Messages in this thread |  | | Subject | Re: [PATCH] sched, fair: Allow a small degree of load imbalance between SD_NUMA domains v2 | From | Valentin Schneider <> | Date | Fri, 20 Dec 2019 15:32:53 +0000 |
| |
On 20/12/2019 14:22, Mel Gorman wrote: >> Now, I have to say I'm not sold on the idle_cpus thing, I'd much rather use >> the number of runnable tasks. We are setting up a threshold for how far we >> are willing to ignore imbalances; if we have overloaded CPUs we *really* >> should try to solve this. Number of tasks is the safer option IMO: when we >> do have one task per CPU, it'll be the same as if we had used idle_cpus, and >> when we don't have one task per CPU we'll load-balance more often that we >> would have with idle_cpus. >> > > I couldn't convince myself to really push back hard on the sum_nr_runnable > versus idle_cpus. If the local group has spare capacity and the busiest > group has multiple tasks stacked on CPUs then it's most likely due to > CPU affinity.
Not necessarily, for instance wakeup balancing (select_idle_sibling()) could end up packing stuff within a node if said node spans more than one LLC domain, which IIRC is the case on some AMD chips.
Or, still with the same LLC < node topology, you could start with the node being completely utilized, then some tasks on some LLC domains terminate but there's an LLC that still has a bunch of tasks running, and then you're left with an imbalance between LLC domains that the wakeup balance cannot solve.
> In that case, there is no guarantee tasks can move to the > local group either. In that case, the difference between sum_nr_running > and idle_cpus is almost moot. There may be common use cases where the > distinction really matters but right now, I'm at the point where I think > such a change could be a separate patch with the use case included and > supporting data on why it must be sum_nr_running. Right now, I feel it's > mostly a cosmetic issue given the context and intent of the patch. >
Let me spin it this way: do we need to push this ignoring of the imbalance as far as possible, or are we okay with it only happening when there's just a few tasks running? The latter is achieved with sum_nr_running and is the safer option IMO.
|  |