feat: don't allocate in UDP recv path #2184
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
mxinden
force-pushed
the
recv-no-alloc-common
branch
from
ea91ae6 to
c2520e2
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2184 +/- ##
=======================================
Coverage 95.39% 95.39%
=======================================
Files 112 112
Lines 36373 36447 +74
=======================================
+ Hits 34697 34769 +72
- Misses 1676 1678 +2 ☔ View full report in Codecov by Sentry. |
Failed Interop TestsQUIC Interop Runner, client vs. server neqo-latest as client
neqo-latest as server
All resultsSucceeded Interop TestsQUIC Interop Runner, client vs. server neqo-latest as client
neqo-latest as server
Unsupported Interop TestsQUIC Interop Runner, client vs. server neqo-latest as client
neqo-latest as server
|
mxinden
requested review from
KershawChang,
martinthomson and
larseggert
as code owners
This was referenced Oct 17, 2024
mxinden
force-pushed
the
recv-no-alloc-common
branch
from
c2520e2 to
0dcc9b2
Compare
Benchmark resultsPerformance differences relative to b5e0374. coalesce_acked_from_zero 1+1 entries: Change within noise threshold. time: [98.917 ns 99.193 ns 99.481 ns]
change: [-0.9459% -0.5340% -0.0999%] (p = 0.02 < 0.05)
coalesce_acked_from_zero 3+1 entries: Change within noise threshold. time: [117.01 ns 117.37 ns 117.75 ns]
change: [-0.8819% -0.5415% -0.2079%] (p = 0.00 < 0.05)
coalesce_acked_from_zero 10+1 entries: Change within noise threshold. time: [116.51 ns 116.85 ns 117.29 ns]
change: [-1.3921% -0.8099% -0.3062%] (p = 0.00 < 0.05)
coalesce_acked_from_zero 1000+1 entries: No change in performance detected. time: [96.717 ns 101.78 ns 112.97 ns]
change: [-2.1810% +0.6453% +6.4132%] (p = 0.83 > 0.05)
RxStreamOrderer::inbound_frame(): Change within noise threshold. time: [111.61 ms 111.67 ms 111.74 ms]
change: [-0.3427% -0.2718% -0.1991%] (p = 0.00 < 0.05)
transfer/pacing-false/varying-seeds: No change in performance detected. time: [25.850 ms 26.759 ms 27.654 ms]
change: [-7.9363% -3.0643% +2.2316%] (p = 0.24 > 0.05)
transfer/pacing-true/varying-seeds: No change in performance detected. time: [33.889 ms 35.649 ms 37.426 ms]
change: [-6.8098% -0.1727% +6.8300%] (p = 0.96 > 0.05)
transfer/pacing-false/same-seed: No change in performance detected. time: [25.294 ms 25.955 ms 26.623 ms]
change: [-6.4483% -2.4512% +1.7271%] (p = 0.25 > 0.05)
transfer/pacing-true/same-seed: No change in performance detected. time: [39.871 ms 41.886 ms 43.943 ms]
change: [-9.8397% -3.6845% +3.0139%] (p = 0.27 > 0.05)
1-conn/1-100mb-resp/mtu-1500 (aka. Download)/client: No change in performance detected. time: [871.83 ms 880.49 ms 889.33 ms]
thrpt: [112.44 MiB/s 113.57 MiB/s 114.70 MiB/s]
change:
time: [-0.7754% +0.5489% +1.9642%] (p = 0.44 > 0.05)
thrpt: [-1.9264% -0.5459% +0.7815%]
1-conn/10_000-parallel-1b-resp/mtu-1500 (aka. RPS)/client: No change in performance detected. time: [316.33 ms 319.33 ms 322.42 ms]
thrpt: [31.015 Kelem/s 31.315 Kelem/s 31.613 Kelem/s]
change:
time: [-1.3503% -0.0743% +1.3366%] (p = 0.92 > 0.05)
thrpt: [-1.3190% +0.0743% +1.3688%]
1-conn/1-1b-resp/mtu-1500 (aka. HPS)/client: No change in performance detected. time: [33.780 ms 34.009 ms 34.259 ms]
thrpt: [29.189 elem/s 29.404 elem/s 29.603 elem/s]
change:
time: [-1.5083% -0.5383% +0.4334%] (p = 0.29 > 0.05)
thrpt: [-0.4315% +0.5412% +1.5314%]
1-conn/1-100mb-resp/mtu-65536 (aka. Download)/client: 💚 Performance has improved. time: [111.51 ms 111.80 ms 112.08 ms]
thrpt: [892.23 MiB/s 894.46 MiB/s 896.77 MiB/s]
change:
time: [-2.0616% -1.5830% -1.0883%] (p = 0.00 < 0.05)
thrpt: [+1.1003% +1.6085% +2.1050%]
1-conn/10_000-parallel-1b-resp/mtu-65536 (aka. RPS)/client: No change in performance detected. time: [315.47 ms 318.63 ms 321.84 ms]
thrpt: [31.072 Kelem/s 31.384 Kelem/s 31.699 Kelem/s]
change:
time: [-1.0625% +0.4502% +1.9110%] (p = 0.55 > 0.05)
thrpt: [-1.8752% -0.4482% +1.0739%]
1-conn/1-1b-resp/mtu-65536 (aka. HPS)/client: No change in performance detected. time: [33.916 ms 34.102 ms 34.308 ms]
thrpt: [29.148 elem/s 29.324 elem/s 29.484 elem/s]
change:
time: [-0.2857% +0.5354% +1.3188%] (p = 0.21 > 0.05)
thrpt: [-1.3016% -0.5325% +0.2865%]
Client/server transfer resultsTransfer of 33554432 bytes over loopback.
|
mxinden
changed the title
mxinden
force-pushed
the
recv-no-alloc-common
branch
from
7c23ff6 to
af47ac2
Compare
|
This is the final pull request in the stack. All commits, but the first (08a0003), have been reviewed and merged through the 3 previous stacked pull requests. @martinthomson and @larseggert can you do one more review? |
Previously the payload type of `Datagram` was `Vec<u8>`, thus the `Datagram` always owned the UDP datagram payload. This change enables `Datagram` to represent both (a) an owned payload allocation (i.e. `Vec<u8>`), but also represent (b) a view into an existing payload (e.g. a long lived receive buffer via `&[u8]`). The default payload type stays `Vec<u8>`, thus not breaking existing usage of `Datagram`.
Previously `process_input` (and the like) took a `Datagram<Vec<u8>>` as input. In other words, it required allocating each UDP datagram payload in a dedicated `Vec<u8>` before passing it to `process_input`. With this patch, `process_input` accepts a `Datagram<&[u8]>`. In other words, it accepts a `Datagram` with a borrowed view into an existing buffer (`&[u8]`), e.g. a long lived receive buffer.
Previously `recv_inner` would return `Datagram<Vec<u8>>`. In other words, it would allocate a new `Vec<u8>` for each UDP datagram payload. Now `recv_inner` reads into a provided buffer and returns `Datagram<&[u8]>`, i.e. it returns a view into the provided buffer without allocating.
Pass a long lived receive buffer to `neqo_udp::recv_inner`, receiving an iterator of `Datagram<&[u8]>`s pointing into that buffer, thus no longer allocating in UDP receive path.
mxinden
force-pushed
the
recv-no-alloc-common
branch
from
af47ac2 to
2b322d7
Compare
|
@mxinden is this good to go? Nothing more pending? |
|
Nothing else pending from my end @larseggert. Please merge in case you have no further comments. I will propose a new release once merged. Excited to get this into Firefox. |
|
Thank you for the help here! For early CPU profiles of this patch in Firefox see #2198 (comment) |
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.
This is the final stacked pull request combining the following commits:
feat(common): make Datagram generic over payload type
Previously the payload type of
DatagramwasVec<u8>, thus theDatagramalways owned the UDP datagram payload.This change enables
Datagramto represent both (a) an owned payload allocation (i.e.Vec<u8>), but also represent (b) a view into an existing payload (e.g. a long lived receive buffer via&[u8]).The default payload type stays
Vec<u8>, thus not breaking existing usage ofDatagram.feat(transport): accept borrowed instead of owned input datagram
Previously
process_input(and the like) took aDatagram<Vec<u8>>as input. In other words, it required allocating each UDP datagram payload in a dedicatedVec<u8>before passing it toprocess_input.With this patch,
process_inputaccepts aDatagram<&[u8]>. In other words, it accepts aDatagramwith a borrowed view into an existing buffer (&[u8]), e.g. a long lived receive buffer.feat(udp): return borrowed Datagram on receive
Previously
recv_innerwould returnDatagram<Vec<u8>>. In other words, it would allocate a newVec<u8>for each UDP datagram payload.Now
recv_innerreads into a provided buffer and returnsDatagram<&[u8]>, i.e. it returns a view into the provided buffer without allocating.feat(bin): don't allocate in UDP recv path
Pass a long lived receive buffer to
neqo_udp::recv_inner, receiving an iterator ofDatagram<&[u8]>s pointing into that buffer, thus no longer allocating in UDP receive path.Extracted out of #2093.
Part of #1693.
Pull request stack: