maglev: Allocate permutations slice ahead of time #14622
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
christarazi
added
kind/performance
maintainer-s-little-helper
bot
added
dont-merge/needs-release-note-label
christarazi
force-pushed
the
pr/christarazi/maglev-perf
branch
from
4031d65
to
67f4e7c
Compare
|
test-me-please |
brb
reviewed
Jan 15, 2021
christarazi
force-pushed
the
pr/christarazi/maglev-perf
branch
3 times, most recently
from
2f91c70
to
be5555a
Compare
|
test-me-please |
|
CI has passed. Pushing fixups to nits, which don't require a re-run of CI. |
christarazi
force-pushed
the
pr/christarazi/maglev-perf
branch
2 times, most recently
from
af12caa
to
1159b37
Compare
aanm
approved these changes
Jan 30, 2021
christarazi
force-pushed
the
pr/christarazi/maglev-perf
branch
2 times, most recently
from
989c1bf
to
b14d077
Compare
brb
reviewed
Feb 1, 2021
christarazi
force-pushed
the
pr/christarazi/maglev-perf
branch
from
b14d077
to
9201cdd
Compare
|
test-me-please |
brb
reviewed
Feb 2, 2021
brb
reviewed
Feb 2, 2021
Previously, the Maglev permutations slice is allocated inside
`getPermutation()`, which means every time `maglev.GetLookupTable()` is
called, the slice will be recreated. That means it would recreated on
every Service creation or update.
This commit allocates the slice ahead of time when the Maglev subsystem
is initialized (`maglev.Init()`). This gives a rough improvement in time
of around 15%.
The slice is allocated based on a heuristic computation, involving the M
size. Nodes running with less than 8GB of RAM do not apply and Cilium
will not be preallocate the slice ahead of time. For nodes with more
than the aforementioned threshold, the formula for the heuristic is:
(M / 100) * 100
This is derived from the maximum backends property from Maglev where
|backends| * 100 < M. This gives the following memory pressure profile
based on M (left-hand side):
251: 0.004806594848632812 MB
509: 0.019766311645507812 MB
1021: 0.07953193664550783 MB
2039: 0.3171936798095703 MB
4093: 1.2781256866455077 MB
8191: 5.118750076293945 MB
16381: 20.472500686645507 MB
32749: 81.82502754211426 MB
65521: 327.5300171661377 MB
131071: 1310.700000076294 MB
If the user has more backends than the preallocated size, we will adjust
the allocation to be |backends| * M. This is because we want to ensure
that the Maglev subsystem isn't thrashing trying to reallocate the slice
on each Service create or update.
Benchmark timings:
Before
```
$ go test -v ./pkg/maglev -check.v -check.b -check.btime 5s -check.bmem
=== RUN Test
PASS: maglev_test.go:91: MaglevTestSuite.BenchmarkGetMaglevTable 50 341255883 ns/op1049633800 B/op 12 allocs/op
OK: 1 passed
--- PASS: Test (17.42s)
PASS
ok github.com/cilium/cilium/pkg/maglev 17.475s
```
After
```
$ go test -v ./pkg/maglev -check.v -check.b -check.btime 5s -check.bmem
=== RUN Test
PASS: maglev_test.go:91: MaglevTestSuite.BenchmarkGetMaglevTable 50 282792516 ns/op 1057899 B/op 11 allocs/op
OK: 1 passed
--- PASS: Test (17.89s)
PASS
ok github.com/cilium/cilium/pkg/maglev 17.943s
```
Suggested-by: Martynas Pumputis <m@lambda.lt>
Signed-off-by: Chris Tarazi <chris@isovalent.com>
christarazi
force-pushed
the
pr/christarazi/maglev-perf
branch
from
9201cdd
to
b15cd54
Compare
|
CI passed on the last run. Pushed a change that only updated the comments so no actual code was changed. |
borkmann
approved these changes
Feb 3, 2021
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Previously, the Maglev permutations slice is allocated inside
getPermutation(), which means every timemaglev.GetLookupTable()iscalled, the slice will be recreated. That means it would recreated on
every Service creation or update.
This commit allocates the slice ahead of time when the Maglev subsystem
is initialized (
maglev.Init()). This gives a rough improvement in timeof around 15%.
The slice is allocated based on a heuristic computation, involving the M
size. Nodes running with less than 8GB of RAM do not apply and Cilium
will not be preallocate the slice ahead of time. For nodes with more
than the aforementioned threshold, the formula for the heuristic is:
This is derived from the maximum backends property from Maglev where
|backends| * 100 < M. This gives the following memory pressure profile
based on M (left-hand side):
If the user has more backends than the preallocated size, we will adjust
the allocation to be |backends| * M. This is because we want to ensure
that the Maglev subsystem isn't thrashing trying to reallocate the slice
on each Service create or update.
Benchmark timings:
Before
After
Suggested-by: Martynas Pumputis m@lambda.lt
FlameGraphs:
Related: #14397