[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Low benchmark performance



Simon Derr wrote: 
> While running a simple thread wake-up benchmark on a 4-way ia64 system I
> noticed an interesting behaviour.
> 
I am also seeing a similar behavior while running a pingpong test
program.
On a 4-way IA32 system, NPTL is much slower than LinuxThreads.
Using NPTL:
./pp -v -n 128 -i 1 -S 32768
256 threads initialised in 0s.10700ms
128 games completed in 25s.648194ms

Using LinuxThreads:
./pp -v -n 128 -i 1 -S 32768
256 threads initialised in 0s.21485ms
128 games completed in 0s.17053ms

The results using NPTL varies a lot, sometimes it takes 15s, 50s,
3s,....

On a UP and 2-way, NPTL results looks good compare to LT, but on a 4-way
and beyond it's pretty unstable. The reason could be futexes in kernel
uses 3 global locks and their contention goes high on a 4-way system.
Any other reasons? Do you think any real life mutexes oriented apps will
face same problem?

- Saurabh





[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]