[Meeting note taken by Mr. M.S.; translated by ikoma]

CE Linux Forum
12th Tokyo Technical Jamboree
12/08/2006

---------------------------------------------------------
Performance Measurement of PPC64 RT Patch
- Tsutomu Owa (Toshiba) 
---------------------------------------------------------

# Cell Reference Set

 * Applied RT patch on PPC64
  -> Measured performance of RT patch to see the effect

 * Hypervisor OS which manages hardware resources

 * Supplied reference set with software


# RT Measurement

 * ping response

 * lpar interrupt response performance

 * netperf


# ping Response

 * With RT patch, variation of response time becomes negligible

 * Minimum response time is slower with RT than with nonRT
   Without RT, in most cases it is faster, but sometimes outrageously slow


# lpar interrupt response

 * lpar means "logical partition"

 * Without load, no difference between RT and nonRT was observed
   With some load, RT version can response fast
   With further load, sometimes lags were observed

 * Experiment to improve RT-ness
  - Remove debug options relating to RT
  - Remove influence of TLB flush in pending
    -> include/asm-ppc64/tlbflush.h
       #define PPC64_TLB_BATCH_NR 192


# netperf

 * 1.6 times faster on RT version

 * Looks like nonRT IRQ threads seem to be effective
   Good throughput at a glance
     -> why? Let me know if you know the answer


# Issues

 * Upgrade kernel

 * Evaluate RT-ness / Improve performance

 * Submit to Linux community/CELF
   Port onto other platforms

---------------------------------------------------------