envoy
envoy copied to clipboard
envoyproxy
subset lb: added multiple keys or multiple selectors support for single host per subset mode
Signed-off-by: wbpcode [email protected]
Commit Message: subset lb: added multiple keys or multiple selectors support for single host per subset mode Additional Description:
Subset lb is great and powerful lb. But in the previous implementation of subset lb, the single host per subset mode is handled separately and can only be used for configurations that contain single selector and single key. This PR does some optimizations to make the single host mode can be used in more scenes.
- Added multiple keys or multiple selectors support for single host per subset mode.
- Use almost completely same codes for single host mode and normal mode to make the code simpler and cleaner.
- Evaluate all hosts for every update to make sure that we can add hosts to different subsets correctly in all cases (metadata only update, hosts added/removed, hybrid update).
- Eliminate unnecessary metadata match (In the previous implementation, all the hosts will match against to metadata for every subset update and an update of endpoint set will result in all the subsets being updated.)
Risk Level: Mid. Core lb update. Testing: Unit tests ware added. Docs Changes: n/a. Release Notes: Added. Platform Specific Features: n/a.
CC @envoyproxy/api-shepherds: Your approval is needed for changes made to (api/envoy/|docs/root/api-docs/).
envoyproxy/api-shepherds assignee is @adisuissa
CC @envoyproxy/api-watchers: FYI only for changes made to (api/envoy/|docs/root/api-docs/).
:cat:
Caused by: https://github.com/envoyproxy/envoy/pull/22644 was opened by wbpcode.
![repokitteh-read-only[bot] avatar](https://avatars.githubusercontent.com/in/16729?v=4 "Published by a pub.dev verified publisher"){kind=small}
Benchmark result after this PR:
-----------------------------------------------------------------------------------
Benchmark Time CPU Iterations
-----------------------------------------------------------------------------------
benchmarkSubsetLoadBalancerCreate/0/50 1.88 ms 1.88 ms 352
benchmarkSubsetLoadBalancerCreate/1/50 1.38 ms 1.38 ms 496
benchmarkSubsetLoadBalancerCreate/0/64 2.16 ms 2.16 ms 327
benchmarkSubsetLoadBalancerCreate/1/64 1.45 ms 1.45 ms 479
benchmarkSubsetLoadBalancerCreate/0/512 14.7 ms 14.7 ms 48
benchmarkSubsetLoadBalancerCreate/1/512 4.15 ms 4.15 ms 162
benchmarkSubsetLoadBalancerCreate/0/2500 156 ms 156 ms 4
benchmarkSubsetLoadBalancerCreate/1/2500 17.5 ms 17.5 ms 43
benchmarkSubsetLoadBalancerUpdate/0/50 0.662 ms 0.662 ms 1150
benchmarkSubsetLoadBalancerUpdate/1/50 0.184 ms 0.184 ms 3694
benchmarkSubsetLoadBalancerUpdate/0/64 0.758 ms 0.758 ms 902
benchmarkSubsetLoadBalancerUpdate/1/64 0.218 ms 0.218 ms 2933
benchmarkSubsetLoadBalancerUpdate/0/512 13.5 ms 13.5 ms 51
benchmarkSubsetLoadBalancerUpdate/1/512 1.90 ms 1.90 ms 376
benchmarkSubsetLoadBalancerUpdate/0/2500 224 ms 224 ms 3
benchmarkSubsetLoadBalancerUpdate/1/2500 10.00 ms 10.00 ms 73
Benchmark result before this PR:
-----------------------------------------------------------------------------------
Benchmark Time CPU Iterations
-----------------------------------------------------------------------------------
benchmarkSubsetLoadBalancerCreate/0/50 2.09 ms 2.09 ms 344
benchmarkSubsetLoadBalancerCreate/1/50 1.34 ms 1.34 ms 511
benchmarkSubsetLoadBalancerCreate/0/64 2.52 ms 2.51 ms 290
benchmarkSubsetLoadBalancerCreate/1/64 1.44 ms 1.44 ms 499
benchmarkSubsetLoadBalancerCreate/0/512 45.8 ms 45.8 ms 17
benchmarkSubsetLoadBalancerCreate/1/512 3.86 ms 3.86 ms 183
benchmarkSubsetLoadBalancerCreate/0/2500 871 ms 871 ms 1
benchmarkSubsetLoadBalancerCreate/1/2500 14.9 ms 14.9 ms 47
benchmarkSubsetLoadBalancerUpdate/0/50 1.13 ms 1.13 ms 628
benchmarkSubsetLoadBalancerUpdate/1/50 0.135 ms 0.135 ms 5385
benchmarkSubsetLoadBalancerUpdate/0/64 1.70 ms 1.70 ms 379
benchmarkSubsetLoadBalancerUpdate/1/64 0.166 ms 0.166 ms 4265
benchmarkSubsetLoadBalancerUpdate/0/512 78.8 ms 78.8 ms 9
benchmarkSubsetLoadBalancerUpdate/1/512 1.48 ms 1.48 ms 470
benchmarkSubsetLoadBalancerUpdate/0/2500 1716 ms 1716 ms 1
benchmarkSubsetLoadBalancerUpdate/1/2500 7.61 ms 7.61 ms 92
This performance got a minor worse in the single host mode and got a big improvement in normal mode.
Considering that for the multiple priorities host set, the new implementation needn't evaluate all priorities' hosts for every update and may get a better performance in single host mode, so I think this result is just ok.
Retrying Azure Pipelines: Check envoy-presubmit didn't fail.
:cat:
Caused by: a https://github.com/envoyproxy/envoy/pull/22644#issuecomment-1211723962 was created by @wbpcode.
I'm going to be out the rest of this week so re-assigning to @zuercher for review.
I think the performance degradation for the single mode is probably ok, but I'm curious what the benchmark shows if you expand it to higher numbers of hosts (I mostly want to see what the algorithmic complexity of the old and new looks like).
I'm going to be out the rest of this week so re-assigning to @zuercher for review.
I think the performance degradation for the single mode is probably ok, but I'm curious what the benchmark shows if you expand it to higher numbers of hosts (I mostly want to see what the algorithmic complexity of the old and new looks like).
I can run some more test to ensure it.
The algorithmic complexity should no big change except in some corner cases (e.g., many priorities and few hosts for every priority). 🤔
More hosts benchmark test result. cc @ggreenway
Benchmark result after this PR:
------------------------------------------------------------------------------------
Benchmark Time CPU Iterations
------------------------------------------------------------------------------------
benchmarkSubsetLoadBalancerCreate/0/50 1.29 ms 1.29 ms 502
benchmarkSubsetLoadBalancerCreate/1/50 0.967 ms 0.967 ms 752
benchmarkSubsetLoadBalancerCreate/0/64 1.44 ms 1.44 ms 492
benchmarkSubsetLoadBalancerCreate/1/64 1.04 ms 1.04 ms 695
benchmarkSubsetLoadBalancerCreate/0/512 10.9 ms 10.9 ms 64
benchmarkSubsetLoadBalancerCreate/1/512 3.08 ms 3.08 ms 225
benchmarkSubsetLoadBalancerCreate/0/4096 374 ms 374 ms 2
benchmarkSubsetLoadBalancerCreate/1/4096 20.1 ms 20.1 ms 35
benchmarkSubsetLoadBalancerCreate/0/10000 2373 ms 2373 ms 1
benchmarkSubsetLoadBalancerCreate/1/10000 50.0 ms 50.0 ms 12
benchmarkSubsetLoadBalancerUpdate/0/50 0.419 ms 0.419 ms 1614
benchmarkSubsetLoadBalancerUpdate/1/50 0.138 ms 0.138 ms 4984
benchmarkSubsetLoadBalancerUpdate/0/64 0.567 ms 0.567 ms 1207
benchmarkSubsetLoadBalancerUpdate/1/64 0.179 ms 0.179 ms 3935
benchmarkSubsetLoadBalancerUpdate/0/512 12.2 ms 12.2 ms 58
benchmarkSubsetLoadBalancerUpdate/1/512 1.60 ms 1.60 ms 451
benchmarkSubsetLoadBalancerUpdate/0/4096 583 ms 583 ms 1
benchmarkSubsetLoadBalancerUpdate/1/4096 13.5 ms 13.5 ms 52
benchmarkSubsetLoadBalancerUpdate/0/10000 3330 ms 3330 ms 1
benchmarkSubsetLoadBalancerUpdate/1/10000 34.1 ms 34.1 ms 21
Benchmark result before this PR:
------------------------------------------------------------------------------------
Benchmark Time CPU Iterations
------------------------------------------------------------------------------------
benchmarkSubsetLoadBalancerCreate/0/50 1.54 ms 1.54 ms 460
benchmarkSubsetLoadBalancerCreate/1/50 0.952 ms 0.952 ms 714
benchmarkSubsetLoadBalancerCreate/0/64 1.86 ms 1.86 ms 362
benchmarkSubsetLoadBalancerCreate/1/64 1.01 ms 1.01 ms 692
benchmarkSubsetLoadBalancerCreate/0/512 46.6 ms 46.6 ms 16
benchmarkSubsetLoadBalancerCreate/1/512 3.14 ms 3.14 ms 235
benchmarkSubsetLoadBalancerCreate/0/4096 2070 ms 2070 ms 1
benchmarkSubsetLoadBalancerCreate/1/4096 19.9 ms 19.9 ms 36
benchmarkSubsetLoadBalancerCreate/0/10000 11388 ms 11388 ms 1
benchmarkSubsetLoadBalancerCreate/1/10000 49.0 ms 49.0 ms 13
benchmarkSubsetLoadBalancerUpdate/0/50 0.961 ms 0.961 ms 728
benchmarkSubsetLoadBalancerUpdate/1/50 0.106 ms 0.106 ms 6478
benchmarkSubsetLoadBalancerUpdate/0/64 1.43 ms 1.43 ms 506
benchmarkSubsetLoadBalancerUpdate/1/64 0.135 ms 0.135 ms 4246
benchmarkSubsetLoadBalancerUpdate/0/512 74.5 ms 74.5 ms 11
benchmarkSubsetLoadBalancerUpdate/1/512 1.26 ms 1.26 ms 558
benchmarkSubsetLoadBalancerUpdate/0/4096 3826 ms 3826 ms 1
benchmarkSubsetLoadBalancerUpdate/1/4096 11.2 ms 11.2 ms 66
benchmarkSubsetLoadBalancerUpdate/0/10000 21114 ms 21114 ms 1
benchmarkSubsetLoadBalancerUpdate/1/10000 27.4 ms 27.4 ms 26
friendly ping @zuercher could you give a LGTM again. Then we can merge this PR. 😸