RE: Reworking push/pull [LWN.net]

RE: Reworking push/pull

Posted Sep 19, 2017 1:35 UTC (Tue) by iamsrp (subscriber, #84011)
Parent article: Notes from the LPC scheduler microconference

Could the problem be solved via delegation? I.e. something like:

CPU[k] goes idle
CPU[k] walks the other CPUs [0..n) to look for the first idle one (possibly itself); CPU[i]
CPU[k] sends an IPI to CPU[i] requesting that it do the search for a busy CPU
CPU[i] walks the CPUs [0..n) looking for an overloaded one, CPU[b], and "forwards" the IPI from CPU[k] to CPU[b]
CPU[i] also remembers "recently" signaled overloaded CPUs and so avoids forwarding subsequent IPIs to them

Here [i] can clearly change, and memory of recently signaled CPUs isn't shared. It's also not going to be as quick owing to the extra hop, and it may well still have pathological cases, but is possibly better..?