答复: [PATCH v3 0/4] Add qemu support setting qos via ovs on ovs interface

Hi Yalan, You are right about it. For other interface type, values in tc rules are calculated by multiply 8*1000 instead of 8*1024. I didn’t notice it. To make them uniform, I will fix it in next patch. Really thanks for your help. Best Regards, Jinsheng Zhang 发件人: Yalan Zhang [mailto:yalzhang@redhat.com] 发送时间: 2021年10月29日 16:52 收件人: Jinsheng Zhang (张金生)-云服务集团 抄送: libvir-list(a)redhat.com; Norman Shen(申嘉童) 主题: Re: [PATCH v3 0/4] Add qemu support setting qos via ovs on ovs interface Hi Jinsheng, I asked as I have mixed the "K" as 1000. Thank you for the explanation, I'm clear now. And I found the "inbound.peak : 200" was calculated to "ceil 1638Kbit", maybe 1600Kbit is more reasonable? as I found for other interface type like nat, it was 1600Kbit from tc output. Please help to confirm, Thank you! # virsh domiftune rhel vnet5 inbound.average: 100 inbound.peak : 200 inbound.burst : 256 ... # tc -d class show dev vnet5 class htb 1:1 parent 1:fffe prio 0 quantum 10240 rate 819200bit ceil 1638Kbit linklayer ethernet burst 256Kb/1 mpu 0b cburst 256Kb/1 mpu 0b level 0 class htb 1:fffe root rate 1638Kbit ceil 1638Kbit linklayer ethernet burst 1499b/1 mpu 0b cburst 1499b/1 mpu 0b level 7 ------- Best Regards, Yalan Zhang IRC: yalzhang On Thu, Oct 28, 2021 at 4:20 PM Jinsheng Zhang (张金生)-云服务集团 <zhangjl02@inspur.com<mailto:zhangjl02@inspur.com>> wrote: Hi Yalan, It seems that there is no output error abount inbound settings from your statistics. 100KB is short for 100 kilobytes, and 1 byte is 8 bit, therefore 100 kilobytes is 800 kilobit and is also 1024*800 bit which is 819200 bit or 800 Kbit for short. Similarly, 200 KB is equal to 1600Kbit. From your test results, inbound.average is set to 400 KB which is 400 * 1024 * 8 bit(approximately 3.2*10^6 bits). outbound.average is set to 100 KB which is approximately 0.8*10^6 bits. Considering peek and burst is larger than average. The netperf test result is meaningful. For the second bug mentioned, after create the ovs-net, tc rules are created. But when attach an interface to an instance, qos settings is not add to port neither in xml or tc . It is a bug, I think. I will think about fixing this. ------- Best Regards, Jinsheng Zhang 发件人: Yalan Zhang [mailto:yalzhang@redhat.com<mailto:yalzhang@redhat.com>] 发送时间: 2021年10月27日 18:35 收件人: Jinsheng Zhang (张金生)-云服务集团 抄送: libvir-list@redhat.com<mailto:libvir-list@redhat.com>; Norman Shen(申嘉童) 主题: Re: [PATCH v3 0/4] Add qemu support setting qos via ovs on ovs interface Hi Jinsheng, Thank you for the explanation. From the statistics above, the tc outputs for outbound matches. But I'm confused about the inbound statistics: # virsh domiftune rhel vnet5 inbound. approximately 3.2*10^6 bits: 100 inbound.peak : 200 inbound.burst : 256 ... # tc -d class show dev vnet5 class htb 1:1 parent 1:fffe prio 0 quantum 10240 rate 819200bit ceil 1638Kbit linklayer ethernet burst 256Kb/1 mpu 0b cburst 256Kb/1 mpu 0b level 0 class htb 1:fffe root rate 1638Kbit ceil 1638Kbit linklayer ethernet burst 1499b/1 mpu 0b cburst 1499b/1 mpu 0b level 7 As the value in libvirt xml is KB, inbound.average: *100 KB* can not match with *"rate 819200bit"* in tc outputs, I supposed it should be 800Kbit. Please help to confirm. And so does "ceil 1638Kbit" (may be it should be 1600Kbit as "inbound.peak : 200"). I have run netperf to test the actual rate, the result is pass. 2 vm connected to the same bridge, set one vm with Qos, see test results below: # virsh domiftune rhel vnet0 inbound.average: 400 inbound.peak : 500 inbound.burst : 125 inbound.floor : 0 outbound.average: 100 outbound.peak : 200 outbound.burst : 256 Throughput for inbound: 3.92 * 10^6bits/sec Throughput for outbound: 0.93 * 10^6bits/sec These patches fixed the bug [1] which closed with deferred resolution. Thank you! And this reminds me of another ovs Qos related bug [2], which was about network. And I tried with the scenarios in [2], there are no changes(not fixed). Just for information. :-) [1] https://bugzilla.redhat.com/show_bug.cgi?id=1510237 [2] https://bugzilla.redhat.com/show_bug.cgi?id=1826168 ------- Best Regards, Yalan Zhang IRC: yalzhang On Tue, Oct 26, 2021 at 3:23 PM Jinsheng Zhang (张金生)-云服务集团 <zhangjl02@inspur.com<mailto:zhangjl02@inspur.com>> wrote: Hi Yalan, 1) For inbound, we can use `ovs-vsctl list qos` and `ovs-vsctl list queue` to check them from the openvswitch side. Values can be found in other_config. Inbound is in kbyte when set qos with `virsh domiftune …`, well it is in bit in ovs, Therefore, when inbound.average is set to 100, the corresponding value will be set to 819200 in ovs. 2) For outbound, it is in kbyte in libvirt and ingress_policing_XX in ovs interface is in kbit. 3) Ovs use tc to set qos, so we can see output from tc command. This patch is to unify the qos control and query on ovs ports. The conversion explanation is added in this patch: https://listman.redhat.com/archives/libvir-list/2021-August/msg00422.html And there are 6 following patches to fix some bugs. See https://listman.redhat.com/archives/libvir-list/2021-August/msg00423.html ------- Best Regards, Jinsheng Zhang 发件人: Yalan Zhang [mailto:yalzhang@redhat.com<mailto:yalzhang@redhat.com>] 发送时间: 2021年10月25日 17:54 收件人: Michal Prívozník; Jinsheng Zhang (张金生)-云服务集团 抄送: libvir-list@redhat.com<mailto:libvir-list@redhat.com>; Norman Shen(申嘉童); zhangjl02 主题: Re: [PATCH v3 0/4] Add qemu support setting qos via ovs on ovs interface Hi Jinsheng, I have tested the patch and have some questions, could you please help to confirm? 1) For inbound, how to check it from the openvswitch side? tc will still show the statistics, is that expected? 2) For outbound, the peak is ignored. I just can not understand the "ingress_policing_burst: 2048", how can it come from the setting "outbound.burst : 256"? 3) Is the output from tc command expected? Test inbound: 1. start vm with setting as below: <interface type='bridge'> <source bridge='ovsbr0'/> <virtualport type='openvswitch'/> <bandwidth> <inbound average='100' peak='200' burst='256'/> </bandwidth> ... </interface> 2. # virsh domiftune rhel vnet5 inbound.average: 100 inbound.peak : 200 inbound.burst : 256 inbound.floor : 0 outbound.average: 0 outbound.peak : 0 outbound.burst : 0 # ip l 17: vnet5: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc htb master ovs-system state UNKNOWN mode DEFAULT group default qlen 1000 link/ether fe:54:00:4d:43:5a brd ff:ff:ff:ff:ff:ff # ovs-vsctl show interface …... ingress_policing_burst: 0 ingress_policing_kpkts_burst: 0 ingress_policing_kpkts_rate: 0 ingress_policing_rate: 0 …... name : vnet5 # tc -d class show dev vnet5 class htb 1:1 parent 1:fffe prio 0 quantum 10240 rate 819200bit ceil 1638Kbit linklayer ethernet burst 256Kb/1 mpu 0b cburst 256Kb/1 mpu 0b level 0 class htb 1:fffe root rate 1638Kbit ceil 1638Kbit linklayer ethernet burst 1499b/1 mpu 0b cburst 1499b/1 mpu 0b level 7 # tc -d filter show dev vnet5 parent ffff: (no outputs) For outbound: # virsh dumpxml rhel | grep /bandwidth -B2 <bandwidth> <outbound average='100' peak='200' burst='256'/> </bandwidth> # virsh domiftune rhel vnet9 inbound.average: 0 inbound.peak : 0 inbound.burst : 0 inbound.floor : 0 outbound.average: 100 outbound.peak : 200 outbound.burst : 256 # ovs-vsctl list interface ingress_policing_burst: 2048 ingress_policing_kpkts_burst: 0 ingress_policing_kpkts_rate: 0 ingress_policing_rate: 800 ... # tc -d filter show dev vnet9 parent ffff: filter protocol all pref 49 basic chain 0 filter protocol all pref 49 basic chain 0 handle 0x1 action order 1: police 0x1 rate 800Kbit burst 256Kb mtu 64Kb action drop/pipe overhead 0b linklayer unspec ref 1 bind 1 # tc -d class show dev vnet9 (no outputs) ------- Best Regards, Yalan Zhang IRC: yalzhang On Mon, Jul 12, 2021 at 3:43 PM Michal Prívozník <mprivozn@redhat.com<mailto:mprivozn@redhat.com>> wrote: On 7/9/21 3:31 PM, Jinsheng Zhang (张金生)-云服务集团 wrote: Perfect. Reviewed-by: Michal Privoznik <mprivozn@redhat.com<mailto:mprivozn@redhat.com>> and pushed. Congratulations on your first libvirt contribution! Michal