next up previous
Next: About this document ...

=90mm \epsfbox{roadrunner7.ps}
Scalable Systems Lab

Benchmarking OS Bypass In MPI Implementations

Bill Lawry
Riley Wilson
Arthur B. Maccabe

31 January 2002

Outline

Introduction

Scalable Systems Lab

Presenting
Background
- High-Performance Computing-

=170mm \epsfbox{comm.ps}

Source: ``Effects of Communication Latency, Overhead, and Bandwidth in a Cluster Architecture''
Martin, Vahdat, Culler, and Anderson;
University of California Berkeley

Background

Performance Metrics

Background

``On a cluster of workstations, applications displayed the strongest sensitivity to network overhead, slowing by as much as a factor of 50 when overhead is increased to roughly 100
microseconds.''

Source: ``Effects of Communication Latency, Overhead, and Bandwidth in a Cluster Architecture''

Martin, Vahdat, Culler, and Anderson;
University of California Berkeley

HP Computing Issues

Sources of overhead:

=170mm \epsfbox{overall.eps}
- Kernel Mode / Context Switching
- Message Control (e.g. Request To Send / Clear To Send Protocol)
- Others

Terms related to overhead: processor availability, user appliction slowdown.

HP Computing Issues
=170mm \epsfbox{workplot.ps}
Cost of Hardware Interrupt
      processor:        500 MHz AMD-K6(tm)-2
      OS:               Linux Kernel 2.2.14-5.0
      network card:     Myrinet M2L-PCI64/2-3.0

HP Computing Issues

Request To Send / Clear To Send Protocol

(long message protocol)

=170mm \epsfbox{RTS_CTS.eps}
- Want aggregate performance metrics
- Want insight into CPU and NIC interaction
(bypass)

HP Computing Issues
Developers aim to reduce overhead at the library level, system level, and network level.
  1. Decoupling Computation and Communication
  2. Host Processor Bypass

=170mm \epsfbox{overall.eps}

There are a variety of systems.

Traditional Metrics & Tools

Performance Metrics

Host Processor Bypass (OS and Application)

Traditional Metrics & Tools

Benchmark Process
Selected Model
=60mm \epsfbox{ch_model.eps}

Benchmark Methods

(ping-pong vs simultaneous messaging)

Benchmark Methods

MPI Standard

A Few Basic Library Calls

Benchmark Methods

=170mm \epsfbox{methods.eps}

Benchmark Methods

Variables

(work with and without message handling)

Benchmark Methods

=170mm \epsfbox{versus.ps}

Benchmark Methods

=170mm \epsfbox{testeffect.ps}

Restricted application bypass

Interpreting Benchmark Results

Typical Results - Poll

=105mm \epsfbox{pPORTband.ps}

=105mm \epsfbox{pPORTuser.ps}

Typical Results - PWW

=105mm \epsfbox{pwwPORTband.ps}

=105mm \epsfbox{pwwPORTuser.ps}

Revisit PWW Timings

=160mm \epsfbox{pww_method.eps}

Durations of Interest:

Portals vs MPICH/GM Time - Jemez/Bulk
=110mm \epsfbox{port_time.ps} =110mm \epsfbox{gm_time.ps}
Wenbin Zhu Modifications
=110mm \epsfbox{bandPoll.ps} =110mm \epsfbox{userPoll.ps}

Wenbin Zhu Modifications
=110mm \epsfbox{timeorig.ps}

=110mm \epsfbox{timercv.ps}

Summary
Benchmarking OS Bypass in MPI Implementations
Attained

Future:

=120mm \epsfbox{roadrunner7.ps}
Scalable Systems Lab




next up previous
Next: About this document ...
2002-02-20