> -----Original Message----- > From: Kevin Traynor [mailto:ktray...@redhat.com] > Sent: Friday, 17 February, 2017 17:38 > > If there are multiple queues in a poll list and only one has packets, > the cycles polling the empty queues for packets will be counted in the > processing time - whereas you'd expect them to be in the idle time. > > Is there much of a performance hit if the cycle_count_intermediate() is > moved into the poll_cnt loop? OTOH, how much would the idle queues skew > the counters? If it's a small amount it may be better than doing a > cycle_count per q.
In our testing an idle iteration over 8 vhostuser and one eth port consumes on average 600 cycles, i.e. ~60 cycles per port. Compared to the typical cost for processing a packet in the range of 1000 cycles, the overhead of polling idle queues is not significant, unless a PMD were to poll hundreds of queues. Also we have seen that an rte_rdtsc() call can cost up to 20 cycles. Below are samples of how the reported load scales with the processed packets in our setup with VXLAN underlay (2 ports busy, 7 ports idle): pkts/s busy cycles/pkt 0 0 - 149216 4.89% 788.6220446 298432 9.61% 774.8487964 447574 14.62% 786.069622 596832 19.20% 774.3068502 745965 23.80% 767.6537934 895200 28.89% 776.7226977 1042880 33.93% 782.986241 1193243 38.14% 769.2920729 1491792 47.86% 772.0912312 1788811.5 56.96% 766.3735553 2086880 66.59% 767.9209188 2385772 75.48% 761.4203964 2684853.5 85.09% 762.6955575 2979551 93.44% 754.7385526 3274118 99.07% 728.1760908 The relation between processed packets and reported processing load is almost perfectly linearly and not significantly disturbed by the presence of idle ports. The slight decrease of cycles/pkt is due to the increased amortization of batch overhead with the growing size of batches. Based on this data, I would propose to keep the original simple approach and not do intermediate cycle sampling per port. BR, Jan _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
