LCH Reed-Solomon Benchmarks

118 Benchmark rows

11,800 Raw samples

100 Repetitions per row

1 KiB Bytes per shard

Two-panel RS throughput chart. Encode compares LCH+AVX2, LCH+GFNI, XDRS, ISA-L, Jerasure, and klauspost. DecodeMax includes those series plus native Leopard for K at least 128. Full numerical medians follow in accessible tables. — Median CPU-time-derived input throughput on the pinned local test host. Higher is better. The y-axis is logarithmic and throughput is measured in GiB/s. On narrow screens, scroll the chart horizontally or use the complete tables below.

Measurement contract

Six series receive the same nine full-length K/R cases and deterministic maximum-erasure workload: two owned LCH backends plus XDRS, ISA-L, Jerasure, and klauspost. Native Leopard receives its five valid high-rate cases. External setup differences are made explicit rather than hidden behind a synthetic common API.

Code shape

K = 8, 16, 32, 64, 128, 192, 224, 240, and 248. Recovery count is R = 256 - K. Native Leopard is shown only where R is not greater than K.

DecodeMax

Exactly R symbols are shuffled across data and recovery positions. Timed throughput uses the paper-compatible K x bytes input metric, not recovered-output volume.

Run discipline

CPU0 pinned, 1 second warmup, at least 0.2 seconds per repetition, and randomized ten-repetition C++/Go chunks. The current harness verifies every row outside timing. This long run predates the native XDRS encode check; all nine XDRS encode configurations were subsequently output-verified under the final harness.

Representative medians

Selected points from the complete dataset, in GiB/s. The graph above contains every measured K value.

Backend	Encode K128	Encode K248	Decode K128	Decode K248
LCH+GFNI	8.09	25.87	2.79	13.96
LCH+AVX2	6.84	23.20	2.55	13.08
XDRS	5.98	19.43	2.45	8.56
Leopard	6.24	20.12	2.44	4.66
ISA-L	1.25	16.36	0.072	0.027
Jerasure	0.151	2.50	0.050	0.204
klauspost	0.494	7.86	0.069	0.185

Complete median data for every plotted point

Encode input throughput (GiB/s)

Backend	K8	K16	K32	K64	K128	K192	K224	K240	K248
LCH+GFNI	0.775	1.421	2.166	4.447	8.093	13.656	15.149	22.458	25.874
LCH+AVX2	0.686	1.168	1.886	3.661	6.844	11.446	13.717	18.634	23.198
XDRS	0.661	1.084	1.754	3.269	5.979	9.507	11.361	15.165	19.434
Leopard	not run	not run	not run	not run	6.240	10.053	11.859	15.954	20.118
ISA-L	0.779	0.828	0.875	1.014	1.252	2.301	4.493	8.992	16.356
Jerasure	0.0853	0.0758	0.0872	0.102	0.151	0.292	0.601	1.239	2.501
klauspost	0.258	0.264	0.283	0.328	0.494	0.975	1.957	3.888	7.861

DecodeMax input throughput (GiB/s)

Backend	K8	K16	K32	K64	K128	K192	K224	K240	K248
LCH+GFNI	0.405	0.849	1.126	2.214	2.792	4.535	6.200	9.567	13.958
LCH+AVX2	0.356	0.656	1.011	1.733	2.552	4.186	5.649	8.634	13.079
XDRS	0.371	0.639	1.012	1.690	2.453	3.921	5.122	6.925	8.558
Leopard	not run	not run	not run	not run	2.441	3.414	3.940	4.364	4.655
ISA-L	6.927	2.610	0.788	0.230	0.0718	0.0384	0.0312	0.0284	0.0271
Jerasure	1.759	0.625	0.255	0.0942	0.0503	0.0407	0.0661	0.110	0.204
klauspost	0.745	0.656	0.349	0.137	0.0693	0.0533	0.0777	0.120	0.185

This WSL/AMD run remains environmentally noisy: CPU-time coefficient of variation ranges from 7.2% to 41.3%. Treat medians as local comparative evidence, not universal hardware rankings. Native XDRS, ISA-L, Jerasure, and klauspost use their own standard-coordinate code families; Leopard uses the compatible Cantor representation.