On Tue, Feb 3, 2015 at 8:21 PM, Andrey Korolyov <andrey at xdel.ru> wrote: >> These patches are to enable DPDK 1.8 only. What 'bulk processing' are you >> referring to? >> By default there is a batch size of 192 in netdev-dpdk for rx from the NIC - >> the linked >> patch doesn't change this, just the DPDK version. > > Sorry, I referred the wrong part there: bulk transmission, which is > clearly not involved in my case. The idea was that the conditionally > enabling prefetch for rx queues (BULK_ALLOC) may help somehow, but > it`s probably will mask issue instead of solving it directly. By my > understanding, strict drop rule should have a zero impact on a main > ovs thread (and this is true) and work just fine with a line rate > (this is not). > >> >> Main things to consider are to isocpu's, pin the pmd thread and keep >> everything >> on 1 NUMA socket. At 11 mpps without packet loss on that processor I suspect >> you are >> doing those things already. > > Yes, with all tuning improvements I was able to do this, but bare > Linux stack on same machine is able to handle 12Mpps and there are > absolutely no hints of what exactly is being congested.
Also both action=NORMAL & action=output:<non-dpdk port> do manage flow control in such a way that the generator side reaches line (14.8Mpps) rate on 60b data packets, though very high drop ratio persists. With action=DROP or action=output:X, where X is another dpdk port, flow control establishes somewhere at the 13Mpps. Of course, using regular host interface or NORMAL action generates a lot of context switches, mainly by miniflow_extract() and emc_..(), the difference in a syscall distribution between congested (line rate is reached) and non-congested link is unobservable.
