Resource limits are fixed because we need to make a good scheduling decision for the initial burst you’re describing (the extremely high cpu at the beginning). Some applications might also need similar cpu on restart. Your workload needs to “burst”, so setting your cpu limit to your startup peak and your cpu request to a reasonable percentage of the long term use is the best way to ensure the scheduler can put you on a node that can accommodate you. No matter what, if you want the cpu at the peak we have to schedule you somewhere you can get the peak cpu.
The longer term approach that makes this less annoying is the feedback loop between actual used resources on a node for running workloads and requested workloads, and the eviction and descheduling agents (which attempt to rebalance nodes by achuffling workloads around). For the specific case of an app where you know for sure the processes will use a fraction of the initial limit, you can always voluntarily limit your own cpu at some time after startup. That could be a side agent that puts a more restrictive cgroup limit in place on the container after it has been up a few minutes. On May 5, 2018, at 9:57 AM, Alan Christie <[email protected]> wrote: I like the idea of placing resource limits on applications running in the cluster but I wonder if there’s any advice for defining CPU “limits" that are more tolerant of application start-up behaviour? Something like the initial delay on a readiness or liveness probe for example? It just seems like a rather obvious property of any limit. The ones available are just too “hard". One example, and I’m sure this must be common to many applications, is an application that consumes a significant proportion of the CPU during initialisation but then, in its steady-state, falls back to a much lower and non-bursty behaviour. I’ve attached a screenshot of one such application. You can see that for a very short period of time, exclusive to initialisation, it consumes many more cycles than the post-initialisation stage. <PastedGraphic-1.png> During initialisation CPU demand tends to fall and memory use tends to rise. I suspect that what I’m seeing during this time is OpenShift “throttling” my app (understandable given the parameters available) and it then fails to pass through initialisation fast enough to satisfy the readiness/liveness probe and then gets restarted. Again, and again. I cannot use any sensible steady-state limit (i.e. one that prevents the normal stead-state behaviour from deviating) without the application constantly being forced to throttle and potentially reboot during initialisation. In this example I’d like to set a a perfectly reasonable CPU limit of something like 5Mi (because, after the first minute of execution it should never deviate from the steady-state level). Sadly I cannot set a low level because OpenShift will not let the application start (for reasons already explained) as its initial but very brief CPU load exceeds any “reasonable" level I set. I can get around this by defining an abnormally large cpu limit but, to me, using an “abnormal” level sort of defeats the purpose of a limit. *Aren’t resource limits missing one vital parameter, “duration" or "initial delay”.* Maybe this is beyond the resources feature and has to be deferred to something like prometheus? But can prometheus take actions rather than just monitor and alert? And, even if it could, employing prometheus may seem to some like "using a sledgehammer to crack a nut”. Any advice on permitting bursty applications within the cluster but also using limits would be most appreciated. Alan Christie _______________________________________________ users mailing list [email protected] http://lists.openshift.redhat.com/openshiftmm/listinfo/users
_______________________________________________ users mailing list [email protected] http://lists.openshift.redhat.com/openshiftmm/listinfo/users
