Dell High Performance Cluster Computing: An Overvie · HPC Clusters Grid Computing Proprietary Standardization Fully Commoditized Simplicity/Volume/Choice. 3 TM Dell HPCC Methodology!

TM

Dell High Performance Cluster Computing:An Overview

Jenwei HsiehDell Computer Corporation

March, 2003 @ SOS7

2TM

Product Maturity Lifecycle in the Open Systems Market

4P servers1P/2P serversAppliance Servers

Network Attached Storage

Project based SANs

Heterogeneous SANs

Direct Attached Storage

RISC systems

8P servers

WorkstationDesktops

HPC Clusters

GridComputing

Proprietary Standardization Fully CommoditizedSimplicity/Volume/Choice

3TM

Dell HPCC Methodology

! Baselining and Benchmarking

! Testing Compatibility

! Tuning Performance of Components

! Developing Tools and Utilities

! Integration-Testing of Software Packages

! Conducting R&D with Key National Labs and Universities

! Partnering with Best of Class Partners

! Sharing Our Findings

4TM

Building Block Approach

InfiniBand

Parallel Benchmarks (NAS, HINT, Linpack…) Parallel Benchmarks (NAS, HINT, Linpack…) and Applicationsand Applications

VIA

Myrinet

GM

Linux Windows

MPI/Pro PVMMPICH MVICH

Quadrics

PlatformPlatformPlatform

InterconnectInterconnectInterconnect

ProtocolProtocolProtocol

OSOSOS

MiddlewareMiddlewareMiddleware

ApplicationsApplicationsApplications

ElanTCP

Fast Ethernet Gigabit Ethernet

Dell PowerEdge Servers (IA32 & IA64)

5TM

HPCC Components and Enabling Technologies

- Custom application benchmarks- Standard benchmarks- Performance studies

Vertical Solutions: application Prototyping / Sizing- Energy/Petroleum - Life Science- Automotives – Manufacturing and Design

Resource Monitoring / ManagementResource dynamic allocationCheckpoint restarting and Job redistributing

Compilers and math libraryPerformance tools- MPI analyzer / profiler- Debugger- Performance analyzer and optimizer

MPI 2.0 / Fault Tolerant MPIMPICH, MPICH-GM, MPI/LAM, PVM

Interconnect Technologies- FE, GbE, 10GE… (RDMA)- Myrinet, Quadrics, Scali- Infiniband Management Hardware

Interconnects Hardware

Interconnect Protocols

Operating Systems

Middleware / API

ClusterHardwareSoftware

Monitoring &Management

Application

Node Monitoring & Management

Benchmark

Development Tools

Job Scheduler

Platform Hardware

ClusterInstallation

ClusterFile System

Cluster monitoring Load analysis andBalancing-Remote access-Web-based GUI

Cluster monitoringDistributed System Performance Monitoring Workload analysis andBalancing-Remote access-Web-based GUI

Remote installation / configurationPXE supportSystem ImagerLinuxBIOS

- Reliable PVFS- GFS , GPFS …- Storage Cluster Solutions

IA-32, IA64 (Processor / Platform) comparisonStandard rack mounted, blade and brick servers / workstations

6TM

In-the-Box Scalability

65% scalability - 2.8 GHz

70% scalability - 2.4 GHz

76% scalability - 2.0 GHz

7TM

4-way vs. 2-way Interleaving using HINT

2.2 GHz vs. 2.4 GHz

4-way vs. 2-way interleaving

8TM

BLAST Performance Comparison

Blast comparison on different Processor types

0

500

1000

1500

2000

2500

3000

3500

4000

1 thread 2 threads 4 threads

No of threads

Tim

e (m

in)

PIII - 1.4 GHzItanium II - 1.0 GHzXeon - 2.4 GHz

Hyper-Threading

9TM

BLAS Comparison on Clusters

64 nodes (128 processors) HPL comparison using different Libraries

0

50

100

150

200

250

300

350

400

450

Linpack number with Goto Linpack number with ATLAS

Gflo

ps

Linpack number with Goto Linpack number with ATLAS

37%37% Improvement Improvement with Goto’s librarywith Goto’s library

10TM

Aggregated Write Bandwidth

11TM

One Million Cell, Implicit, Black-Oil Model

Source: Landmark Graphics

12TM

Price/Performance ComparisonUNIX vs. Xeon Clusters (W2K or LINUX)

US$50,00016 Processor/8 Gbyte W2K Cluster

US$300,00016 Processor/8 Gbyte Unix

Price Comparison67 seconds67 seconds16 2.2 GHz LINUX Processors

95 seconds95 seconds8 2.2 GHz LINUX Processors

63 seconds56 seconds16 2.4 GHz W2K Processors

100 seconds92 seconds8 2.4 GHz W2K Processors

221 seconds220 seconds8 Processor UNIX Machine B

355 seconds320 seconds8 Processor UNIX Machine A

Elapsed TimeMax CPU TimeCPU Type

1 Million Cell Model

Source: Landmark Graphics

13TM

Sample of Dell HPCC Partners

• OS/Management Tools/ISV’s– CGG– Platform Computing– Fluent– Landmark Graphics– MSC.Software– Intel: Compilers– Microsoft – RedHat

• Hardware Partners– Intel– Myricom– Extreme Networks

• Integration/Consultants– Cray– MPI Software Technology, Inc– Cornell Theory Center– Scali– SCS– TurboWorx

• Universities and National Lab– Georgia Tech, College of Computing – Oak Ridge National Lab– Penn State University– University of Texas: Center of Petroleum

& Geo-Systems Engineering – University of Houston Computer Science,

High-performance Compilers

14TM

Product Offerings

• Two classes of HPCC products: Standard and Custom• Standard:

– Low to medium size opportunities whose requirements can be generalized and packaged

– To date, we have pre-tested/validated configurations of 8, 16, 32, 64 and 128 node configurations

– Supports PIII and XEON technologies both– Fast and Gigabit Ethernet and Myrinet for intra-cluster

communication– Fast Ethernet for management fabric– Software stack for building a generic HPC stack– Professional services from pre-sales to post-sales

• Custom: – Case-by-case, larger or strategic opportunities that have unique

customer requirements and have to be handled individually.

15TM

Dell Centers for Research Excellence Awards

• Award created by Michael Dell to recognize innovative uses of High Performance Compute Clusters

– Innovation in HPCC applications or solutions: organizations thatdevelop technical enhancements that further the standardization and simplify the use of cluster computing for data intensive applications.

– Size and scope of the cluster: organizations that have HPCC deployments that achieve new levels of performance and capabilities.

– Applications and types of research: organizations that use a HPCC cluster to perform groundbreaking commercial and government research or research for the betterment of society.

16TM

Technical Computing Market

HPQ31.5%

Sun17.8%

Dell5.4%

SGI 5.5%

Cray1.7% Others

1.4%

IBM 36.7%CY’01 Dell was part of the “Others” group: We have a

Great deal of work left to do!

Source: IDC High Performance Technical Computer QView, Q4’02

TM

Thank you for your time!

Questions?