On 25/10/2016 22:45, Emilio G. Cota wrote: > On Tue, Oct 25, 2016 at 16:35:48 -0400, Pranith Kumar wrote: >> On Tue, Oct 25, 2016 at 4:02 PM, Paolo Bonzini <pbonz...@redhat.com> wrote: >>> >>> >>>> I've written a patch (see below) to take the per-bucket sequence locks. >>> >>> What's the performance like? >>> >> >> Applying only this patch, the perf numbers are similar to the 128 >> cache line alignment you suggested. > > That makes sense. Having a single seqlock per bucket is simple and fast; > note that bucket chains should be very short (we use good hashing and > automatic resize for this purpose).
But why do we get such worse performance in the 100% reader case? (And even more puzzling, why does Pranith's original patch improve performance instead of causing more cache misses?) Thanks, Paolo
