On 08/25/2015 07:48 PM, Michael S. Tsirkin wrote: > On Fri, Aug 21, 2015 at 05:05:49PM +0800, Jason Wang wrote: >> > We used to use mmio for notification. This could be slow on some arch >> > (e.g on x86 without EPT). So this patch introduces pio bar and a pio >> > notification cap for modern device. This ability is enabled through >> > property "modern-pio-notify" for virtio pci devices and was disabled >> > by default. Management can enable when it thinks it was needed. >> > >> > Benchmarks shows almost no obvious difference with legacy device. >> > Thanks Wenli Quan <wq...@redhat.com> for the benchmarking. >> > >> > Cc: Michael S. Tsirkin <m...@redhat.com> >> > Signed-off-by: Jason Wang <jasow...@redhat.com> > I don't really care much about non-EPT hosts, but if you propose > a patch to optimize them, it should be accompanied by numbers > showing the performance difference. >
According to the test, PIO is a little bit faster than fast mmio in some specific TCP_RR case: modern device fast mmio vs modern device pio: TCP_RR: size/session/+transaction rate%/+cpu%/-+per cpu%/ 64/1/[+1.5646%]/+5.6604%/-4.3415%/ 64/25/+0.3003%/-0.4517%/+0.7486%/ ... 256/1/[+1.0046%]/[-6.5238%]/[+7.0673%]/ So the improvement is almost as much as previous patch.
