How To Compare Amazon Ec2 To A Supercomputer For Scientific Applications



Similar documents
Performance of HPC Applications on the Amazon Web Services Cloud

Cost-Benefit Analysis of Cloud Computing versus Desktop Grids

Volunteer Computing, Grid Computing and Cloud Computing: Opportunities for Synergy. Derrick Kondo INRIA, France

Cloud Computing. Alex Crawford Ben Johnstone

A PERFORMANCE COMPARISON USING HPC BENCHMARKS: WINDOWS HPC SERVER 2008 AND RED HAT ENTERPRISE LINUX 5

A Study on the Scalability of Hybrid LS-DYNA on Multicore Architectures

Volunteer Computing and Cloud Computing: Opportunities for Synergy

Science in the Cloud Exploring Cloud Computing for Science Shane Canon. Moab Con May 11, 2011

Kashif Iqbal - PhD Kashif.iqbal@ichec.ie

HETEROGENEOUS HPC, ARCHITECTURE OPTIMIZATION, AND NVLINK

HPC and Big Data. EPCC The University of Edinburgh. Adrian Jackson Technical Architect

Building Clusters for Gromacs and other HPC applications

Enabling Technologies for Distributed and Cloud Computing

Cost-Benefit Analysis of Cloud Computing versus Desktop Grids. By : Paritosh Heera( MT )

Debunking some Common Misconceptions of Science in the Cloud

Performance Evaluation of Amazon EC2 for NASA HPC Applications!

High Performance Computing in CST STUDIO SUITE

Performance Analysis of High Performance Computing Applications on the Amazon Web Services Cloud

Parallel Programming Survey

Performance Analysis of High Performance Computing Applications on the Amazon Web Services Cloud

Appro Supercomputer Solutions Best Practices Appro 2012 Deployment Successes. Anthony Kenisky, VP of North America Sales

SURFsara HPC Cloud Workshop

benchmarking Amazon EC2 for high-performance scientific computing

Trends in High-Performance Computing for Power Grid Applications

Performance Across the Generations: Processor and Interconnect Technologies

Building a Private Cloud with Eucalyptus

StACC: St Andrews Cloud Computing Co laboratory. A Performance Comparison of Clouds. Amazon EC2 and Ubuntu Enterprise Cloud

Understanding the Performance and Potential of Cloud Computing for Scientific Applications

OpenMP Programming on ScaleMP

COMP/CS 605: Intro to Parallel Computing Lecture 01: Parallel Computing Overview (Part 1)

Department of Computer Sciences University of Salzburg. HPC In The Cloud? Seminar aus Informatik SS 2011/2012. July 16, 2012

A Holistic Model of the Energy-Efficiency of Hypervisors

A Performance and Cost Analysis of the Amazon Elastic Compute Cloud (EC2) Cluster Compute Instance

Comparing Cloud Computing Resources for Model Calibration with PEST

Defense Technical Information Center Compilation Part Notice

System Models for Distributed and Cloud Computing

10- High Performance Compu5ng

Enabling Technologies for Distributed Computing

The CNMS Computer Cluster

HPC Update: Engagement Model

IBM Platform Computing Cloud Service Ready to use Platform LSF & Symphony clusters in the SoftLayer cloud

The Green Index: A Metric for Evaluating System-Wide Energy Efficiency in HPC Systems

Introduction to High Performance Cluster Computing. Cluster Training for UCL Part 1

SURFsara HPC Cloud Workshop

Sun Constellation System: The Open Petascale Computing Architecture

David Rioja Redondo Telecommunication Engineer Englobe Technologies and Systems

CPU Benchmarks Over 600,000 CPUs Benchmarked

Performance Analysis of a Hybrid MPI/OpenMP Application on Multi-core Clusters

High Performance Computing (HPC)

1 Bull, 2011 Bull Extreme Computing

1 DCSC/AU: HUGE. DeIC Sekretariat /RB. Bilag 1. DeIC (DCSC) Scientific Computing Installations

Comparing the performance of the Landmark Nexus reservoir simulator on HP servers

Clusters: Mainstream Technology for CAE

A Very Brief Introduction To Cloud Computing. Jens Vöckler, Gideon Juve, Ewa Deelman, G. Bruce Berriman

Magellan A Test Bed to Explore Cloud Computing for Science Shane Canon and Lavanya Ramakrishnan Cray XE6 Training February 8, 2011

Optimization of a parallel permutation testing function for the SPRINT R package

GPU System Architecture. Alan Gray EPCC The University of Edinburgh

Seeking Opportunities for Hardware Acceleration in Big Data Analytics

FLOW-3D Performance Benchmark and Profiling. September 2012

Building a Top500-class Supercomputing Cluster at LNS-BUAP

Parallel Computing. Introduction

Performance Evaluation of the XDEM framework on the OpenStack Cloud Computing Middleware

Performance measurement of a private Cloud in the OpenCirrus Testbed

Kriterien für ein PetaFlop System

Workshop on Parallel and Distributed Scientific and Engineering Computing, Shanghai, 25 May 2012

Power Efficiency Metrics for the Top500. Shoaib Kamil and John Shalf CRD/NERSC Lawrence Berkeley National Lab

Cluster Computing at HRI

Cost-Benefit Analysis of Cloud Computing versus Desktop Grids

PARALLEL & CLUSTER COMPUTING CS 6260 PROFESSOR: ELISE DE DONCKER BY: LINA HUSSEIN

PassMark - CPU Mark Multiple CPU Systems - Updated 17th of July 2012

wu.cloud: Insights Gained from Operating a Private Cloud System

The Assessment of Benchmarks Executed on Bare-Metal and Using Para-Virtualisation

High-Performance, Low-Cost Computational Chemistry: Servers in a Stick, Box, and Cloud. Nathan Vance Polik Group Hope College February 19, 2015

MAGELLAN 54 S CIDAC REVIEW S PRING 2010 WWW. SCIDACREVIEW. ORG

Toward a practical HPC Cloud : Performance tuning of a virtualized HPC cluster

Streamline Integration using MPI-Hybrid Parallelism on a Large Multi-Core Architecture

A Quantitative Analysis of High Performance Computing with Amazon s EC2 Infrastructure: The Death of the Local Cluster?

Improved LS-DYNA Performance on Sun Servers

Using the Windows Cluster

Dutch HPC Cloud: flexible HPC for high productivity in science & business

Home Software Hardware Benchmarks Services Store Support Forums About Us. CPU Mark Price Performance (Click to select desired chart)

Mississippi State University High Performance Computing Collaboratory Brief Overview. Trey Breckenridge Director, HPC

On-Demand Supercomputing Multiplies the Possibilities

Systems, Storage and Software in the National Supercomputing Service. CSCS User Assembly, Luzern, 26 th March 2010 Neil Stringfellow

Grid Computing vs Cloud

High Performance. CAEA elearning Series. Jonathan G. Dudley, Ph.D. 06/09/ CAE Associates

SR-IOV In High Performance Computing

VON/K: A Fast Virtual Overlay Network Embedded in KVM Hypervisor for High Performance Computing

Cluster Scalability of ANSYS FLUENT 12 for a Large Aerodynamics Case on the Darwin Supercomputer

Neptune. A Domain Specific Language for Deploying HPC Software on Cloud Platforms. Chris Bunch Navraj Chohan Chandra Krintz Khawaja Shams

Overview of HPC Resources at Vanderbilt

GRIDCENTRIC VMS TECHNOLOGY VDI PERFORMANCE STUDY

Performance In the Cloud. White paper

Home Software Hardware Benchmarks Services Store Support Forums About Us

Cloud computing for fire engineering. Chris Salter Hoare Lea, London, United Kingdom,

LS-DYNA Best-Practices: Networking, MPI and Parallel File System Effect on LS-DYNA Performance

PSE Molekulardynamik

Cloud Computing through Virtualization and HPC technologies

Clearing the Clouds. Understanding cloud computing. Ali Khajeh-Hosseini ST ANDREWS CLOUD COMPUTING CO-LABORATORY. Cloud computing

Numerix CrossAsset XL and Windows HPC Server 2008 R2

Transcription:

Amazon Cloud Performance Compared David Adams

Amazon EC2 performance comparison How does EC2 compare to traditional supercomputer for scientific applications? "Performance Analysis of High Performance Computing Applications on the Amazon Web Services Cloud", 2010

HPC Benchmarking NERSC - benchmark framework MPI Head node, worker nodes File server implemented with EBS IPM - MPI communication monitor Compared: Amazon EC2 - N node, m1.large instance 4xEC2 compute units 1-1.2 ghz opteron or xeon per unit Carver - 400 node, 2 x intel quad 2.67 nehalem / node Franklin - 9660 node cray xt4, quad 2.3 opteron / node Lawrencium - 198 node 2x intel xeon quad 2.66 / node

NERSC Benchmark Suite CAM Community Atmosphere Model Stresses processor data movement and MPI interconnect p2p bandwidth Gamess General Atomic and molecular electronic structure system memory access and bandwidth, collective interconnect performance GTC Stresses indirect addressing and random access memory IMPACT-T Integrated Map and Particle Accelerator Tracking Time sensitive to memory bandwidth and MPI performance

NERSC Benchmark Suite Cont MAESTRO Stresses memory performance, latency and global communications MILC Stresses memory bandwidth, prefetching and processing power Paratec Parallel Total Energy Code Stresses global communication bandwidth, processing power HPCC 7 synthetic benchmarks Targets computation, communications

Performance: Application Runtime Metrics take into account cluster size Performance Analysis of High Performance Computing Applications on the Amazon Web Services Cloud

Performance: Percentage runtime communicating using IPM Performance Analysis of High Performance Computing Applications on the Amazon Web Services Cloud

Performance: Sustained Flops Performance Analysis of High Performance Computing Applications on the Amazon Web Services Cloud

Considerations using EC2 Heterogeneous cpus: Intel Xeon E5430 2.66GHz quad-core processor AMD Opteron 270 2.0GHz dual-core processor AMD Opteron 2218 HE 2.6GHz dual-core processor Cannot optimize code High performance variability Sharing hardware with other vms Slow node communication Gigabit ethernet "Transient errors" Failure to boot, network misconfigurations, virtual machine hangs Not always able to acquire requested cores 256+ cores require scheduling/reservation

Cost/Performance compared to Desktop Grid How does Amazon EC2 compare to Grid Computing? "Cost-Benefit Analysis of Cloud Computing versus Desktop Grids", 2009

Desktop Grid/Volunteer Computing Fastest virtual supercomputers (From wikipedia) Bitcoin network 168.26 PFLOPS BOINC 5.634 PFLOPS Folding@Home 5 PFLOPS MilkyWay@Home 1.6 PFLOPS SETI@Home 730 TFLOPS Einstein@Home 210 TFLOPS Amazon HPC 240 teraflops 17024 cores

Considerations using VC Slow acquisition of computing resources 7.8 days to achieve 1000 cloud node equivalent Slow task deployment time = (reconnections * # tasks) / # clients 1000 tasks to 10000 nodes about 45 min Slow completion times deadlines, priorities, 96+% completion rate Average 9 days vs < 4 hours on dedicated

When would you use VC over Cloud? Cost-Benefit Analysis of Cloud Computing versus Desktop Grids

Cloud Power attainable given VC Costs Resources Per Month Given 12k/Month Processing Storage SETI 514 TeraFLOPS 7.7 TB Amazon 2 TeraFLOPS* 80 TB* *One or the other

Cloud-VC Hybrid Approach Cost-Benefit Analysis of Cloud Computing versus Desktop Grids Host VC Server on Cloud: Cost Breakdown

Storage vs Bandwidth Storage vs Bandwidth for a fixed budget Cost-Benefit Analysis of Cloud Computing versus Desktop Grids

Conclusions VC outperform clouds on cost for large long term and highly parallel projects Projects on the order of weeks VC needs a certain number of volunteer nodes before cost effectiveness High startup costs make short term projects not cost effective 1 small EC2 instance is equivalent to 2.83 VC hosts Hybrid approach can lower startup and monthly costs of VC 40% savings on SETI