Poem new GPU app update

[zombie67]

Finished two tasks (at a time) on my TITAN. Took 8400 seconds each. 6500 credits each. That works out to ~133k/day.

For one of my dual 7970 machines, I am also running two tasks per card. With the two tasks, it is pegged at 99% flat. Unlike the TITAN, I need to reserve only .5 CPU thread per task.

[myshortpencil]

Finished two tasks (at a time) on my TITAN. Took 8400 seconds each. 6500 credits each. That works out to ~133k/day.

For one of my dual 7970 machines, I am also running two tasks per card. With the two tasks, it is pegged at 99% flat. Unlike the TITAN, I need to reserve only .5 CPU thread per task.

So that means the credits for Poem GPUs are far less than for the retired version.

Shoot, they're less than DiRT and Collatz, now, aren't they?

[zombie67]

Finished two tasks (at a time) on my TITAN. Took 8400 seconds each. 6500 credits each. That works out to ~133k/day.

Running 6 at a time on the TITAN:

Task run time = 15,500 seconds. That works out to ~218k/day. So 6 is better than 2. I will drop it to see what 4 does.

[zombie67]

Finished two tasks (at a time) on my TITAN. Took 8400 seconds each. 6500 credits each. That works out to ~133k/day.

Running 6 at a time on the TITAN:

Task run time = 15,500 seconds. That works out to ~218k/day. So 6 is better than 2. I will drop it to see what 4 does.

Running 4 at a time on the TITAN:

Task run time = 10,500 seconds. That works out to ~214k/day. So 4 is better the same as 6, and no added value running more than 4. I will drop it to see what 3 does.

[zombie67]

Finished two tasks (at a time) on my TITAN. Took 8400 seconds each. 6500 credits each. That works out to ~133k/day.

Running 6 at a time on the TITAN:

Task run time = 15,500 seconds. That works out to ~218k/day. So 6 is better than 2. I will drop it to see what 4 does.

Running 4 at a time on the TITAN:

Task run time = 10,500 seconds. That works out to ~214k/day. So 4 is the same as 6, and no added value running more than 4. I will drop it to see what 3 does.

Now this is odd!

Running 3 at a time on the TITAN:

Task run time = 6,900 seconds. That works out to ~243k/day. So 3 is better than 4. But how is 3 faster per task than 2???. I think I may have run that 2-per set with CPU tasks also running. I think I will try re-sunning 2-per w/o CPU tasks, to get apples to apples comparison.

[zombie67]

Now this is odd!

Running 3 at a time on the TITAN:

Task run time = 6,900 seconds. That works out to ~243k/day. So 3 is better than 4. But how is 3 faster per task than 2???. I think I may have run that 2-per set with CPU tasks also running. I think I will try re-sunning 2-per w/o CPU tasks, to get apples to apples comparison.

Running 2 at a time on the TITAN, this time without any CPU tasks running:

Task run time = 5,500 seconds. That works out to ~202k/day. So 3 is the best.

In comparison, running with 2 GPU tasks 1 full thread each AND filling up the remaining threads with CPU tasks was 133k/day. That is a SIGNIFICANT hit to production. Now I need to run some more tests to see if I can run some number of CPU tasks before the slow down kicks in.

[John P. Myers]

Running 2 at a time on the TITAN, this time without any CPU tasks running:

Task run time = 5,500 seconds. That works out to ~202k/day. So 3 is the best.

In comparison, running with 2 GPU tasks 1 full thread each AND filling up the remaining threads with CPU tasks was 133k/day. That is a SIGNIFICANT hit to production. Now I need to run some more tests to see if I can run some number of CPU tasks before the slow down kicks in.

Just a guess, but i'm gonna say the slowdown begins at the point where the number of cores used by the GPUs, plus the number of cores used by CPU WUs exceeds the number of cores the CPU has, either pushing the GPU tasks to a HT thread, or trying to cram it together on the same core/thread as a CPU WU.