Re: condvar wakeups

[Saurabh Desai <sdesai austin ibm com>]

Ingo Molnar wrote:
> 
> could you please also test all these benchmarks on a UP kernel?
> 
I ran on IA32- UP, 2p and 4p and results are included here.
There are some noticable degradation on UP and 2p. However, on 2p
as number of threads increases, the requeue patch helps. On 4p, it
helps in a big way. Without this patch, its very unstable and there
is no scalability at all and overall system performs poorly.

System: 4xPIII 700MHz

./cond-perf -r 100 -n 200:      1p        2p          4p
Default NPTL:                 0m0.120s   0m0.211s    3m57.407s
requeue NPTL:                 0m0.124s   0m0.156s    0m0.040s

./cond-perf -r 1000 -n 100:
Default NPTL:                 0m0.276s   0m0.412s     0m0.530s
requeue NPTL:                 0m0.349s   0m0.503s     0m0.550s

./pp -v -n 128 -i 1000 -S 32768:
Default NPTL: 128 games in    1s.11892ms  1s.270430ms  16s.894208ms
(varies)
requeue NPTL: 128 games in    1s.111067ms 1s.959749ms  2s.426947ms

./pp -v -n 1024 -i 10 -S 32768:
Default NPTL: 1024 games in   0s.181636ms 0s.394705ms  incompleted for
>2m+
requeue NPTL: 1024 games in   0s.166086ms 0s.254007ms  0s.341913ms

 
- Saurabh