Re: [slurm-users] slurm power save question
Thanks and no worries for the time it took to reply. Sounds good then, and it's consistent with what the documentation says, namely "prevent those nodes from being powered down". As you said "keep that number of nodes up" is a different thing, and yes, it would be nice to have. For that purpose, I'm looking at my logs of workload and mulling if I should make a cron job (submitting dummy slurm jobs) to force slurm bringing nodes up if not enough idle ones are up, to reduce wait in queue for users jobs. Thanks again On Wed, Nov 29, 2023 at 8:43 AM Brian Andrus wrote: > Sorry for the late reply. > > For my site, I used the optional ":" separator to ensure at least 4 nodes > were up. Eg: nid[10-20]:4 > This means at least 4 nodes.. those nodes do not have to be the same 4 at > any time, so if one is down that used to be idle, but 4 are up, that 1 will > not be brought back up. I don't see this setting having much of anything to > do with bringing nodes up at all with the exception of when you first start > slurmctld and the settings are not met. Once there are jobs running on any > of the listed nodes, they count toward the number. That is my experience > with the small numbers I used. YMMV. > > I have also explicitly stated nodes without the separator, which does > work. I do that when I am trying to look at a node that is idle without a > job on it. That stops slurm from shutting it down while I am looking at it. > > Although, I do agree, the functionality of being able to have "keep at > least X nodes up and idle" would be nice, that is not how I see this > documented or working. > > Brian Andrus > On 11/23/2023 5:12 AM, Davide DelVento wrote: > > Thanks for confirming, Brian. That was my understanding as well. Do you > have it working that way on a machine you have access to? If so, I'd be > interested to see the config file, because that's not the behavior I am > experiencing in my tests. > In fact, in my tests Slurm will not bring down those "X nodes" but will > not bring them up either, *unless* there is a job targeted to those. I may > have something misconfigured, and I'd love to fix that. > > Thanks! > > On Wed, Nov 22, 2023 at 5:46 PM Brian Andrus wrote: > >> As I understand it, that setting means "Always have at least X nodes up", >> which includes running jobs. So it stops any wait time for the first X jobs >> being submitted, but any jobs after that will need to wait for the power_up >> sequence. >> >> Brian Andrus >> On 11/22/2023 6:58 AM, Davide DelVento wrote: >> >> I've started playing with powersave and have a question about >> SuspendExcNodes. The documentation at >> https://slurm.schedmd.com/power_save.html says >> >> For example nid[10-20]:4 will prevent 4 usable nodes (i.e IDLE and not >> DOWN, DRAINING or already powered down) in the set nid[10-20] from being >> powered down. >> >> I initially interpreted that as "Slurm will try to keep 4 nodes idle on >> as much as possible", which would have reduced the wait time for new jobs >> targeting those nodes. Instead, it appears to mean "Slurm will not shut off >> the last 4 nodes which are idle in that partition, however it will not turn >> on nodes which it shut off earlier unless jobs are scheduled on them" >> >> Most notably if the 4 idle nodes will be allocated to other jobs (and so >> they are no idle anymore) slurm does not turn on any nodes which have been >> shut off earlier, so it's possible (and depending on workloads perhaps even >> common) to have no idle nodes on regardless of the SuspendExcNode settings. >> >> Is that how it works, or do I have anything else in my setting which is >> causing this unexpected-to-me behavior? I think I can live with it, but >> IMHO it would have been better if slurm attempted to turn on nodes >> preemptively trying to match the requested SuspendExcNodes, rather than >> waiting for job submissions. >> >> Thanks and Happy Thanksgiving to people in the USA >> >>
Re: [slurm-users] slurm power save question
Sorry for the late reply. For my site, I used the optional ":" separator to ensure at least 4 nodes were up. Eg: nid[10-20]:4 This means at least 4 nodes.. those nodes do not have to be the same 4 at any time, so if one is down that used to be idle, but 4 are up, that 1 will not be brought back up. I don't see this setting having much of anything to do with bringing nodes up at all with the exception of when you first start slurmctld and the settings are not met. Once there are jobs running on any of the listed nodes, they count toward the number. That is my experience with the small numbers I used. YMMV. I have also explicitly stated nodes without the separator, which does work. I do that when I am trying to look at a node that is idle without a job on it. That stops slurm from shutting it down while I am looking at it. Although, I do agree, the functionality of being able to have "keep at least X nodes up and idle" would be nice, that is not how I see this documented or working. Brian Andrus On 11/23/2023 5:12 AM, Davide DelVento wrote: Thanks for confirming, Brian. That was my understanding as well. Do you have it working that way on a machine you have access to? If so, I'd be interested to see the config file, because that's not the behavior I am experiencing in my tests. In fact, in my tests Slurm will not bring down those "X nodes" but will not bring them up either, *unless* there is a job targeted to those. I may have something misconfigured, and I'd love to fix that. Thanks! On Wed, Nov 22, 2023 at 5:46 PM Brian Andrus wrote: As I understand it, that setting means "Always have at least X nodes up", which includes running jobs. So it stops any wait time for the first X jobs being submitted, but any jobs after that will need to wait for the power_up sequence. Brian Andrus On 11/22/2023 6:58 AM, Davide DelVento wrote: I've started playing with powersave and have a question about SuspendExcNodes. The documentation at https://slurm.schedmd.com/power_save.html says For example |nid[10-20]:4| will prevent 4 usable nodes (i.e IDLE and not DOWN, DRAINING or already powered down) in the set |nid[10-20]| from being powered down. I initially interpreted that as "Slurm will try to keep 4 nodes idle on as much as possible", which would have reduced the wait time for new jobs targeting those nodes. Instead, it appears to mean "Slurm will not shut off the last 4 nodes which are idle in that partition, however it will not turn on nodes which it shut off earlier unless jobs are scheduled on them" Most notably if the 4 idle nodes will be allocated to other jobs (and so they are no idle anymore) slurm does not turn on any nodes which have been shut off earlier, so it's possible (and depending on workloads perhaps even common) to have no idle nodes on regardless of the SuspendExcNode settings. Is that how it works, or do I have anything else in my setting which is causing this unexpected-to-me behavior? I think I can live with it, but IMHO it would have been better if slurm attempted to turn on nodes preemptively trying to match the requested SuspendExcNodes, rather than waiting for job submissions. Thanks and Happy Thanksgiving to people in the USA
Re: [slurm-users] slurm power save question
Thanks for confirming, Brian. That was my understanding as well. Do you have it working that way on a machine you have access to? If so, I'd be interested to see the config file, because that's not the behavior I am experiencing in my tests. In fact, in my tests Slurm will not bring down those "X nodes" but will not bring them up either, *unless* there is a job targeted to those. I may have something misconfigured, and I'd love to fix that. Thanks! On Wed, Nov 22, 2023 at 5:46 PM Brian Andrus wrote: > As I understand it, that setting means "Always have at least X nodes up", > which includes running jobs. So it stops any wait time for the first X jobs > being submitted, but any jobs after that will need to wait for the power_up > sequence. > > Brian Andrus > On 11/22/2023 6:58 AM, Davide DelVento wrote: > > I've started playing with powersave and have a question about > SuspendExcNodes. The documentation at > https://slurm.schedmd.com/power_save.html says > > For example nid[10-20]:4 will prevent 4 usable nodes (i.e IDLE and not > DOWN, DRAINING or already powered down) in the set nid[10-20] from being > powered down. > > I initially interpreted that as "Slurm will try to keep 4 nodes idle on as > much as possible", which would have reduced the wait time for new jobs > targeting those nodes. Instead, it appears to mean "Slurm will not shut off > the last 4 nodes which are idle in that partition, however it will not turn > on nodes which it shut off earlier unless jobs are scheduled on them" > > Most notably if the 4 idle nodes will be allocated to other jobs (and so > they are no idle anymore) slurm does not turn on any nodes which have been > shut off earlier, so it's possible (and depending on workloads perhaps even > common) to have no idle nodes on regardless of the SuspendExcNode settings. > > Is that how it works, or do I have anything else in my setting which is > causing this unexpected-to-me behavior? I think I can live with it, but > IMHO it would have been better if slurm attempted to turn on nodes > preemptively trying to match the requested SuspendExcNodes, rather than > waiting for job submissions. > > Thanks and Happy Thanksgiving to people in the USA > >
Re: [slurm-users] slurm power save question
As I understand it, that setting means "Always have at least X nodes up", which includes running jobs. So it stops any wait time for the first X jobs being submitted, but any jobs after that will need to wait for the power_up sequence. Brian Andrus On 11/22/2023 6:58 AM, Davide DelVento wrote: I've started playing with powersave and have a question about SuspendExcNodes. The documentation at https://slurm.schedmd.com/power_save.html says For example |nid[10-20]:4| will prevent 4 usable nodes (i.e IDLE and not DOWN, DRAINING or already powered down) in the set |nid[10-20]| from being powered down. I initially interpreted that as "Slurm will try to keep 4 nodes idle on as much as possible", which would have reduced the wait time for new jobs targeting those nodes. Instead, it appears to mean "Slurm will not shut off the last 4 nodes which are idle in that partition, however it will not turn on nodes which it shut off earlier unless jobs are scheduled on them" Most notably if the 4 idle nodes will be allocated to other jobs (and so they are no idle anymore) slurm does not turn on any nodes which have been shut off earlier, so it's possible (and depending on workloads perhaps even common) to have no idle nodes on regardless of the SuspendExcNode settings. Is that how it works, or do I have anything else in my setting which is causing this unexpected-to-me behavior? I think I can live with it, but IMHO it would have been better if slurm attempted to turn on nodes preemptively trying to match the requested SuspendExcNodes, rather than waiting for job submissions. Thanks and Happy Thanksgiving to people in the USA
[slurm-users] slurm power save question
I've started playing with powersave and have a question about SuspendExcNodes. The documentation at https://slurm.schedmd.com/power_save.html says For example nid[10-20]:4 will prevent 4 usable nodes (i.e IDLE and not DOWN, DRAINING or already powered down) in the set nid[10-20] from being powered down. I initially interpreted that as "Slurm will try to keep 4 nodes idle on as much as possible", which would have reduced the wait time for new jobs targeting those nodes. Instead, it appears to mean "Slurm will not shut off the last 4 nodes which are idle in that partition, however it will not turn on nodes which it shut off earlier unless jobs are scheduled on them" Most notably if the 4 idle nodes will be allocated to other jobs (and so they are no idle anymore) slurm does not turn on any nodes which have been shut off earlier, so it's possible (and depending on workloads perhaps even common) to have no idle nodes on regardless of the SuspendExcNode settings. Is that how it works, or do I have anything else in my setting which is causing this unexpected-to-me behavior? I think I can live with it, but IMHO it would have been better if slurm attempted to turn on nodes preemptively trying to match the requested SuspendExcNodes, rather than waiting for job submissions. Thanks and Happy Thanksgiving to people in the USA
