HyperThreading Aware Process Scheduling Heuristics

1
Hyper-Threading Aware Process Scheduling
Heuristics

• 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

3
Effect on Scheduling

• HT abstracted as independent logical processors
• Processes can have mutual detrimental effect
• Varies with workloads, large range of combined
  performance
• Can improve performance by scheduling processes
  that perform well together
• e.g. avoid co-scheduling two cache-thrashers
• More important in multi-package/core systems

4
What is good performance?

• Optimise for performance - how to measure?
• Maximising IPC is biased towards high-IPC tasks
• Want to compare performance under HT to what we
  would have had with no HT
• Fair to high-IPC and low-IPC tasks
• HT performance ratio pi,b for code block s in
  interval i
• System performance ratio is sum of running thread
  performance ratios

Time to execute s with no HT Time to execute s
with HT and code block b running on the other
thread
5
Performance Estimation

• Can only get pi,b exactly by measuring both cases
• Use online method to estimate
• Processor performance counter data
• Estimate pi,b as f (perf counter increments)
• f is linear function of increments
• Coefficients learned by measuring HT and no-HT
  cases for a training set
• Break tasks into multiple, short code blocks
• Time code block running no-HT and HT against
  various other tasks
• Multiple linear regression on results

6
Scheduling (1)

• Modify scheduling decision to try to maximise
  system performance ratio
• Estimate performance ratio of task pairs
• Rolling average for each pairing that occurs
• Small cache indexed by hash(pid1, pid2)
• Use this to influence scheduling
• e.g. avoid co-scheduling a pair of tasks that
  performed badly together in the past
• e.g. select best performing task for one logical
  processor given the task running on the other
  logical processor

7
Scheduling (2)

• Implementation tryhard
• Modify dynamic priority of a task depending on
  the task running on the other logical processor
• Implementation plan
• Gang schedule tasks based on a scheduling plan
  created occasionally based on recorded
  performance data
• Allow other pairings to run occasionally
• Dont want to miss a potentially good pairing
• Use heuristics rather than rigid rules
• Respect static priorities, starvation avoidance,
  etc.

8
Results

• Linux 2.4.29
• P4 Xeon HT 2.4GHz
• Benchmarks from SPEC CPU2000
• Speedup is akin to system performance ratio
• basic is native package, rather than logical
  processor, affinity
• ipc is a simple IPC-maximising scheme
• Checked for fairness and respect of static
  priority

9
Summary Conclusions

• Performance on Hyper-Threading can be estimated
  using performance counters
• More fair than counting IPC
• Performance gains are possible with HT-aware
  scheduling
• Gains are relatively small and incur some
  complexity
• May not be worthwhile
• Further work
• Improved estimation
• Application to other schedulers, e.g. Linux 2.6

10
Related Work

• Thread-sensitive scheduling Parekh 2000
• Maximise a metric (e.g. IPC, cache hit rate) per
  scheduling quantum
• Symbiotic jobscheduling Snavely 2000, 2002
• Sample, optimize, symbios
• Hyper-Threading in Linux Nakajima 2002
• Longer period performance counter sampling,
  modify hard CPU affinity of tasks
• Chip multithreading Fedorova 2004, 2005
• L2 cache-conscious scheduling to increase hit rate