.pptx

Microsoft PowerPoint - Aqua-Sim.pptx Peng Xie, Zhong Zhou, Zheng Peng, Hai Yan, Tiansi Hu, Jun-Hong Cui, Zhijie Shi, Yunsi Fei, Shengli Zhou Underwater Sensor Network Lab 1 Outline Motivations System Overview Aqua-Sim Components Experimental