Characterizing NAS Benchmark Performance

1
Characterizing NAS Benchmark Performance

• Supercomputing Lab.
• Lee Hwang Jik

2
Contents

• Introduction
• Experimental setup
• Resource utilization
• Processor utilization
• Bandwidth usage
• Memory usage
• Performance of NPB
• Computation loads
• Reduced from 100 to 10Mbps
• Congested links and limited memory
• Conclusion Result

3
Introduction(1/2)

• Motive
• Resource selection in shared heterogeneous
  enviroments
• Before
• Systems to measure and predict the availability
  of resources on a network
• Examples NWS and Remos
• Assume to fit a known simple profile
• In practice
• Diverse structures to be difficult to quantify
• The Broad Goal
• To develop application profiles to estimate
  performance in different resource

4
Introduction(2/2)

• Basic result
• To characterize the performance of the NAS
  Parallel benchmarks on shared networks
• To establish system level measurements on a
  controlled(limited) networking testbed
• Approach
• To measure the core execution parameters of a
  program (ex CPU utilization)
• Results
• The basis of the estimation of the expected
  performance of the program

5
Experimental setup(1/2)
Measure performance with resource sharing
Executable Application Code
Benchmarking on a controlled testbed
ethernet switch (crossbar) 100 Mbps links

• capture patterns of CPU loads and traffic
  during execution

500MHz Pentium Duos
6
Experimental setup(2/2)

• Testbed
• NAS Benchmark
• EP, BT, CG, IS(C??), LU, MG, SP
• Class A size, 4 nodes, g77, gcc
• Measure tools
• CPU and memory utilization
• To use Unix top utility
• Bandwidth usage
• To employ tcpdump
• Etc..

7
Resource utilization

• Processor utilization

• CPU utilization
• under 100
• expect MG
• Difference
• CG and IS 20, 10
• LOW CPU utilization
• gt Expect CPU sharing

8
Resource utilization

• Bandwidth usage(1/2)

• Some application
• all links are busy
• - Others
• certain links are busy
• next page..
• - Avg. Pkt Size
• IS 1030 vs LU 734
• Larger message

9
Resource utilization

• Bandwidth usage(2/2)
• Dominant communication patterns in the programs

10
Resource utilization

• Memory usage

• The amount of memory utilized varies from 2MBytes
  to 117MBytes

11
Performance of NPB

• Computation Loads(1/2)

• Types of Loads is the most busy CPU, Least busy
  CPU and all nodes CPU
• Correlation between low CPU utilization and
  smaller increase in execution time (e.g. MG shows
  only 60 CPU utilization)

12
Performance of NPB

• Computation Loads(2/2)

• Increase beyond 50 due to lack of coordinated
  (gang) scheduling and synchronization
• Execution time is lower if least busy node has a
  competing load (20 difference in the busyness
  level for CG)

13
Performance of NPB

• Reduced from 100 to 10Mbps(one)

• Close correlation between link utilization and
  performance with a shared or slow link

14
Performance of NPB

• Congested links and limited memory
• Be Simply that NPB programs need the amount of
  memory.

15
Conclusion

• CPU utilization and network bandwidth
• To characterize an application in terms of its
  communication and load distribution structure and
  suggest strategies for placement.
• CPU utilization
• Some node having Competing jobs is a key
  indicator of application performance
• Bandwidth
• Closely correlated to reduced performance when
  limited bandwidth is available

16
Result