search

LLNL / blockbuster

Automatically exported from code.google.com/p/blockbuster
1 stars 1 forks source link

blockbuster threads and cache mechanism is broken #125

Closed wealthychef1 closed 9 years ago

wealthychef1 commented 15 years ago

This is a biggee

Original comment by: wealthychef1

wealthychef1 commented 15 years ago

smBase is probably the bottleneck right now. When I removed all mutexes from it, it ran way faster. I was able to get 200% cpu usage with 2 threads on my Mac this way. Of course, the images are garbled and eventually it hung, but it's a proof of concept illustrating the bottleneck.

Original comment by: wealthychef1

wealthychef1 commented 15 years ago

I no longer regard this as a bottleneck for us. However, it would be good to understand the cache and preload mechanisms.

Original comment by: wealthychef1

wealthychef1 commented 15 years ago

BTW, the gist of this is that the nwin parameter in smlib is mapped onto the number of threads right now, which seems to be the way to go, but preloading is not working right, I think.

Original comment by: wealthychef1

wealthychef1 commented 15 years ago

Unless we are I/O bound, you would think a cache would be a speedup, or at least smooth out playback. It does not help, in fact using -cache 2 slows things down. And why is there a -preload and a -cache? What's the diff?

Original comment by: wealthychef1

wealthychef1 commented 10 years ago

Original comment by: wealthychef1