Performance trends

Medians from the nightly runs, oldest at the left. Latest: e0414b65 on GitHub-hosted runners at 2026-10-28 04:36:34 UTC.

Each runner class is charted on its own axes and never joined into one line. A GitHub-hosted runner has noisy neighbours and no fixed CPU, so its figures are comparable with each other over time but not with the retired lab node’s. The README shields read the GitHub-hosted series.

Each panel carries one message size on its own scale. Bandwidth at 4 KiB and at 1 MiB differ by more than an order of magnitude, so a shared axis would flatten the small sizes onto the baseline and hide exactly the movement a regression shows up in. Bandwidth and latency are likewise never combined: they share no scale.

Send bandwidth over time

Median bw_peak_GBs for send (higher is better).

Measured on GitHub-hosted runners:

Last 10 runs, in GB/s:

Run	4 KiB	64 KiB	1 MiB
4479fc5d	1.15	10.12	22.92
dcc9ac9f	0.81	7.53	17.55
151379b0	1.39	9.83	14.45
95cb7235	1.49	11.42	24.34
b901e7e6	0.86	6.4	17.42
c9bef647	1	5.53	13.3
a969a3bb	–	–	–
9da810e2	1.65	11.44	23.72
2f0a80cc	1.23	8.11	21.07
e0414b65	1.49	13.01	24.79

Measured on the retired self-hosted lab node:

Last 3 runs, in GB/s:

Run	4 KiB	64 KiB	1 MiB
56b302de	0.36	2.56	6.02
9a44f18e	0.32	2.4	4.46
5e0516af	0.39	2.45	4.72

Send latency over time

Median lat_typical_us for send (lower is better).

Measured on GitHub-hosted runners:

Last 10 runs, in µs:

Run	4 KiB	64 KiB	1 MiB
4479fc5d	326.37	308.8	718.05
dcc9ac9f	335.1	361.95	942.58
151379b0	266.95	285.23	838.47
95cb7235	256.37	343.49	651.16
b901e7e6	346.94	376.11	920.28
c9bef647	362.29	391.12	1020.02
a969a3bb	–	–	–
9da810e2	254.58	259.91	513.93
2f0a80cc	347.34	273.77	594.8
e0414b65	329.68	327.93	538.22

Measured on the retired self-hosted lab node:

Last 3 runs, in µs:

Run	4 KiB	64 KiB	1 MiB
56b302de	321.46	325.81	468.44
9a44f18e	244.11	328.34	474.77
5e0516af	242.48	247.64	393.74

NVMe-oF read bandwidth over time

Median bw_avg_GBs for nvmeof (higher is better).

Measured on GitHub-hosted runners:

Last 10 runs, in GB/s:

Run	4 KiB	64 KiB	1 MiB
4479fc5d	0.104818	1.65297	3.13766
dcc9ac9f	0.035783	0.619257	1.89117
151379b0	0.112564	0.932681	3.35846
95cb7235	0.060744	0.736472	2.02672
b901e7e6	0.094514	1.10373	3.33257
c9bef647	0.037943	0.457612	1.91172
a969a3bb	0.034286	0.503941	1.28226
9da810e2	0.087094	1.33646	1.22671
2f0a80cc	0.088004	0.979161	1.35417
e0414b65	0.071585	0.603212	2.12094

S3-over-RDMA GET bandwidth over time

Median bw_avg_GBs for s3 (higher is better).

Measured on GitHub-hosted runners:

Last 10 runs, in GB/s:

Run	4 KiB	64 KiB	1 MiB
4479fc5d	0.005372	0.081574	1.15589
dcc9ac9f	0.00545	0.093053	1.2696
151379b0	0.003812	0.081243	0.733097
95cb7235	0.007331	0.116682	1.52898
b901e7e6	0.007785	0.107186	1.8163
c9bef647	0.008456	0.119994	1.95431
a969a3bb	0.00553	0.085895	1.17389
9da810e2	0.005324	0.084193	1.25818
2f0a80cc	0.004907	0.079662	1.23548
e0414b65	0.004922	0.085865	1.21928

P4 fp16-to-fp8 conversion bandwidth over time

Median bw_avg_GBs for p4cvt (higher is better).

Measured on GitHub-hosted runners:

Last 10 runs, in GB/s:

Run	4 KiB	64 KiB
4479fc5d	–	–
dcc9ac9f	–	–
151379b0	–	–
95cb7235	–	–
b901e7e6	–	–
c9bef647	0.013483	0.026705
a969a3bb	0.015136	0.038104
9da810e2	0.009988	0.018601
2f0a80cc	0.014568	0.037291
e0414b65	0.013607	0.026102