![]() Not asking for backup CPU work with less than 3 mins left and GPUs suspended - Try 1īoinc is running thee cpu tasks from Seti_beta and one gpu task from various projects. Not asking for backup CPU work with less than 3 mins left and GPUs suspended - Try 2 Second test for Keith Myers, #1677, with app_config for SETI. No cpu tasks are able to run with both and 16 in use When project has less work cached than current request level. ![]() Revisiting GPU Exclusions - Work Fetch incorrectly detecting idle GPUs as not idle Scheduler interaction between long duration short due date tasks and short duration long due date tasks ![]() To explore REC calculation for different projects run on devices of different speeds. Machine has fetched one new task, "1.00 seconds 0.40 devices" every minute for the last 20 minutes. Machine was running on 3 CPU cores - fourth core used for SETI GPU support. Test for #3076 - work buffering with max concurrent. Has no work cached, can fetch for CPU, but isn't. NumberFields has app_config with max_concurrent. max_concurrent for all projects obeyedįor #3076. Haven't seen anything out of the ordinary to report. Testing: GPU exclusion leaves GPU starvedįor #3076. General test of rec_half_life_days set to 1 Starting state for work fetch bug exploration. Minimal scenario (one project, no discrete GPU, intel GPU not used, no app_config file) for investigating MT work fetch.Īfter test run locally - client actually overfetched. (I add project_max_concurrent to Rosetta and world community grid projects, but this isn't a perfect workaround). The CPU tasks then typically end up getting skipped instead. " 4:53:58 PM | collatz | skipping GPU job collatz_sieve_4404021a-de64-41b2-bbb3-0b8228e66814_0 CPU committedĪdd app_config.xml with project_max_concurrent to limit CPU tasks for any project enough let the GPU tasks get scheduled first. On systems with more than one GPU, this intermittently results in a GPU task being skipped with an error like: Scheduler debug log reporting all CPUs in use and skipping tasks.Īmaranth work fetch test with Rosetta and LHCīOINC Client 7.16.11 (And 7.16.14 on OS X) schedule too many cores (Example has tasks (1 GPU + 1CPU and 12 CPU Tasks) on 12 core system. Only about 25% of CPU capacity being utilised. When i disable leave non-gpu tasks in memory and suspend computations apps stay in memory. Using the most recent relevant GitHub artifact under Linux (#5251), CPU over-commit is seen, as reported in issue #5254. Investigating concurrent running of GPU, MT, and CPU tasks.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |