MP scheduler performance scaling

I’ve collected scaling data on several machines and it looks good!

Executive summary:

All your core are belong to us!

See http://gnuradio.org/trac/wiki/MPSchedulerPerformance for pretty
pictures and links to the code and raw data.

Feel free to send more in. Use run_synthetic.py to gather it.
We could definitely use data on quad-core intel and quad-core AMD.

Does anyone have access to an SMP machine with more than 8 cores?
If so, can you try running the mp-sched code on it?

Eric

On Sun, Jul 20, 2008 at 12:08 AM, Eric B. [email protected] wrote:

I’ve collected scaling data on several machines and it looks good!

Executive summary:

All your core are belong to us!

I are more than pleased.

:wink:

Eric

Bob

On Sat, Jul 19, 2008 at 09:08:39PM -0700, Eric B. wrote:

See http://gnuradio.org/trac/wiki/MPSchedulerPerformance for pretty
pictures and links to the code and raw data.

I’ve added graphs showing Altivec performance.

The MP scaling on the PS3 is much better with Altivec than without
it. I suspect it has something to with the interaction between the
dual-thread dispatch and the VSU pipeline. (I’d always thought that
SMT / hyper-threading was a crock, but at least in this one case it
pays off).

Eric