alts: receive low watermark support #8513
Open
Conversation
This file contains hidden or 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
It's generally useful to have new-and-improved Go. One specific useful feature is `b.Loop()`, which makes benchmarking easier.
It's only called in one place, and is effectively a method on conn. Part of grpc#8510.
Increases large write speed by 9.62% per BenchmarkLargeMessage. Detailed benchmarking numbers below. Rather than use different sizes for the maximum read record, write record, and write buffer, just use 1MB for all of them. Using larger records reduces the amount of payload splitting and the number of syscalls made by ALTS. Part of grpc#8510. SO_RCVLOWAT and TCP receive zerocopy are only effective with larger payloads, and so ALTS can't be limiting payload sizes to 4 KiB. SO_RCVLOWAT and zerocopy are on the receive side, but for benchmarking purposes we need ALTS to send large messages. Benchmarks: $ benchstat large_msg_old.txt large_msg.txt goos: linux goarch: amd64 pkg: google.golang.org/grpc/credentials/alts/internal/conn cpu: AMD Ryzen Threadripper PRO 3945WX 12-Cores │ large_msg_old.txt │ large_msg.txt │ │ sec/op │ sec/op vs base │ LargeMessage-12 68.88m ± 1% 62.25m ± 0% -9.62% (p=0.002 n=6)
SO_RCVLOWAT is *not* enabled by default. Users must turn it on via an option, as not everyone will want the CPU/throughput tradeoff. Part of grpc#8510.
The implementation of setRcvlowat is based on the gRCP C++ library implementation. Part of grpc#8510.
Part of grpc#8510. For large payloads, we see about a 36% reduction in CPU usage and a 2.5% reduction in throughput. This is expected, and has been observed in the C++ gRPC library as well. The expectation is that with TCP receive zerocopy also enabled, we'll see both a reduction in CPU usage and an increase in throughput. For users not using zerocopy, they can choose whether the CPU/throughput tradeoff is worthwhile. SO_RCVLOWAT is unused for small payloads, where its impact would be insignificant (but would cost cycles to make syscalls). Enabling it has no effect on CPU usage or throughput of small payloads, and so is omitted from the below benchmarks so as not to water down the impact on both CPU usage and throughput. Benchmark numbers: $ benchstat -col "/rcvlowat" -filter "/size:(64_KiB OR 512_KiB OR 1_MiB OR 4_MiB) .unit:(Mbps OR cpu-usec/op)" ~/lowat_numbers.txt goos: linux goarch: amd64 pkg: google.golang.org/grpc/credentials/alts/internal/conn cpu: AMD Ryzen Threadripper PRO 3945WX 12-Cores │ false │ true │ │ Mbps │ Mbps vs base │ Rcvlowat/size=64_KiB-12 47.44 ± 0% 47.32 ± 0% -0.24% (p=0.015 n=6) Rcvlowat/size=512_KiB-12 299.2 ± 0% 293.6 ± 0% -1.90% (p=0.002 n=6) Rcvlowat/size=1_MiB-12 482.1 ± 0% 468.1 ± 0% -2.88% (p=0.002 n=6) Rcvlowat/size=4_MiB-12 887.4 ± 1% 842.3 ± 0% -5.08% (p=0.002 n=6) geomean 279.1 272.0 -2.54% │ false │ true │ │ cpu-usec/op │ cpu-usec/op vs base │ Rcvlowat/size=64_KiB-12 992.2 ± 1% 666.1 ± 1% -32.87% (p=0.002 n=6) Rcvlowat/size=512_KiB-12 7.431k ± 1% 4.660k ± 0% -37.30% (p=0.002 n=6) Rcvlowat/size=1_MiB-12 14.720k ± 1% 9.192k ± 0% -37.56% (p=0.002 n=6) Rcvlowat/size=4_MiB-12 59.19k ± 1% 37.50k ± 3% -36.64% (p=0.002 n=6) geomean 8.953k 5.719k -36.12%
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #8513 +/- ##
==========================================
- Coverage 82.40% 81.76% -0.64%
==========================================
Files 414 414
Lines 40531 40575 +44
==========================================
- Hits 33399 33178 -221
- Misses 5770 6026 +256
- Partials 1362 1371 +9
🚀 New features to boost your workflow:
|
|
I believe it should be possible to set this using a custom dialer, without any code changes. Have you considered that approach? |
I don't think using a custom dialer would work since we need to update the value of socket option before every read. Maybe we should consider directly implementing this in the transport. |
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.
Reduce CPU usage significantly via
SO_RCVLOWAT. There is a smallthroughput penalty, so
SO_RCVLOWATis not enabled by default.Users must turn it on via an option, as not everyone will want the
CPU/throughput tradeoff.
Part of #8510.
For large payloads, we see about a 36% reduction in CPU usage and a 2.5%
reduction in throughput. This is expected, and has been observed in the
C++ gRPC library as well. The expectation is that with TCP receive
zerocopy also enabled, we'll see both a reduction in CPU usage and an
increase in throughput. For users not using zerocopy, they can choose
whether the CPU/throughput tradeoff is worthwhile.
SO_RCVLOWAT is unused for small payloads, where its impact would be
insignificant (but would cost cycles to make syscalls). Enabling it in
grpc-go has no effect on CPU usage or throughput of small payloads, and
so is omitted from the below benchmarks so as not to water down the
impact on both CPU usage and throughput.
Benchmarks are of the ALTS layer alone.
Note that this PR includes #8512. GitHub doesn't support proper commit chains / stacked PRs, so I'm doing this in several PRs with some (annoyingly) redundant commits. Let me know if this isn't a good workflow for you and I'll change things up.
Benchmark numbers: