Le Sat, Oct 24, 2026 at 12:32:38AM +0800, Peng Wang a écrit : > On Fri, Oct 23, 2026 at 05:00:22PM +0200, Frederic Weisbecker wrote: > > Le Sat, Oct 17, 2026 at 09:13:41PM +0800, Peng Wang a 殂rit : > > > Cloud hosts may pin vCPU threads one-to-one to host CPUs. At full > > > allocation this includes the housekeeping CPUs, which always retain > > > their non-idle tick. Operators want comparable performance for otherwise > > > equivalent vCPUs. Retaining the non-idle tick on nohz_full CPUs as well > > > avoids one placement-dependent difference in host tick interference. > > > With spare capacity, selected CPUs may stop the tick to reduce host > > > scheduler-tick interference with vCPUs. > > > > > > Allocation changes as VMs are created or destroyed while other VMs > > > remain running. A vCPU thread may be the only runnable task in either > > > state, so existing tick dependencies do not express this policy. > > > > I'm not sure I understand correctly what you are trying to solve. > > Are you saying that you want to keep the tick running when there > > is more than one vCPU per CPU? Because the scheduler already takes > > care of that. > > Hi Frederic, > > No. We pin vCPU threads one-to-one to host CPUs, without > oversubscription. By "full allocation" I meant that all allocatable > host CPUs are in use, including CPUs outside the nohz_full set. > > For example, on an eight-CPU host booted with isolcpus=nohz,1-7, > CPU0 is outside the nohz_full set: > > - With eight pinned vCPUs, CPU0 retains its non-idle tick while > CPUs1-7 may stop theirs. Retaining the tick on CPUs1-7 too avoids > this placement-dependent difference in tick interference between > otherwise equivalent vCPUs. > > - With seven or fewer pinned vCPUs, we can leave CPU0 without a > vCPU and keep them on CPUs1-7. We can then allow tick suppression > on CPUs1-7, subject to the existing dependencies. This gives > these vCPUs a common tick policy while reducing host tick > interference. > > If the vCPU on CPU0 is released while the other seven remain > running, we allow CPUs1-7 to suppress their ticks. If a new vCPU > is later assigned to CPU0, we retain the ticks on CPUs1-7 again > to avoid the difference in tick interference. > > Throughout these transitions, CPU1 can still have its original > pinned vCPU thread as its only runnable task. Its runnable task > count does not change merely because CPU0 gains or loses a vCPU, > so the scheduler's existing tick dependency cannot express this > allocation policy.
Ok now I understand what you want to do, I just don't understand why. Why it matters to turn off nohz_full on all CPUs when vCPUs run on housekeeping? > > > Add a per-CPU POLICY dependency within the boot-time nohz_full set. > > > Permission defaults to enabled. The policy can be changed without > > > taking CPUs offline. Revoking permission uses the existing asynchronous > > > kick; granting it leaves all other dependencies effective. Preserve > > > POLICY across CPU hotplug and report it in tick_stop tracing. > > > > > > Idle tick suppression, nohz_full membership, context tracking, RCU > > > offloading and housekeeping configuration are unchanged. > > > > This still keeps the overhead of nohz_full even if the tick stays on, right? > > (RCU nocb, kernel <-> user overhead, etc...). Isn't it a problem for you? > > Yes. Our setup uses only the nohz flag of isolcpus, without > additional housekeeping affinity tuning. Context tracking and > RCU callback offloading remain enabled when the tick is kept > running. > > We have observed a performance benefit from this nohz_full setup > with our database workload running on pinned vCPUs. Do you mean that your database runs better when hosts CPUs are in nohz_full? But how does this new interface help? Ie: why is it worse to have all CPUs in nohz_full except the housekeeping rather than shutting down nohz_full altogether? Or to put it another way: why is it better to have either: - all vCPUs on nohz_full host CPUs or - all vCPUs on ticking CPUs And why does it hurt to have a single vCPU running on a ticking CPU. Thanks. -- Frederic Weisbecker SUSE Labs
