perf(bin/server): increase msg size and don't allocate msg per resp

[mxinden]

Previously neqo-server would respond to a request by repeatedly sending a static 440 byte message (Major-General's Song). Instead of sending 440 bytes, increase the batch size to 4096 bytes. This also matches the neqo-client receive buffer size.

neqo/neqo-bin/src/bin/client/http3.rs

Line 165 in 76630a5

let mut data = vec![0; 4096];

Previously ResponseData::repeat would convert the provided buf: &[u8] to Vec<u8>, i.e. re-allocate the buf. Instead keep a reference to the original buf, thus removing the allocation.

---

Preliminary benchmark comparison:

main:

1-conn/1-100mb-resp (aka. Download)/client
                        time:   [768.10 ms 1.0894 s 1.5517 s]
                        thrpt:  [64.443 MiB/s 91.794 MiB/s 130.19 MiB/s]

this pull request:

1-conn/1-100mb-resp (aka. Download)/client
                        time:   [674.30 ms 837.63 ms 1.0564 s]
                        thrpt:  [94.661 MiB/s 119.38 MiB/s 148.30 MiB/s]

Also visible in qviz where client reads MANY small (440 bytes) chunks from server:

I suggest merging #1758 first. We can then validate this pull request using the CI benchmark server.

I will take a look whether a long lived receive buffer on the client side has any impact next.

neqo/neqo-bin/src/bin/client/http3.rs

Line 165 in 76630a5

let mut data = vec![0; 4096];

[larseggert]

Fix clippy, but otherwise LGTM.

[github-actions]

Benchmark results

Performance differences relative to 76630a5.

• drain a timer quickly time: [391.24 ns 399.16 ns 406.56 ns]
  change: [-1.6765% +0.0373% +1.7660%] (p = 0.97 > 0.05)
  No change in performance detected.

• coalesce_acked_from_zero 1+1 entries
  time: [194.34 ns 194.80 ns 195.30 ns]
  change: [-0.2246% +0.3216% +1.0490%] (p = 0.39 > 0.05)
  No change in performance detected.

• coalesce_acked_from_zero 3+1 entries
  time: [234.86 ns 235.47 ns 236.11 ns]
  change: [-1.3377% -0.5925% +0.5348%] (p = 0.24 > 0.05)
  No change in performance detected.

• coalesce_acked_from_zero 10+1 entries
  time: [234.13 ns 234.80 ns 235.60 ns]
  change: [-1.4329% -0.8859% -0.3309%] (p = 0.00 < 0.05)
  Change within noise threshold.

• coalesce_acked_from_zero 1000+1 entries
  time: [213.97 ns 214.12 ns 214.28 ns]
  change: [-0.1185% +0.7805% +1.7211%] (p = 0.08 > 0.05)
  No change in performance detected.

• RxStreamOrderer::inbound_frame()
  time: [117.61 ms 117.70 ms 117.79 ms]
  change: [+0.1268% +0.2385% +0.3501%] (p = 0.00 < 0.05)
  Change within noise threshold.

• transfer/Run multiple transfers with varying seeds
  time: [117.35 ms 117.59 ms 117.83 ms]
  thrpt: [33.947 MiB/s 34.016 MiB/s 34.085 MiB/s]
  change:
  time: [-0.4683% -0.1775% +0.1387%] (p = 0.25 > 0.05)
  thrpt: [-0.1385% +0.1778% +0.4705%]
  No change in performance detected.

• transfer/Run multiple transfers with the same seed
  time: [118.27 ms 118.42 ms 118.57 ms]
  thrpt: [33.736 MiB/s 33.778 MiB/s 33.820 MiB/s]
  change:
  time: [-0.2297% -0.0505% +0.1355%] (p = 0.59 > 0.05)
  thrpt: [-0.1353% +0.0505% +0.2303%]
  No change in performance detected.

Client/server transfer results

Transfer of 134217728 bytes over loopback.

Client	Server	CC	Pacing	Mean [ms]	Min [ms]	Max [ms]	Relative
msquic	msquic			816.5 ± 345.1	372.3	1326.9	1.00
neqo	msquic	reno	on	1944.3 ± 30.1	1905.8	1997.6	1.00
neqo	msquic	reno		2110.8 ± 206.9	1909.8	2471.1	1.00
neqo	msquic	cubic	on	2010.2 ± 273.3	1753.1	2720.3	1.00
neqo	msquic	cubic		1973.4 ± 198.9	1812.2	2509.8	1.00
msquic	neqo	reno	on	4490.0 ± 306.8	4171.6	5159.4	1.00
msquic	neqo	reno		4444.6 ± 192.6	4164.6	4844.3	1.00
msquic	neqo	cubic	on	4571.1 ± 266.9	4272.5	5179.5	1.00
msquic	neqo	cubic		4431.7 ± 242.7	4142.3	4947.5	1.00
neqo	neqo	reno	on	3586.7 ± 237.4	3284.1	3950.7	1.00
neqo	neqo	reno		3630.4 ± 255.0	3102.2	4007.3	1.00
neqo	neqo	cubic	on	4428.5 ± 341.1	3779.6	4992.5	1.00
neqo	neqo	cubic		4365.0 ± 293.2	4071.9	4994.8	1.00

⬇️ Download logs