Screen Version

School of Computer Science & Engineering
University of New South Wales
 Advanced Operating Systems 
 COMP9242 2002/S2 

Locking: Performance Considerations

• Small lock granularity
  • decreases lock contention,
  • increases potential parallelism.
  • Also increased scope for stuffing up.
• Even with careful design hard to avoid bottlenecks (``convoys'').
  • Important to measure lock contention.
  • Instrument lock ops to keep stats.

Illustrative example

Windows-NT Kernel dispatcher lock[PS96].

• DEC people investigated performance problems of Microsoft's SQL server running on Alphas under NT.
• No access to source code.
• Used tool to patch executable code (OS and apps).
• Instrumented code logged change of control flow to memory buffer.
• Reconstructed instruction trace from log.
• Visualised results.

Result for 4-CPU system

Summary of results:

• Second box shows convoy effect on KiDispatcherLock.
  • Lock held for 200-900 cycles.
    • Partially due to interrupts being enabled during critical section.
    • Disk interrupt serviced while holding lock.
  • Lock held for about 45% of total time.
  • 16% of time spent spinning.
• This one lock limits OS scalability to about 6 CPUs!
• Shows the necessity of keeping critical sections short.