Here are some of my log reports for the past few days running a RX 5600 XT at 3x tasks and no CPU tasks.
So the recent diversity of completion times for FGRP task also is happening on AMD GPUs. The standard deviation of a day's worth of task times used to be just a few seconds, while now it is typically over a minute.
Another difference with these new workunits is that I'm now getting occasional computation errors; either very short times or very long times. I can't rule out, however, that something is crapping out only card or system.
For the past month or so individual task times (one third of completion time), is between ~3:10 and ~ 4:30, while with the earlier workunits it was very close to 6:10. I can't complain about the variability or computation errors though because the faster calculations have boosted by RAC for that card from ~800k to >1.1M!
Daily metrics of completion times for tasks running @3X:
2021-Mar-05 14:52:58; >>> SUMMARY count for the past 1d: 374
Task Times: mean 00:11:34, range [00:09:18 - 00:13:44],
stdev 00:01:13, total 3d 00:07:06
2021-Mar-06 14:54:24; >>> SUMMARY count for the past 1d: 328
Task Times: mean 00:13:13, range [00:02:28 - 15:41:39],
stdev 00:51:26, total 3d 00:15:04
2021-Mar-07 14:55:50; >>> SUMMARY count for the past 1d: 325
Task Times: mean 00:10:22, range [00:01:58 - 00:13:35],
stdev 00:01:22, total 2d 08:14:25
2021-Mar-08 14:57:17; >>> SUMMARY count for the past 1d: 374
Task Times: mean 00:14:02, range [00:09:06 - 15:41:39],
stdev 00:48:06, total 3d 15:30:57
2021-Mar-09 14:58:43; >>> SUMMARY count for the past 1d: 373
Task Times: mean 00:11:34, range [00:09:15 - 00:13:58],
stdev 00:01:09, total 2d 23:56:14
Ideas are not fixed, nor should they be; we live in model-dependent reality.
Einstein has pushed out another data file for GR tasks. 4001 series.
these behave more similarly to older GR tasks, and with nvidia cards, it's uses the full GPU (97-98% utilization) and has better production at 1x than 2x. I switched my 2080tis back to 1x. run times are a bit slower than the most recent 300x series, but still faster than the historical speed. expect to see RAC drop a bit.
Einstein has pushed out another data file for GR tasks. 4001 series.
<snip>
expect to see RAC drop a bit.
Quite a bit.
On the three machines here, for a small sample size, the increase in elapsed time for 400n tasks over 300n tasks is right about 45%. The machines all run modern AMD boards: a 5700, a 6800 Xt, and a plain 6800.
As my rough and ready speedup estimate from before to 300n was 3 to 2, these are slowing down to something very similar to what I saw before.
Here are some of my log
)
Here are some of my log reports for the past few days running a RX 5600 XT at 3x tasks and no CPU tasks.
So the recent diversity of completion times for FGRP task also is happening on AMD GPUs. The standard deviation of a day's worth of task times used to be just a few seconds, while now it is typically over a minute.
Another difference with these new workunits is that I'm now getting occasional computation errors; either very short times or very long times. I can't rule out, however, that something is crapping out only card or system.
For the past month or so individual task times (one third of completion time), is between ~3:10 and ~ 4:30, while with the earlier workunits it was very close to 6:10. I can't complain about the variability or computation errors though because the faster calculations have boosted by RAC for that card from ~800k to >1.1M!
Daily metrics of completion times for tasks running @3X:
2021-Mar-05 14:52:58; >>> SUMMARY count for the past 1d: 374 Task Times: mean 00:11:34, range [00:09:18 - 00:13:44], stdev 00:01:13, total 3d 00:07:06 2021-Mar-06 14:54:24; >>> SUMMARY count for the past 1d: 328 Task Times: mean 00:13:13, range [00:02:28 - 15:41:39], stdev 00:51:26, total 3d 00:15:04 2021-Mar-07 14:55:50; >>> SUMMARY count for the past 1d: 325 Task Times: mean 00:10:22, range [00:01:58 - 00:13:35], stdev 00:01:22, total 2d 08:14:25 2021-Mar-08 14:57:17; >>> SUMMARY count for the past 1d: 374 Task Times: mean 00:14:02, range [00:09:06 - 15:41:39], stdev 00:48:06, total 3d 15:30:57 2021-Mar-09 14:58:43; >>> SUMMARY count for the past 1d: 373 Task Times: mean 00:11:34, range [00:09:15 - 00:13:58], stdev 00:01:09, total 2d 23:56:14Ideas are not fixed, nor should they be; we live in model-dependent reality.
You 2 longest errors were due
)
You 2 longest errors were due BOINC aborted too long executing task. Perhaps they lost OpenCL context and could hang indefinitely otherwise.
Shortest error is:
Einstein has pushed out
)
Einstein has pushed out another data file for GR tasks. 4001 series.
these behave more similarly to older GR tasks, and with nvidia cards, it's uses the full GPU (97-98% utilization) and has better production at 1x than 2x. I switched my 2080tis back to 1x. run times are a bit slower than the most recent 300x series, but still faster than the historical speed. expect to see RAC drop a bit.
_________________________________________________________________________
Ian&Steve C. wrote:Einstein
)
Quite a bit.
On the three machines here, for a small sample size, the increase in elapsed time for 400n tasks over 300n tasks is right about 45%. The machines all run modern AMD boards: a 5700, a 6800 Xt, and a plain 6800.
As my rough and ready speedup estimate from before to 300n was 3 to 2, these are slowing down to something very similar to what I saw before.
The RAC bonus party is over.
The Party is over now......
)
The Party is over now......
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)