Bug 698541
| Summary: | vhost-net/kvm need optimizations for UDP R from guest to host . | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Product: | Red Hat Enterprise Linux 7 | Reporter: | Quan Wenli <wquan> | ||||||
| Component: | kernel | Assignee: | Vlad Yasevich <vyasevic> | ||||||
| Status: | CLOSED WONTFIX | QA Contact: | Virtualization Bugs <virt-bugs> | ||||||
| Severity: | medium | Docs Contact: | |||||||
| Priority: | medium | ||||||||
| Version: | 7.0 | CC: | ailan, amit.shah, juzhang, knoel, mkenneth, virt-bugs, virt-maint, vyasevic, wquan | ||||||
| Target Milestone: | rc | ||||||||
| Target Release: | 7.0 | ||||||||
| Hardware: | Unspecified | ||||||||
| OS: | Unspecified | ||||||||
| Whiteboard: | |||||||||
| Fixed In Version: | Doc Type: | Bug Fix | |||||||
| Doc Text: | Story Points: | --- | |||||||
| Clone Of: | Environment: | ||||||||
| Last Closed: | 2016-06-22 18:00:22 UTC | Type: | --- | ||||||
| Regression: | --- | Mount Type: | --- | ||||||
| Documentation: | --- | CRM: | |||||||
| Verified Versions: | Category: | --- | |||||||
| oVirt Team: | --- | RHEL 7.3 requirements from Atomic Host: | |||||||
| Cloudforms Team: | --- | Target Upstream Version: | |||||||
| Embargoed: | |||||||||
| Bug Depends On: | |||||||||
| Bug Blocks: | 1113511 | ||||||||
| Attachments: |
|
||||||||
Basically I think to improve UDP receive speed we need something like GRO for UDP. Another thing that is reported to help here is the event index feature. In both cases we need the code upstream first as maintaining a forked version in rhel would be very painful. Deferring to RHEL7. It seems as if we need to add GRO like functionality for UDP, which is too much for RHEL6. First do it upstream. Created attachment 923057 [details]
UDP results with 3.10.0-138.el7.x86_64/qemu-kvm-1.5.3-66.el7.x86_64
Re-tested it on latest rhel7 host/guest with 3.10.0-138.el7.x86_64/qemu-kvm-1.5.3-66.el7.x86_64 From atteched file UDP_Rhel7-3.10.0-138.html, we could see : - for samll packet ( <= 4096), 0% packet drop in both virtio-net/vhost, but host sends more UDP through virtio-net than vhost, so there is 20% around degradation through vhost. - for large packet ( > 4096), although host sends more (around 20%) UDP through virtio-net than vhost, but guest receive UDP more through vhost than virtio-net. Vlad, any comments about above results, do you think we still have UDP issue with vhost? Things improved, but still requires some follow-up/investigation. Deferring to 7.2. |
Created attachment 493728 [details] rhel6.1-snp1-vhost-net vs virtio-net megabyte/cpu Description of problem: 1> When application doesn't flow control, megabyte/cpu shows regression (drop ) for all UDP R for guest to host.Detail info ,please check the attachment. 2> When application does flow control, for example by rate-limiting the stream to within 10% and 30% packet drops, vhost-net doesn't have performance improvement for UDP R for guest to host compared with virtio-net. Lost packet rate = 10% vhost-net Socket Message Elapsed Messages Size Size Time Okay Errors Throughput bytes bytes secs # # 10^6bits/sec 124928 1460 60.00 468000 0 91.10 124928 60.00 420001 81.76 virtio-net Socket Message Elapsed Messages Size Size Time Okay Errors Throughput bytes bytes secs # # 10^6bits/sec 124928 1460 60.00 468000 0 91.10 124928 60.00 420003 81.76 Lost packet rate = 30% vhost-net Socket Message Elapsed Messages Size Size Time Okay Errors Throughput bytes bytes secs # # 10^6bits/sec 124928 1460 60.00 600000 0 116.80 124928 60.00 420122 81.78 virtio-net Socket Message Elapsed Messages Size Size Time Okay Errors Throughput bytes bytes secs # # 10^6bits/sec 124928 1460 60.00 600000 0 116.80 124928 60.00 420029 81.77 Version-Release number of selected component (if applicable): host: kernel-2.6.32-125.el6.x86_64 qemu-kvm-0.12.1.2-2.152.el6.x86_64 guest: kernel-2.6.32-125.el6.x86_64 How reproducible: Steps to Reproduce: 1.steps for description 1 run netserver on the guest run netperf ci on the host like : mpstat -P ALL 1 &>${CPU_FILE} & ssh $client $client_path -C -c -H $server -l 30 -t UDP_STREAM -- -m $packet_size && kill -9 `pgrep mpstat` Calculate megabyte/cpu for each packets sizes 2.steps for description 2 Rate limit the sender and see at what rate of sender do you get some predefined loss ratio (e.g control the lost packet rate within 10% and 30%). configurate netperf with "--enable-intervals=yes " Lost packet rate = 10% vhost-net:/root/tool/netperf-2.4.5/src/netperf -w 1 -b 78 -H 192.168.0.13 -l 60 -t UDP_STREAM -- -m 1460 virtio-net:/root/tool/netperf-2.4.5/src/netperf -w 1 -b 78 -H 192.168.0.13 -l60 -t UDP_STREAM -- -m 1460 Lost packet rate = 30% vhost-net:/root/tool/netperf-2.4.5/src/netperf -w 1 -b 100 -H 192.168.0.13 -l 60 -t UDP_STREAM -- -m 1460 virtio-net:/root/tool/netperf-2.4.5/src/netperf -w 1 -b 100 -H 192.168.0.13 -l 60 -t UDP_STREAM -- -m 1460 Actual results: Expected results: Additional info: