| 1 | = Automated estimation of job and app version characteristics = |
| 2 | |
| 3 | == Goals == |
| 4 | |
| 5 | * eliminate the need for projects to supply FLOPs estimate for jobs |
| 6 | * eliminate the need for projects to supply FLOPS estimates for app versions (in app_plan()) |
| 7 | |
| 8 | == Outline == |
| 9 | |
| 10 | === Server === |
| 11 | For each app, maintain |
| 12 | * flops_avg: the estimated average number of FLOPs used by the app's jobs |
| 13 | * flops_stdev: the standard deviation of the above |
| 14 | |
| 15 | Note: if a project has different types of jobs for a given app, |
| 16 | with widely differing durations, |
| 17 | they should create separate apps for them. |
| 18 | That will reduce the variance of the estimate. |
| 19 | |
| 20 | Initially this is set to a high value (e.g. 1 GFLOP-day). |
| 21 | |
| 22 | Update: whenever a complete job is reported, |
| 23 | let x = duration * (host's flops_est). |
| 24 | Update flops_avg in a way that favors decrease over increase. |
| 25 | That way, hosts that execute the app efficiently (close to peak hardware speed) |
| 26 | have a larger weight in the estimate. |
| 27 | |
| 28 | Job completion time estimate: |
| 29 | app.flops_avg / (host's flops_est for this app version) |
| 30 | |
| 31 | |
| 32 | === client === |
| 33 | for each app version, maintain |
| 34 | |
| 35 | * flops_est: dynamic estimate of the real FLOPS of the app version on this host. |
| 36 | |
| 37 | Initially this is based on the peak hardware speed, i.e. ngpus*(GPU peak FLOPs) + avg_ncus * (whetstone). |
| 38 | Update: when a job finishes, let x = (job.flops / duration). |
| 39 | Update accordingly (but cap at peak hardware speed). |
| 40 | |
| 41 | Note: this replaces "duration correction factor". |
| 42 | |
| 43 | === protocol === |
| 44 | |
| 45 | Request message: add flops_est for each app version |
| 46 | |
| 47 | == Credit == |
| 48 | |
| 49 | Grant each validated job credit proportional to app.flops_avg |
| 50 | |
| 51 | |
| 52 | |