Multiprogramming Performance of the Pentium 4 with HyperThreading

1
Multiprogramming Performance of the Pentium 4
with Hyper-Threading

• James Bulpin and Ian Pratt
• University of Cambridge
• Computer Laboratory

2
Introduction

• Simultaneous Multithreading (SMT)
• Fine-grain hardware multithreading
• Processor resources dynamically shared
• Extracts thread-level parallelism
• Intel Hyper-Threading (HT)
• 2 heavyweight threads
• Some static partitioning, dynamic core with some
  fairness mechanisms
• Shared caches, free-for-all

3
Motivation to Measure

• HT abstracted as independent processors
• Processes can have mutual detrimental effect
• Can improve performance by scheduling processes
  that perform well together
• More important in multi-package/core systems
• Need to know circumstances where performance lost

4
Duplicating and Deconstructing

• Initial Observations of the Simultaneous
  Multithreading Pentium 4 Processor. N. Tuck and
  D. M. Tullsen PACT03
• Various multithreaded and multiprogrammed
  workloads
• We have slightly different hardware and compiler
• We go further on multiprogrammed
• Bias between processes performance
• More analysis of performance counter data
• Compare Hyper-Threading to SMP

5
Experimental Method

• Aim to measure mutual performance effect of a set
  of compute-intensive workloads
• Dual package 2.4GHz HT Xeon. Linux 2.4.19
• SPEC CPU2000 Benchmarks (INT and FP)
• Measure standalone run times
• Measure runs times for crossproduct of
  simultaneously executing pairs

6
Hyper-Threading Results (1)

• Performance for an individual benchmark
• Standalone runtime / runtime in HT pair
• System speedup
• Sum of both performance figures. 1.0 is
  equivalent to serial execution
• System speedups
• Lowest 0.86 swim vs. mgrid
• Highest 1.50 mcf vs. mesa
• Mean 1.20

7
1.6
1.5
1.4
1.3
Multiprogrammed speedup
1.2
1.1
1
0.9
art
vpr
gcc
eon
mcf
gap
gzip
apsi
twolf
bzip2
swim
applu
mesa
crafty
mgrid
ammp
parser
vortex
equake
sixtrack
perlbmk
wupwise
8
Hyper-Threading Results (2)
9
Hyper-Threading vs. SMP

• Mean speedups HT 20, SMP 77
• SMP 48 improvement on HT
• (HT is worth 65 of SMP)

10
Conclusions

• Individual benchmark in a pair up to 40 faster
  than sequential execution
• High variance, lots of different behaviour
• Biased performance gains/losses
• Cache contention plays big part
• High L2 miss rates hurt the other threads
• High L1 miss rates benefit from Hyper-Threading

11
(No Transcript)
12
Hyper-Threading Results (1)
13
Best HT System Throughput
14
Worst HT System Throughput
15
SMP Result Matrix