DokuWiki

User Tools

Site Tools

Trace:
cluster:167

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
Next revision 		Previous revision
Next revision Both sides next revision
cluster:167 [2018/06/27 17:32]
hmeij07 [CPU vs GPU] 		cluster:167 [2018/06/28 14:00]
hmeij07
Line 10: Line 10:
   * That period covers 600 hours of time   * That period covers 600 hours of time
   * Assume 99% utilization of cpu core or gpu device   * Assume 99% utilization of cpu core or gpu device
-  * Available time is measured per cpu core but by gpu device+  * Available Hours is measured per physical CPU core but by GPU device (exclusivity and persistence modes on)
   * There is no good/bad metric   * There is no good/bad metric
   * Never collated such data before   * Never collated such data before
-  * The GPU usage is based on detecting gpu reservations (gpu= flag)+  * The GPU jobs are detected based on GPU resource reservations (gpu= flag)
  
  
-^ Metric ^ CPU ^ Ratio ^ GPU ^ Notes ^+^ Metric ^ CPU ^ Ratio ^ GPU ^ Notes June 2018 ^
 | Device Count | 72  |  3:1  |  24 | cpu all intel, gpu all nvidia | | Device Count | 72  |  3:1  |  24 | cpu all intel, gpu all nvidia |
 | Core Count | 1,192 |  1:54  |  64,300 | physical only  |  | Core Count | 1,192 |  1:54  |  64,300 | physical only  | 
 | Memory | 7,408  |  51:1  |  144 | GB | | Memory | 7,408  |  51:1  |  144 | GB |
 | Teraflops | 38  |  1.5:1  |  25 | double precision, floating point, theoretical | | Teraflops | 38  |  1.5:1  |  25 | double precision, floating point, theoretical |
-| Avail Hours | 715,200  |  50:1  |  14,400 | total cpu cores, total gpus | +| Job Count | 2,834  |  3:1  |  1,045 | processed jobs irregardless of exit status | 
-| Job Count | 2,834  |  3:1  |  1,045 | scheduled jobs irregardless of exit status | +| Avail Hours | 715,200  |  50:1  |  14,400 | total for cpu cores, total for gpus | 
-| Job Hours | 221,136  |  77:1  |  2,872 | cummulative hours of consumed usage | +| Job Hours | 221,136  |  77:1  |  2,872 | cumulative hours of consumed usage | 
-| Avail Hrs Util% | 31  |  6:1  |  5 | weeping... |+| Job Hours % | 31  |  6:1  |  5 | as a percentage |
 | Avail Hours2 | 561,600  |  39:1  |  14,400 | total cpu cores - hp12's 256 cores, total gpus | | Avail Hours2 | 561,600  |  39:1  |  14,400 | total cpu cores - hp12's 256 cores, total gpus |
-| Avail Hrs2 Util% | 39  |  8:1  |  5 | more realistic... |+| Job Hours % | 39  |  8:1  |  5 | more realistic...hp12 rarely used in June18|
  
-The logs showing gpu %util confirm the extremely low GPU usage. When concatenating the four gpu %util values int a string, since 01Jan2017, the string '0000' has occurred 10 million times out of 16 million observations. (GPUs are polled every 10 mins). Sad. The surprising strong GPU job count is due to the Amber group launches lots of small GPU jobs.+The logs showing gpu %util confirm the extremely low GPU usage. When concatenating the four gpu %util values into a string, since 01Jan2017, the string '0000' has occurred 10 million times out of 16 million observations. (GPUs are polled every 10 mins). Sad. The surprising strong GPU job count is due to the Amber group launching lots of small GPU jobs.
  
 So were these 25 days in June 2018 an oddity? March is Honors' Theses time so lets look at Jul17 so we can compare that to Jul18 in august. So were these 25 days in June 2018 an oddity? March is Honors' Theses time so lets look at Jul17 so we can compare that to Jul18 in august.
Line 36: Line 36:
 |313,303|273,051|128,390|111,224|280,101|51,727|306,453|222,585|437,959|262,227|294,724| |313,303|273,051|128,390|111,224|280,101|51,727|306,453|222,585|437,959|262,227|294,724|
  
 +^ Metric ^ CPU ^ Ratio ^ GPU ^ Notes July 2017 ^
 +| Device Count | 72  |  4:1  |  20 | cpu all intel, gpu all nvidia |
 +| Core Count | 1,192 |  1:42  |  50,000 | physical only  | 
 +| Memory | 7,408  |  74:1  |  100 | GB |
 +| Teraflops | 38  |  1.7:1  |  23 | double precision, floating point, theoretical |
 +| Job Count | 12,798  |  18:1  |  722 | processed jobs irregardless of exit status |
 +| Avail Hours | 886,848  |  60:1  |  14,880 | total cpu cores, total gpus |
 +| Job Hours |  260,997  |  69:1  |  3,805 | cumulative hours of consumed usage |
 +| Job Hours % | 30  |  1:1  |  26 | as a percentage |
 +| Avail Hours2 | 696,384  |  47:1  |  14,880 | total for cpu cores minus hp12's 256 cores, total for gpus |
 +| Job Hours % | 37  |  1.5:1  |  26 | more realistic...hp12 rarely used in June18|
 +
 +  * Some noise in this data with the inability to match start and end of job (~15% of records)
 +  * The assumption that ''hp12'' was barely used might not be correct
 +
 +Based on Jul17 we process about 60-70 times more CPU jobs than GPU jobs, that seems consistent with Jul18. The metric of Job Hours consumed versus Available Hours in %, the picture is probably more like Jul17...30-40% of CPU cycles are consumed and 25% of GPU cycles. We shall wait for Jul18 metrics.
 + 
  
 **[[cluster:0|Back]]** **[[cluster:0|Back]]**
cluster/167.txt · Last modified: 2018/08/01 14:11 by hmeij07

Page Tools

Except where otherwise noted, content on this wiki is licensed under the following license: CC Attribution-Share Alike 4.0 International
CC Attribution-Share Alike 4.0 International Donate Powered by PHP Valid HTML5 Valid CSS Driven by DokuWiki