Objet : Developers list for StarPU
Archives de la liste
- From: Jean Etienne Ndamlabin mboula <jean-etienne.ndamlabin-mboula@inria.fr>
- To: starpu-devel@inria.fr
- Cc: Berenger Bramas <berenger.bramas@inria.fr>
- Subject: [starpu-devel] Issues while using of hypervisors
- Date: Fri, 8 Jul 2022 15:35:12 +0200 (CEST)
- Authentication-results: mail2-relais-roc.national.inria.fr; dkim=none (message not signed) header.i=none; spf=SoftFail smtp.mailfrom=jean-etienne.ndamlabin-mboula@inria.fr; spf=None smtp.helo=postmaster@zcs-store1.inria.fr
Hello,
I'am working (with Bérenger in Cc) on the scheduling of several applications using scheduling contexts and hypervisors.
I would like to know the followings:
- What does it mean when worker = -2 in the poped_task callback of the hypervisor?
*for instance in feft_lp_handle_poped_task callback of the feft_lp_policy, the first line is "if(worker == -2) return;"
*we also faced some situation where worker = -2.
- It seems that empty task are passed unto the poped_task callback of the hypervisor. Is it possible?
*we have flops = 0 for all tasks,
*we have workerid = 0 for all tasks (even while 0 < worker < STARPU_NMAXWORKERS)
*we haveworkerids_len = 0 for all tasks (even while 0 < worker < STARPU_NMAXWORKERS)
*etc.
- When we try to count the poped tasks globally (poped_task callback of the hypervisor), it is not up to the total tasks executed by the workers. Are we missing something?
- How to insure the execution of the following hypervisor's methods: .handle_submitted_job, .handle_post_exec_hook, .handle_idle_cycle
*I filled the structure of the hypervisor :
struct sc_hypervisor_policy rsched_policy;
rsched_policy.name = "rsched";
void *perf_counters = sc_hypervisor_init(&rsched_policy);
*register the contexts (previously reated):
sc_hypervisor_register_ctx(apps_ctx[ctx_idx], 0.0);
starpu_sched_ctx_set_perf_counters(apps_ctx[ctx_idx], perf_counters);
*and:
sc_hypervisor_ctl(sched_ctxs[ctx_idx],
SC_HYPERVISOR_MIN_WORKERS, 0,
SC_HYPERVISOR_MAX_WORKERS, ncpus + ngpu,
SC_HYPERVISOR_MIN_TASKS, 10,
SC_HYPERVISOR_MAX_IDLE, ctx_workerids, ctx_nworkers / 2, 5000.0,
SC_HYPERVISOR_MAX_IDLE, ctx_workerids, ctx_nworkers, 10000.0,
SC_HYPERVISOR_NEW_WORKERS_MAX_IDLE, 100.0,
//SC_HYPERVISOR_EMPTY_CTX_MAX_IDLE, ctx_workerids, ctx_nworkers, 500.0,
//SC_HYPERVISOR_TIME_TO_APPLY, 1,
//SC_HYPERVISOR_GRANULARITY, 2,
SC_HYPERVISOR_NULL);
*But these methods are never called. Is there something to do for them to be, or is there a filter that is the problem?
Thank you!
Best regards,
Etienne
- [starpu-devel] Issues while using of hypervisors, Jean Etienne Ndamlabin mboula, 08/07/2022
Archives gérées par MHonArc 2.6.19+.