Hello Maxime, > Le 9 nov. 2016 ? 15:51, Maxime Coquelin <maxime.coquelin at redhat.com> a > ?crit : > > Hi Pierre, > > On 11/09/2016 01:42 PM, Pierre Pfister (ppfister) wrote: >> Hello Maxime, >> >> Sorry for the late reply. >> >> >>> Le 8 nov. 2016 ? 10:44, Maxime Coquelin <maxime.coquelin at redhat.com> a >>> ?crit : >>> >>> Hi Pierre, >>> >>> On 11/08/2016 10:31 AM, Pierre Pfister (ppfister) wrote: >>>> Current virtio driver advertises VERSION_1 support, >>>> but does not handle device's VERSION_1 support when >>>> sending packets (it looks for ANY_LAYOUT feature, >>>> which is absent). >>>> >>>> This patch enables 'can_push' in tx path when VERSION_1 >>>> is advertised by the device. >>>> >>>> This significantly improves small packets forwarding rate >>>> towards devices advertising VERSION_1 feature. >>> I think it depends whether offloading is enabled or not. >>> If no offloading enabled, I measured significant drop. >>> Indeed, when no offloading is enabled, the Tx path in Virtio >>> does not access the virtio header before your patch, as the header is >>> memset to zero at device init time. >>> With your patch, it gets memset to zero at every transmit in the hot >>> path. >> >> Right. On the virtio side that is true, but on the device side, we have to >> access the header anyway. > No more now, if no offload features have been negotiated. > I have done a patch that landed in v16.11 to skip header parsing in > this case. > That said, we still have to access its descriptor. > >> And accessing two descriptors (with the address resolution and memory fetch >> which comes with it) >> is a costy operation compared to a single one. >> In the case indirect descriptors are used, this is 1 desc access instead or >> 3. > I agree this is far from being optimal. > >> And in the case chained descriptors are used, this doubles the number of >> packets that you can put in your queue. >> >> Those are the results in my PHY -> VM (testpmd) -> PHY setup >> Traffic is flowing bidirectionally. Numbers are for lossless-rates. >> >> When chained buffers are used for dpdk's TX: 2x2.13Mpps >> When indirect descriptors are used for dpdk's TX: 2x2.38Mpps >> When shallow buffers are used for dpdk's TX (with this patch): 2x2.42Mpps > When I tried it, I also did PVP 0% benchmark, and I got opposite results. > Chained and indirect cases were significantly better. > > My PVP setup was using a single NIC and single Virtio PMD, and NIC2VM > forwarding was IO mode done with testpmd on host, and Rx->Tx forwarding > was macswap mode on guest side. > > I also saw some perf regression when running simple tespmd test on both > ends. > > Yuanhan, did you run some benchmark with your series enabling > ANY_LAYOUT?
It was enabled. But the specs specify that VERSION_1 includes ANY_LAYOUT. Therefor, Qemu removes ANY_LAYOUT when VERSION_1 is set. We can keep arguing about which is fastest. I guess we have different setups and different results, so we probably are deadlocked here. But in any case, the current code is inconsistent, as it uses single descriptor when ANY_LAYOUT is set, but not when VERSION_1 is set. I believe it makes sense to use single-descriptor any time it is possible, but you are free to think otherwise. Please make a call and make the code consistent (removes single-descriptors all together, or use them when VERSION_1 is set too). Otherwise it just creates yet-another testing headache. Thanks, - Pierre > >> >> I must also note that qemu 2.5 does not seem to deal with VERSION_1 and >> ANY_LAYOUT correctly. >> The patch I am proposing here works for qemu 2.7, but with qemu 2.5, testpmd >> still behaves as if ANY_LAYOUT (or VERSION_1) was not available. This is not >> catastrophic. But just note that you will not see performance in some cases >> with qemu 2.5. > > Thanks for the info. > > Regards, > Maxime
