Hello, we noticed some unexpected performance regressions in the scheduler by switching the guest CPU topology from "-smp 2,sockets=2,cores=1" to "-smp 2,sockets=1,cores=2".
With sockets=2,cores=1 localhost message passing (pipes, AF_UNIX etc..) runs serially at 100% CPU load of a single vcpu with optimal performance. With sockets=1,cores=2 the load is spread across both vcpus and performance is reduced. With SCHED_MC=n on older kernels the problem goes away (but that's far from ideal for heavily multithreaded workloads which then regress) because that basically disables the last part of select_idle_sibling(). On bare metal with SCHED_MC=y on any recent multicore CPU the scheduler (as expected) behaves like sockets=1,cores=2, so it won't run the tasks serially. The reason is that select_idle_sibling() can disregard the "sync" hint and all decisions done up to that point and at the last minute it can decide to move the waken task to an arbitrary idle core. To test the above theory I implemented this patch which seems to confirm the reason the tasks won't run serially anymore with sockets=1,cores=2 is select_idle_sibling() overriding the "sync" hint. You worked on the wake_affine() before so you may want to review this issue, if you agree these sync workloads should run serially even in presence of idle cores in the system. I don't know if the current behavior is on purpose but if it is, it'd be interesting to know why. This is just a RFC. To test I used this trivial program. /* * pipe-loop.c * * Copyright (C) 2019 Red Hat, Inc. * * This work is licensed under the terms of the GNU GPL, version 2. */ #include <stdio.h> #include <unistd.h> #include <sys/types.h> #include <sys/wait.h> int main(int argc, char ** argv) { char buf[1]; int n = 1000000; int pipe1[2], pipe2[2]; pipe(pipe1); pipe(pipe2); if (fork()) { while (n--) { read(pipe1[0], buf, 1); write(pipe2[1], buf, 1); } wait(NULL); } else { while (n--) { write(pipe1[1], buf, 1); read(pipe2[0], buf, 1); } } return 0; } Andrea Arcangeli (1): sched/fair: skip select_idle_sibling() in presence of sync wakeups kernel/sched/fair.c | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-)