PDA

View Full Version : No app_info for Moo?



trigggl
08-30-12, 11:45 AM
On AMD, should it just run straight up, or is there some value to running 2 at a time? Does it run considerably faster if a core is left open?

Currently, I'm running it straight up, one task at a time and not reserving a core. I'm getting less credit than when I was running MW one at a time reserving a core.

Slicker
08-30-12, 12:39 PM
On AMD, should it just run straight up, or is there some value to running 2 at a time? Does it run considerably faster if a core is left open?

Currently, I'm running it straight up, one task at a time and not reserving a core. I'm getting less credit than when I was running MW one at a time reserving a core.

The Paperboy Metaphor...

Let's say you want to wake up as soon as the paper boy delivers the morning paper. He normally comes around 7 a.m. so you set your alarm for 6:59 a.m. When the alarm goes off, you wake up and check for the paper. At 7:01 a.m. you give up and hit the snooze and go back to sleep. At 7:03 the paper arrives, but since your snooze runs for 9 minutes, you don't wake up again until 7:10. You wasted 7 minutes that could have been used reading the paper. That's kind of how most asynchronous GPU applications are developed. While snoozing the app uses very little CPU, but it doesn't always wake up when it should.

A more complex version would be to pay the paper boy and extra $5 per month so that when he delivers the paper, he will ring the doorbell to wake you up. That way, you don't have to count on the alarm clock and snooze control. You will get notified as soon as the paper boy delivers the paper and no sooner. So long as he always rings the doorbell, you will always wake up at the exact proper time. However, if the paperboy is sick, he won't ring the bell and you will be late for work or possibly sleep through the entire work day (those $3.95 all you can drink Thursday night specials used to be killers!). Likewise, GPU apps that use event callbacks risk hanging infinately unless they still set some type of alarm.

Expanding upon the paperboy scenario further... The Sunday paper is usually 3-4 times the size of the weekday papers. That means the paperboy can only carry 25% as many papers in his basket on Sunday since he has to spend more time riding back and forth to refill his bicycle basket. But, if there were 2 or more paperboys helping him on Sundays, one could be delivering papers while the others was loading their baskets. Then, one of them would always be delivering papers. So, depending upon the "size" of the GPU app, running two at a time makes sense if they spend more time loading than running. But, like delivering the weekday papers, there's no benefit to multiple paperboys if the app is small.

Since GPU utilization is dependant upon the app, your CPU, and your GPU, it is hard to determine whether it makes sense or not to run multiple WUs for a project. POEM yes. Collatz? Not since the 2.xx apps as the app info can be used to get the utilization over 99% (to the point where the video response is sluggish). Personally, I've never bothered running multiple WUs with Moo since it runs hot and at 98+ % GPU utilization already.

Mike029
08-30-12, 04:24 PM
So how many Moo wu's can the paperboy run again? 8-}

Duke of Buckingham
08-30-12, 05:04 PM
http://sjat.files.wordpress.com/2010/01/moo.jpg

trigggl
08-30-12, 05:36 PM
The Paperboy Metaphor...
...
...Personally, I've never bothered running multiple WUs with Moo since it runs hot and at 98+ % GPU utilization already.


So how many Moo wu's can the paperboy run again? 8-}

This is kind of what I got out of it.