On Wed, Jul 07, 2021 at 11:25:50PM +0000, Wang, Wei W wrote: > On Thursday, July 8, 2021 12:45 AM, Peter Xu wrote: > > On Wed, Jul 07, 2021 at 12:45:32PM +0000, Wang, Wei W wrote: > > > Btw, what would you think if we change mutex to QemuSpin? It will also > > > reduce > > the overhead, I think. > > > > As I replied at the other place, the bottleneck we encountered is the lock > > taking > > not sleeping, so I'm afraid spinlock will have the same issue. Thanks, > > I suspect the overhead you observed might be caused by the syscalls for > mutex. Per-page syscall might be too much. > If possible, you could have another test of the 3GB guest migration using > QemuSpin.
I'm a bit confused.. Why taking a mutex with no contention would need a syscall? Please feel free to strace on a general version glibc impl of pthread mutex which qemu uses, I believe it didn't need to. Thanks, -- Peter Xu
