On Mon, 2017-02-13 at 16:34 -0800, Alexander Duyck wrote: > On Mon, Feb 13, 2017 at 4:22 PM, Eric Dumazet <eduma...@google.com> wrote: > > On Mon, Feb 13, 2017 at 3:47 PM, Alexander Duyck > > <alexander.du...@gmail.com> wrote: > > > >> Actually it depends on the use case. In the case of pktgen packets > >> they are usually dropped pretty early in the receive path. Think > >> something more along the lines of a TCP syn flood versus something > >> that would be loading up a socket. > > > > So I gave another spin an it, reducing the MTU on the sender to 500 > > instead of 1500 to triple the load (in term of packets per second) > > since the sender seems to hit some kind of limit around 30Gbit. > > > > Number of packets we process on one RX queue , and one TCP flow. > > > > Current patch series : 6.3 Mpps > > > > Which is not too bad ;) > > > > This number does not change putting your __page_frag_cache_drain() trick. > > Well the __page_frag_cache_drain by itself isn't going to add much of > anything. You really aren't going to be using it except for in the > slow path. I was talking about the bulk page count update by itself. > All __page_frag_cache_drain gets you is it cleans up the > page_frag_sub(n-1); put_page(1); code calls. > > The approach I took for the Intel drivers isn't too different than > what we did for the skbuff page frag cache. Basically I update the > page count once every 65535 frames setting it to 64K, and holding no > more than 65535 references in the pagecnt_bias. Also in my code I > don't update the count until after the first recycle call since the > highest likelihood of us discarding the frame is after the first > allocation anyway so we wait until the first recycle to start > repopulating the count.
Alex, be assured that I implemented the full thing, of course. ( The pagecnt_bias field, .. refilled every 16K rounds ) And I repeat : I got _no_ convincing results to justify this patch being sent in this series. This series is not about adding 1% or 2% performance changes, but adding robustness to RX allocations.
