Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Michael S. Tsirkin m...@redhat.com wrote on 10/12/2010 10:39:07 PM: Sorry for the delay, I was sick last couple of days. The results with your patch are (%'s over original code): Code BW% CPU% RemoteCPU MQ (#txq=16) 31.4% 38.42% 6.41% MQ+MST (#txq=16) 28.3% 18.9% -10.77% The patch helps CPU utilization but didn't help single stream drop. Thanks, What other shared TX/RX locks are there? In your setup, is the same macvtap socket structure used for RX and TX? If yes this will create cacheline bounces as sk_wmem_alloc/sk_rmem_alloc share a cache line, there might also be contention on the lock in sk_sleep waitqueue. Anything else? The patch is not introducing any locking (both vhost and virtio-net). The single stream drop is due to different vhost threads handling the RX/TX traffic. I added a heuristic (fuzzy) to determine if more than one flow is being used on the device, and if not, use vhost[0] for both tx and rx (vhost_poll_queue figures this out before waking up the suitable vhost thread). Testing shows that single stream performance is as good as the original code. __ #txqs = 2 (#vhosts = 3) # BW1 BW2 (%) CPU1CPU2 (%) RCPU1 RCPU2 (%) __ 1 77344 74973 (-3.06) 172 143 (-16.86) 358 324 (-9.49) 2 20924 21107 (.87) 107 103 (-3.73)220 217 (-1.36) 4 21629 32911 (52.16) 214 391 (82.71)446 616 (38.11) 8 21678 34359 (58.49) 428 845 (97.42)892 1286 (44.17) 1622046 34401 (56.04) 841 1677 (99.40) 17852585 (44.81) 2422396 35117 (56.80) 12722447 (92.37) 26673863 (44.84) 3222750 35158 (54.54) 17193233 (88.07) 35695143 (44.10) 4023041 35345 (53.40) 22193970 (78.90) 44786410 (43.14) 4823209 35219 (51.74) 27074685 (73.06) 53867684 (42.66) 6423215 35209 (51.66) 36396195 (70.23) 720610218 (41.79) 8023443 35179 (50.06) 46337625 (64.58) 905112745 (40.81) 9624006 36108 (50.41) 56359096 (61.41) 10864 15283 (40.67) 128 23601 35744 (51.45) 747512104 (61.92) 14495 20405 (40.77) __ SUM: BW: (37.6) CPU: (69.0) RCPU: (41.2) __ #txqs = 8 (#vhosts = 5) # BW1 BW2(%) CPU1 CPU2 (%) RCPU1 RCPU2 (%) __ 1 77344 75341 (-2.58) 172 171 (-.58) 358 356 (-.55) 2 20924 26872 (28.42) 107 135 (26.16)220 262 (19.09) 4 21629 33594 (55.31) 214 394 (84.11)446 615 (37.89) 8 21678 39714 (83.19) 428 949 (121.72) 892 1358 (52.24) 1622046 39879 (80.88) 841 1791 (112.96) 17852737 (53.33) 2422396 38436 (71.61) 12722111 (65.95) 26673453 (29.47) 3222750 38776 (70.44) 17193594 (109.07) 35695421 (51.89) 4023041 38023 (65.02) 22194358 (96.39) 44786507 (45.31) 4823209 33811 (45.68) 27074047 (49.50) 53866222 (15.52) 6423215 30212 (30.13) 36393858 (6.01)72065819 (-19.24) 8023443 34497 (47.15) 46337214 (55.70) 905110776 (19.05) 9624006 30990 (29.09) 56355731 (1.70)10864 8799 (-19.00) 128 23601 29413 (24.62) 74757804 (4.40)14495 11638 (-19.71) __ SUM: BW: (40.1) CPU: (35.7) RCPU: (4.1) ___ The SD numbers are also good (same table as before, but SD instead of CPU: __ #txqs = 2 (#vhosts = 3) # BW% SD1 SD2 (%)RSD1 RSD2 (%) __ 1 -3.06)5 4 (-20.00) 21 19 (-9.52) 2 .87 6 6 (0) 27 27 (0) 4 52.16 26 32 (23.07) 108 103 (-4.62) 8 58.49 103 146 (41.74)431 445 (3.24) 1656.04 407 514 (26.28)1729 1586 (-8.27) 2456.80 934 1161 (24.30) 3916 3665 (-6.40) 3254.54 16682160 (29.49) 6925 6872 (-.76) 4053.40 26553317 (24.93) 1071210707 (-.04) 4851.74 39204486 (14.43) 1559814715 (-5.66) 6451.66 70968250 (16.26) 2809927211 (-3.16) 8050.06 11240 12586 (11.97) 4391342070 (-4.19) 9650.41 16342 16976
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Thu, Oct 14, 2010 at 01:28:58PM +0530, Krishna Kumar2 wrote: Michael S. Tsirkin m...@redhat.com wrote on 10/12/2010 10:39:07 PM: Sorry for the delay, I was sick last couple of days. The results with your patch are (%'s over original code): Code BW% CPU% RemoteCPU MQ (#txq=16) 31.4% 38.42% 6.41% MQ+MST (#txq=16) 28.3% 18.9% -10.77% The patch helps CPU utilization but didn't help single stream drop. Thanks, What other shared TX/RX locks are there? In your setup, is the same macvtap socket structure used for RX and TX? If yes this will create cacheline bounces as sk_wmem_alloc/sk_rmem_alloc share a cache line, there might also be contention on the lock in sk_sleep waitqueue. Anything else? The patch is not introducing any locking (both vhost and virtio-net). The single stream drop is due to different vhost threads handling the RX/TX traffic. I added a heuristic (fuzzy) to determine if more than one flow is being used on the device, and if not, use vhost[0] for both tx and rx (vhost_poll_queue figures this out before waking up the suitable vhost thread). Testing shows that single stream performance is as good as the original code. ... This approach works nicely for both single and multiple stream. Does this look good? Thanks, - KK Yes, but I guess it depends on the heuristic :) What's the logic? -- MST -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Michael S. Tsirkin m...@redhat.com What other shared TX/RX locks are there? In your setup, is the same macvtap socket structure used for RX and TX? If yes this will create cacheline bounces as sk_wmem_alloc/sk_rmem_alloc share a cache line, there might also be contention on the lock in sk_sleep waitqueue. Anything else? The patch is not introducing any locking (both vhost and virtio-net). The single stream drop is due to different vhost threads handling the RX/TX traffic. I added a heuristic (fuzzy) to determine if more than one flow is being used on the device, and if not, use vhost[0] for both tx and rx (vhost_poll_queue figures this out before waking up the suitable vhost thread). Testing shows that single stream performance is as good as the original code. ... This approach works nicely for both single and multiple stream. Does this look good? Thanks, - KK Yes, but I guess it depends on the heuristic :) What's the logic? I define how recently a txq was used. If 0 or 1 txq's were used recently, use vq[0] (which also handles rx). Otherwise, use multiple txq (vq[1-n]). The code is: /* * Algorithm for selecting vq: * * ConditionReturn * RX vqvq[0] * If all txqs unused vq[0] * If one txq used, and new txq is same vq[0] * If one txq used, and new txq is differentvq[vq-qnum] * If 1 txqs used vq[vq-qnum] * Where used means the txq was used in the last 'n' jiffies. * * Note: locking is not required as an update race will only result in * a different worker being woken up. */ static inline struct vhost_virtqueue *vhost_find_vq(struct vhost_poll *poll) { if (poll-vq-qnum) { struct vhost_dev *dev = poll-vq-dev; struct vhost_virtqueue *vq = dev-vqs[0]; unsigned long max_time = jiffies - 5; /* Some macro needed */ unsigned long *table = dev-jiffies; int i, used = 0; for (i = 0; i dev-nvqs - 1; i++) { if (time_after_eq(table[i], max_time) ++used 1) { vq = poll-vq; break; } } table[poll-vq-qnum - 1] = jiffies; return vq; } /* RX is handled by the same worker thread */ return poll-vq; } void vhost_poll_queue(struct vhost_poll *poll) { struct vhost_virtqueue *vq = vhost_find_vq(poll); vhost_work_queue(vq, poll-work); } Since poll batches packets, find_vq does not seem to add much to the CPU utilization (or BW). I am sure that code can be optimized much better. The results I sent in my last mail were without your use_mm patch, and the only tuning was to make vhost threads run on only cpus 0-3 (though the performance is good even without that). I will test it later today with the use_mm patch too. Thanks, - KK -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Krishna Kumar2/India/IBM wrote on 10/14/2010 02:34:01 PM: void vhost_poll_queue(struct vhost_poll *poll) { struct vhost_virtqueue *vq = vhost_find_vq(poll); vhost_work_queue(vq, poll-work); } Since poll batches packets, find_vq does not seem to add much to the CPU utilization (or BW). I am sure that code can be optimized much better. The results I sent in my last mail were without your use_mm patch, and the only tuning was to make vhost threads run on only cpus 0-3 (though the performance is good even without that). I will test it later today with the use_mm patch too. There's a significant reduction in CPU/SD utilization with your patch. Following is the performance of ORG vs MQ+mm patch: _ Org vs MQ+mm patch txq=2 # BW% CPU/RCPU% SD/RSD% _ 1 2.26-1.16.27 -20.00 0 2 35.07 29.9021.81 0 -11.11 4 55.03 84.5737.66 26.92 -4.62 8 73.16 118.69 49.21 45.63 -.46 1677.43 98.8147.89 24.07 -7.80 2471.59 105.18 48.44 62.84 18.18 3270.91 102.38 47.15 49.22 8.54 4063.26 90.5841.00 85.27 37.33 4845.25 45.9911.23 14.31 -12.91 6442.78 41.825.50 .43-25.12 8031.40 7.31 -18.6915.78 -11.93 9627.60 7.79 -18.5417.39 -10.98 128 23.46 -11.89 -34.41-.41 -25.53 _ BW: 40.2 CPU/RCPU: 29.9,-2.2 SD/RSD: 12.0,-15.6 Following is the performance of MQ vs MQ+mm patch: _ MQ vs MQ+mm patch # BW% CPU% RCPU%SD% RSD% _ 1 4.98-.58 .84 -20.000 2 5.17 2.96 2.29 0 -4.00 4 -.18 .25 -.16 3.12 .98 8 -5.47-1.36 -1.98 17.1816.57 16-1.90-6.64 -3.54 -14.83 -12.12 24-.01 23.63 14.65 57.6146.64 32 .27 -3.19 -3.11-22.98 -22.91 40-1.06-2.96 -2.96-4.18-4.10 48-.28 -2.34 -3.71-2.41-3.81 64 9.71 33.77 30.6581.4477.09 80-10.69-31.07-31.70 -29.22 -29.88 96-1.14 5.98 .56 -11.57 -16.14 128 -.93 -15.60 -18.31 -19.89 -22.65 _ BW: 0 CPU/RCPU: -4.2,-6.1 SD/RSD: -13.1,-15.6 _ Each test case is for 60 secs, sum over two runs (except when number of netperf sessions is 1, which has 7 runs of 10 secs each), numcpus=4, numtxqs=8, etc. No tuning other than taskset each vhost to cpus 0-3. Thanks, - KK -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Krishna Kumar2/India/IBM wrote on 10/14/2010 05:47:54 PM: Sorry, it should read txq=8 below. - KK There's a significant reduction in CPU/SD utilization with your patch. Following is the performance of ORG vs MQ+mm patch: _ Org vs MQ+mm patch txq=2 # BW% CPU/RCPU% SD/RSD% _ 1 2.26-1.16.27 -20.00 0 2 35.07 29.9021.81 0 -11.11 4 55.03 84.5737.66 26.92 -4.62 8 73.16 118.69 49.21 45.63 -.46 1677.43 98.8147.89 24.07 -7.80 2471.59 105.18 48.44 62.84 18.18 3270.91 102.38 47.15 49.22 8.54 4063.26 90.5841.00 85.27 37.33 4845.25 45.9911.23 14.31 -12.91 6442.78 41.825.50 .43-25.12 8031.40 7.31 -18.6915.78 -11.93 9627.60 7.79 -18.5417.39 -10.98 128 23.46 -11.89 -34.41-.41 -25.53 _ BW: 40.2 CPU/RCPU: 29.9,-2.2 SD/RSD: 12.0,-15.6 Following is the performance of MQ vs MQ+mm patch: _ MQ vs MQ+mm patch # BW% CPU% RCPU%SD% RSD% _ 1 4.98-.58 .84 -20.000 2 5.17 2.96 2.29 0 -4.00 4 -.18 .25 -.16 3.12 .98 8 -5.47-1.36 -1.98 17.1816.57 16-1.90-6.64 -3.54 -14.83 -12.12 24-.01 23.63 14.65 57.6146.64 32 .27 -3.19 -3.11-22.98 -22.91 40-1.06-2.96 -2.96-4.18-4.10 48-.28 -2.34 -3.71-2.41-3.81 64 9.71 33.77 30.6581.4477.09 80-10.69-31.07-31.70 -29.22 -29.88 96-1.14 5.98 .56 -11.57 -16.14 128 -.93 -15.60 -18.31 -19.89 -22.65 _ BW: 0 CPU/RCPU: -4.2,-6.1 SD/RSD: -13.1,-15.6 _ Each test case is for 60 secs, sum over two runs (except when number of netperf sessions is 1, which has 7 runs of 10 secs each), numcpus=4, numtxqs=8, etc. No tuning other than taskset each vhost to cpus 0-3. Thanks, - KK -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Michael S. Tsirkin m...@redhat.com wrote on 10/06/2010 07:04:31 PM: On Fri, Sep 17, 2010 at 03:33:07PM +0530, Krishna Kumar wrote: For 1 TCP netperf, I ran 7 iterations and summed it. Explanation for degradation for 1 stream case: I thought about possible RX/TX contention reasons, and I realized that we get/put the mm counter all the time. So I write the following: I haven't seen any performance gain from this in a single queue case, but maybe this will help multiqueue? Sorry for the delay, I was sick last couple of days. The results with your patch are (%'s over original code): Code BW% CPU% RemoteCPU MQ (#txq=16) 31.4% 38.42% 6.41% MQ+MST (#txq=16) 28.3% 18.9% -10.77% The patch helps CPU utilization but didn't help single stream drop. Thanks, - KK -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Tuesday 05 October 2010, Krishna Kumar2 wrote: After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. I did two more tests: 1. Pin vhosts to same CPU: - BW drop is much lower for 1 stream case (- 5 to -8% range) - But performance is not so high for more sessions. 2. Changed vhost to be single threaded: - No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%). - BW degrades after that, all the way till 128 netperf sessions. - But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions): txq=4: BW: (-2.3) CPU: (-16.5)RCPU: (-5.3) txq=16: BW: (-1.9) CPU: (-24.9)RCPU: (-9.6) I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: ___ #netperf ORG NEW BW (#retr)BW (#retr) ___ 1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 12823251 (2) 31676 (3185) ___ This smells like it could be related to a problem that Ben Greear found recently (see macvlan: Enable qdisc backoff logic). When the hardware is busy, used to just drop the packet. With Ben's patch, we return -EAGAIN to qemu (or vhost-net) to trigger a resend. I suppose what we really should do is feed that condition back to the guest network stack and implement the backoff in there. Arnd -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Fri, Sep 17, 2010 at 03:33:07PM +0530, Krishna Kumar wrote: For 1 TCP netperf, I ran 7 iterations and summed it. Explanation for degradation for 1 stream case: I thought about possible RX/TX contention reasons, and I realized that we get/put the mm counter all the time. So I write the following: I haven't seen any performance gain from this in a single queue case, but maybe this will help multiqueue? Thanks, Michael S. Tsirkin (2): vhost: put mm after thread stop vhost-net: batch use/unuse mm drivers/vhost/net.c |7 --- drivers/vhost/vhost.c | 16 ++-- 2 files changed, 10 insertions(+), 13 deletions(-) -- 1.7.3-rc1 -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Michael S. Tsirkin m...@redhat.com wrote on 10/06/2010 07:04:31 PM: Michael S. Tsirkin m...@redhat.com 10/06/2010 07:04 PM To Krishna Kumar2/India/i...@ibmin cc ru...@rustcorp.com.au, da...@davemloft.net, kvm@vger.kernel.org, a...@arndb.de, net...@vger.kernel.org, a...@redhat.com, anth...@codemonkey.ws Subject Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net On Fri, Sep 17, 2010 at 03:33:07PM +0530, Krishna Kumar wrote: For 1 TCP netperf, I ran 7 iterations and summed it. Explanation for degradation for 1 stream case: I thought about possible RX/TX contention reasons, and I realized that we get/put the mm counter all the time. So I write the following: I haven't seen any performance gain from this in a single queue case, but maybe this will help multiqueue? Great! I am on vacation tomorrow, but will test with this patch tomorrow night. Thanks, - KK -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Arnd Bergmann a...@arndb.de wrote on 10/06/2010 05:49:00 PM: I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: ___ #netperf ORG NEW BW (#retr)BW (#retr) ___ 1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 12823251 (2) 31676 (3185) ___ This smells like it could be related to a problem that Ben Greear found recently (see macvlan: Enable qdisc backoff logic). When the hardware is busy, used to just drop the packet. With Ben's patch, we return -EAGAIN to qemu (or vhost-net) to trigger a resend. I suppose what we really should do is feed that condition back to the guest network stack and implement the backoff in there. Thanks for the pointer. I will take a look at this as I hadn't seen this patch earlier. Is there any way to figure out if this is the issue? Thanks, - KK -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Michael S. Tsirkin m...@redhat.com wrote on 10/05/2010 11:53:23 PM: Any idea where does this come from? Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs? 4. Identify reasons for single netperf BW regression. After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. Right. Can we fix it? I am not sure how to. My initial patch had one thread but gave small gains and ran into limitations once number of sessions became large. I did two more tests: 1. Pin vhosts to same CPU: - BW drop is much lower for 1 stream case (- 5 to -8% range) - But performance is not so high for more sessions. 2. Changed vhost to be single threaded: - No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%). - BW degrades after that, all the way till 128 netperf sessions. - But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions): txq=4: BW: (-2.3) CPU: (-16.5)RCPU: (-5.3) txq=16: BW: (-1.9) CPU: (-24.9)RCPU: (-9.6) I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: Hmm, ok, and do you see any errors? I haven't seen any in any statistics, messages, etc. Also no retranmissions for txq=1. Single netperf case didn't have any retransmissions so that is not the cause for drop. I tested ixgbe (MQ): ___ #netperf ixgbe ixgbe (pin intrs to cpu#0 on both server/client) BW (#retr) BW (#retr) ___ 1 3567 (117) 6000 (251) 2 4406 (477) 6298 (725) 4 6119 (1085) 7208 (3387) 8 6595 (4276) 7381 (15296) 16 6651 (11651)6856 (30394) Interesting. You are saying we get much more retransmissions with physical nic as well? Yes, with ixgbe. I re-ran with 16 netperfs running for 15 secs on both ixgbe and cxgb3 just now to reconfirm: ixgbe: BW: 6186.85 SD/Remote: 135.711, 339.376 CPU/Remote: 79.99, 200.00, Retrans: 545 cxgb3: BW: 8051.07 SD/Remote: 144.416, 260.487 CPU/Remote: 110.88, 200.00, Retrans: 0 However 64 netperfs for 30 secs gave: ixgbe: BW: 6691.12 SD/Remote: 8046.617, 5259.992 CPU/Remote: 1223.86, 799.97, Retrans: 1424 cxgb3: BW: 7799.16 SD/Remote: 2589.875, 4317.013 CPU/Remote: 480.39 800.64, Retrans: 649 # ethtool -i eth4 driver: ixgbe version: 2.0.84-k2 firmware-version: 0.9-3 bus-info: :1f:00.1 # ifconfig output: RX packets:783241 errors:0 dropped:0 overruns:0 frame:0 TX packets:689533 errors:0 dropped:0 overruns:0 carrier:0 collisions:0 txqueuelen:1000 # lspci output: 1f:00.0 Ethernet controller: Intel Corporation 82599EB 10-Gigabit Network Connec tion (rev 01) Subsystem: Intel Corporation Ethernet Server Adapter X520-2 Flags: bus master, fast devsel, latency 0, IRQ 30 Memory at 9890 (64-bit, prefetchable) [size=512K] I/O ports at 2020 [size=32] Memory at 98a0 (64-bit, prefetchable) [size=16K] Capabilities: [40] Power Management version 3 Capabilities: [50] MSI: Enable- Count=1/1 Maskable+ 64bit+ Capabilities: [70] MSI-X: Enable+ Count=64 Masked- Capabilities: [a0] Express Endpoint, MSI 00 Capabilities: [100] Advanced Error Reporting Capabilities: [140] Device Serial Number 00-1b-21-ff-ff-40-4a-b4 Capabilities: [150] Alternative Routing-ID Interpretation (ARI) Capabilities: [160] Single Root I/O Virtualization (SR-IOV) Kernel driver in use: ixgbe Kernel modules: ixgbe I haven't done this right now since I don't have a setup. I guess it would be limited by wire speed and gains may not be there. I will try to do this later when I get the setup. OK but at least need to check that it does not hurt things. Yes, sure. Summary: 1. Average BW increase for regular I/O is best for #txq=16 with the least CPU utilization increase. 2. The average BW for 512 byte I/O is best for lower #txq=2. For higher #txqs, BW increased only after a particular #netperf sessions - in my testing that limit was 32 netperf sessions. 3. Multiple txq for guest by itself doesn't seem to have any issues. Guest CPU% increase is slightly higher than BW improvement. I think it is true for all mq drivers since more paths run in parallel upto the device instead of sleeping and allowing one thread to send all packets via qdisc_restart. 4. Having high number of txqs gives better gains
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Wednesday 06 October 2010 19:14:42 Krishna Kumar2 wrote: Arnd Bergmann a...@arndb.de wrote on 10/06/2010 05:49:00 PM: I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: ___ #netperf ORG NEW BW (#retr)BW (#retr) ___ 1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 12823251 (2) 31676 (3185) ___ This smells like it could be related to a problem that Ben Greear found recently (see macvlan: Enable qdisc backoff logic). When the hardware is busy, used to just drop the packet. With Ben's patch, we return -EAGAIN to qemu (or vhost-net) to trigger a resend. I suppose what we really should do is feed that condition back to the guest network stack and implement the backoff in there. Thanks for the pointer. I will take a look at this as I hadn't seen this patch earlier. Is there any way to figure out if this is the issue? I think a good indication would be if this changes with/without the patch, and if you see -EAGAIN in qemu with the patch applied. Arnd -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Wed, Oct 06, 2010 at 11:13:31PM +0530, Krishna Kumar2 wrote: Michael S. Tsirkin m...@redhat.com wrote on 10/05/2010 11:53:23 PM: Any idea where does this come from? Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs? 4. Identify reasons for single netperf BW regression. After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. Right. Can we fix it? I am not sure how to. My initial patch had one thread but gave small gains and ran into limitations once number of sessions became large. Sure. We will need multiple RX queues, and have a single thread handle a TX and RX pair. Then we need to make sure packets from a given flow on TX land on the same thread on RX. As flows can be hashed differently, for this to work we'll have to expose this info in host/guest interface. But since multiqueue implies host/guest ABI changes anyway, this point is moot. BTW, an interesting approach could be using bonding and multiple virtio-net interfaces. What are the disadvantages of such a setup? One advantage is it can be made to work in existing guests. I did two more tests: 1. Pin vhosts to same CPU: - BW drop is much lower for 1 stream case (- 5 to -8% range) - But performance is not so high for more sessions. 2. Changed vhost to be single threaded: - No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%). - BW degrades after that, all the way till 128 netperf sessions. - But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions): txq=4: BW: (-2.3) CPU: (-16.5)RCPU: (-5.3) txq=16: BW: (-1.9) CPU: (-24.9)RCPU: (-9.6) I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: Hmm, ok, and do you see any errors? I haven't seen any in any statistics, messages, etc. Herbert, could you help out debugging this increase in retransmissions please? Older mail on netdev in this thread has some numbers that seem to imply that we start hitting retransmissions much more as # of flows goes up. Also no retranmissions for txq=1. While it's nice that we have this parameter, the need to choose between single stream and multi stream performance when you start the vm makes this patch much less interesting IMHO. -- MST -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Michael S. Tsirkin m...@redhat.com wrote on 09/19/2010 06:14:43 PM: Could you document how exactly do you measure multistream bandwidth: netperf flags, etc? All results were without any netperf flags or system tuning: for i in $list do netperf -c -C -l 60 -H 192.168.122.1 /tmp/netperf.$$.$i done wait Another script processes the result files. It also displays the start time/end time of each iteration to make sure skew due to parallel netperfs is minimal. I changed the vhost functionality once more to try to get the best model, the new model being: 1. #numtxqs=1 - #vhosts=1, this thread handles both RX/TX. 2. #numtxqs1 - vhost[0] handles RX and vhost[1-MAX] handles TX[0-n], where MAX is 4. Beyond numtxqs=4, the remaining TX queues are handled by vhost threads in round-robin fashion. Results from here on are with these changes, and only tuning is to set each vhost's affinity to CPUs[0-3] (taskset -p f vhost-pids). Any idea where does this come from? Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs? 4. Identify reasons for single netperf BW regression. After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. I did two more tests: 1. Pin vhosts to same CPU: - BW drop is much lower for 1 stream case (- 5 to -8% range) - But performance is not so high for more sessions. 2. Changed vhost to be single threaded: - No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%). - BW degrades after that, all the way till 128 netperf sessions. - But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions): txq=4: BW: (-2.3) CPU: (-16.5)RCPU: (-5.3) txq=16: BW: (-1.9) CPU: (-24.9)RCPU: (-9.6) I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: ___ #netperf ORG NEW BW (#retr)BW (#retr) ___ 1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 12823251 (2) 31676 (3185) ___ Single netperf case didn't have any retransmissions so that is not the cause for drop. I tested ixgbe (MQ): ___ #netperf ixgbe ixgbe (pin intrs to cpu#0 on both server/client) BW (#retr) BW (#retr) ___ 1 3567 (117) 6000 (251) 2 4406 (477) 6298 (725) 4 6119 (1085) 7208 (3387) 8 6595 (4276) 7381 (15296) 16 6651 (11651)6856 (30394) ___ 5. Test perf in more scenarious: small packets 512 byte packets - BW drop for upto 8 (sometimes 16) netperf sessions, but increases with #sessions: ___ # BW1 BW2 (%) CPU1CPU2 (%)RCPU1 RCPU2 (%) ___ 1 40433800 (-6.0) 50 50 (0) 86 98 (13.9) 2 83587485 (-10.4)153 178 (16.3) 230 264 (14.7) 4 20664 13567 (-34.3) 448 490 (9.3) 530 624 (17.7) 8 25198 17590 (-30.1) 967 1021 (5.5) 10851257 (15.8) 16 23791 24057 (1.1) 19042220 (16.5) 21562578 (19.5) 24 23055 26378 (14.4)28073378 (20.3) 32253901 (20.9) 32 22873 27116 (18.5)37484525 (20.7) 43075239 (21.6) 40 22876 29106 (27.2)47055717 (21.5) 53886591 (22.3) 48 23099 31352 (35.7)56426986 (23.8) 64758085 (24.8) 64 22645 30563 (34.9)75279027 (19.9) 861910656 (23.6) 80 22497 31922 (41.8)937511390 (21.4)10736 13485 (25.6) 96 22509 32718 (45.3)11271 13710 (21.6)12927 16269 (25.8) 128 22255 32397 (45.5)15036 18093 (20.3)17144 21608 (26.0) ___ SUM:BW: (16.7) CPU: (20.6) RCPU: (24.3) ___ host - guest ___ # BW1 BW2 (%) CPU1CPU2 (%)RCPU1
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Tue, Oct 05, 2010 at 04:10:00PM +0530, Krishna Kumar2 wrote: Michael S. Tsirkin m...@redhat.com wrote on 09/19/2010 06:14:43 PM: Could you document how exactly do you measure multistream bandwidth: netperf flags, etc? All results were without any netperf flags or system tuning: for i in $list do netperf -c -C -l 60 -H 192.168.122.1 /tmp/netperf.$$.$i done wait Another script processes the result files. It also displays the start time/end time of each iteration to make sure skew due to parallel netperfs is minimal. I changed the vhost functionality once more to try to get the best model, the new model being: 1. #numtxqs=1 - #vhosts=1, this thread handles both RX/TX. 2. #numtxqs1 - vhost[0] handles RX and vhost[1-MAX] handles TX[0-n], where MAX is 4. Beyond numtxqs=4, the remaining TX queues are handled by vhost threads in round-robin fashion. Results from here on are with these changes, and only tuning is to set each vhost's affinity to CPUs[0-3] (taskset -p f vhost-pids). Any idea where does this come from? Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs? 4. Identify reasons for single netperf BW regression. After testing various combinations of #txqs, #vhosts, #netperf sessions, I think the drop for 1 stream is due to TX and RX for a flow being processed on different cpus. Right. Can we fix it? I did two more tests: 1. Pin vhosts to same CPU: - BW drop is much lower for 1 stream case (- 5 to -8% range) - But performance is not so high for more sessions. 2. Changed vhost to be single threaded: - No degradation for 1 session, and improvement for upto 8, sometimes 16 streams (5-12%). - BW degrades after that, all the way till 128 netperf sessions. - But overall CPU utilization improves. Summary of the entire run (for 1-128 sessions): txq=4: BW: (-2.3) CPU: (-16.5)RCPU: (-5.3) txq=16: BW: (-1.9) CPU: (-24.9)RCPU: (-9.6) I don't see any reasons mentioned above. However, for higher number of netperf sessions, I see a big increase in retransmissions: Hmm, ok, and do you see any errors? ___ #netperf ORG NEW BW (#retr)BW (#retr) ___ 1 70244 (0) 64102 (0) 4 21421 (0) 36570 (416) 8 21746 (0) 38604 (148) 16 21783 (0) 40632 (464) 32 22677 (0) 37163 (1053) 64 23648 (4) 36449 (2197) 12823251 (2) 31676 (3185) ___ Single netperf case didn't have any retransmissions so that is not the cause for drop. I tested ixgbe (MQ): ___ #netperf ixgbe ixgbe (pin intrs to cpu#0 on both server/client) BW (#retr) BW (#retr) ___ 1 3567 (117) 6000 (251) 2 4406 (477) 6298 (725) 4 6119 (1085) 7208 (3387) 8 6595 (4276) 7381 (15296) 16 6651 (11651)6856 (30394) Interesting. You are saying we get much more retransmissions with physical nic as well? ___ 5. Test perf in more scenarious: small packets 512 byte packets - BW drop for upto 8 (sometimes 16) netperf sessions, but increases with #sessions: ___ # BW1 BW2 (%) CPU1CPU2 (%)RCPU1 RCPU2 (%) ___ 1 40433800 (-6.0) 50 50 (0) 86 98 (13.9) 2 83587485 (-10.4)153 178 (16.3) 230 264 (14.7) 4 20664 13567 (-34.3) 448 490 (9.3) 530 624 (17.7) 8 25198 17590 (-30.1) 967 1021 (5.5) 10851257 (15.8) 16 23791 24057 (1.1) 19042220 (16.5) 21562578 (19.5) 24 23055 26378 (14.4)28073378 (20.3) 32253901 (20.9) 32 22873 27116 (18.5)37484525 (20.7) 43075239 (21.6) 40 22876 29106 (27.2)47055717 (21.5) 53886591 (22.3) 48 23099 31352 (35.7)56426986 (23.8) 64758085 (24.8) 64 22645 30563 (34.9)75279027 (19.9) 861910656 (23.6) 80 22497 31922 (41.8)937511390 (21.4)10736 13485 (25.6) 96 22509 32718 (45.3)11271 13710 (21.6)12927 16269 (25.8) 128 22255 32397 (45.5)15036 18093 (20.3)17144 21608 (26.0)
Re: [v2 RFC PATCH 0/4] Implement multiqueue virtio-net
On Fri, Sep 17, 2010 at 03:33:07PM +0530, Krishna Kumar wrote: For 1 TCP netperf, I ran 7 iterations and summed it. Explanation for degradation for 1 stream case: Could you document how exactly do you measure multistream bandwidth: netperf flags, etc? 1. Without any tuning, BW falls -6.5%. Any idea where does this come from? Do you see more TX interrupts? RX interrupts? Exits? Do interrupts bounce more between guest CPUs? 2. When vhosts on server were bound to CPU0, BW was as good as with original code. 3. When new code was started with numtxqs=1 (or mq=off, which is the default), there was no degradation. Next steps: --- 1. MQ RX patch is also complete - plan to submit once TX is OK (as well as after identifying bandwidth degradations for some test cases). 2. Cache-align data structures: I didn't see any BW/SD improvement after making the sq's (and similarly for vhost) cache-aligned statically: struct virtnet_info { ... struct send_queue sq[16] cacheline_aligned_in_smp; ... }; 3. Migration is not tested. 4. Identify reasons for single netperf BW regression. 5. Test perf in more scenarious: small packets host - guest guest - external in last case: find some other way to measure host CPU utilization, try multiqueue and single queue devices 6. Use above to figure out what is a sane default for numtxqs. Review/feedback appreciated. Signed-off-by: Krishna Kumar krkum...@in.ibm.com --- -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
[v2 RFC PATCH 0/4] Implement multiqueue virtio-net
Following patches implement transmit MQ in virtio-net. Also included is the user qemu changes. MQ is disabled by default unless qemu specifies it. 1. This feature was first implemented with a single vhost. Testing showed 3-8% performance gain for upto 8 netperf sessions (and sometimes 16), but BW dropped with more sessions. However, adding more vhosts improved BW significantly all the way to 128 sessions. Multiple vhost is implemented in-kernel by passing an argument to SET_OWNER (retaining backward compatibility). The vhost patch adds 173 source lines (incl comments). 2. BW - CPU/SD equation: Average TCP performance increased 23% compared to almost 70% for earlier patch (with unrestricted #vhosts). SD improved -4.2% while it had increased 55% for the earlier patch. Increasing #vhosts has it's pros and cons, but this patch lays emphasis on reducing CPU utilization. Another option could be a tunable to select number of vhosts threads. 3. Interoperability: Many combinations, but not all, of qemu, host, guest tested together. Tested with multiple i/f's on guest, with both mq=on/off, vhost=on/off, etc. Changes from rev1: -- 1. Move queue_index from virtio_pci_vq_info to virtqueue, and resulting changes to existing code and to the patch. 2. virtio-net probe uses virtio_config_val. 3. Remove constants: VIRTIO_MAX_TXQS, MAX_VQS, all arrays allocated on stack, etc. 4. Restrict number of vhost threads to 2 - I get much better cpu/sd results (without any tuning) with low number of vhost threads. Higher vhosts gives better average BW performance (from average of 45%), but SD increases significantly (90%). 5. Working of vhost threads changes, eg for numtxqs=4: vhost-0: handles RX vhost-1: handles TX[0] vhost-0: handles TX[1] vhost-1: handles TX[2] vhost-0: handles TX[3] Enabling MQ on virtio: --- When following options are passed to qemu: - smp 1 - vhost=on - mq=on (new option, default:off) then #txqueues = #cpus. The #txqueues can be changed by using an optional 'numtxqs' option. e.g. for a smp=4 guest: vhost=on - #txqueues = 1 vhost=on,mq=on - #txqueues = 4 vhost=on,mq=on,numtxqs=8 - #txqueues = 8 vhost=on,mq=on,numtxqs=2 - #txqueues = 2 Performance (guest - local host): --- System configuration: Host: 8 Intel Xeon, 8 GB memory Guest: 4 cpus, 2 GB memory, numtxqs=4 All testing without any system tuning, and default netperf Results split across two tables to show SD and CPU usage: TCP: BW vs CPU/Remote CPU utilization: #BW1BW2 (%)CPU1CPU2 (%) RCPU1 RCPU2 (%) 169971 65376 (-6.56) 134 170 (26.86) 322376 (16.77) 220911 24839 (18.78) 107 139 (29.90) 217264 (21.65) 421431 28912 (34.90) 213 318 (49.29) 444541 (21.84) 821857 34592 (58.26) 444 859 (93.46) 9011247 (38.40) 16 22368 33083 (47.90) 899 1523 (69.41) 1813 2410 (32.92) 24 22556 32578 (44.43) 1347 2249 (66.96) 2712 3606 (32.96) 32 22727 30923 (36.06) 1806 2506 (38.75) 3622 3952 (9.11) 40 23054 29334 (27.24) 2319 2872 (23.84) 4544 4551 (.15) 48 23006 28800 (25.18) 2827 2990 (5.76)5465 4718 (-13.66) 64 23411 27661 (18.15) 3708 3306 (-10.84) 7231 5218 (-27.83) 80 23175 27141 (17.11) 4796 4509 (-5.98) 9152 7182 (-21.52) 96 23337 26759 (14.66) 5603 4543 (-18.91) 10890 7162 (-34.23) 128 22726 28339 (24.69) 7559 6395 (-15.39) 14600 10169 (-30.34) Summary:BW: 22.8%CPU: 1.9%RCPU: -17.0% TCP: BW vs SD/Remote SD: #BW1BW2 (%)SD1 SD2 (%)RSD1RSD2 (%) 169971 65376 (-6.56) 4 6 (50.00)21 26 (23.80) 220911 24839 (18.78) 6 7 (16.66)27 28 (3.70) 421431 28912 (34.90) 26 31(19.23)108 111(2.77) 821857 34592 (58.26) 106 135 (27.35)432 393(-9.02) 16 22368 33083 (47.90) 431 577 (33.87)17421828 (4.93) 24 22556 32578 (44.43) 972 1393 (43.31)39154479 (14.40) 32 22727 30923 (36.06) 17232165 (25.65)69086842 (-.95) 40 23054 29334 (27.24) 27742761 (-.46) 10874 8764 (-19.40) 48 23006 28800 (25.18)