Fibre MPMC benchmarks

many producers, many consumers, work-sharing

Throughput in Melem/s: Million items per second per item sent, median of samples. The best result in each row is highlighted. In batched tables, the bracketed figure is the gain over the same implementation's single-item row. A dash means the implementation has no channel for that cell. Measured on MacBook M4 Pro (14 cores).

sync

Capacity	P×C	fibre	crossbeam	crossfire	flume	kanal
rendezvous	1x1	0.565	0.539	-	0.448	6.33
rendezvous	1x16	0.615	0.168	-	0.504	1.23
rendezvous	1x64	0.526	0.155	-	0.462	0.463
rendezvous	16x1	0.519	0.165	-	0.196	1.42
rendezvous	16x16	0.545	0.286	-	0.330	1.43
rendezvous	64x1	0.635	0.167	-	0.209	0.474
rendezvous	64x64	0.558	0.274	-	0.339	1.54
1	1x1	1.69	7.03	13.0	0.649	4.28
1	1x16	0.174	0.189	2.13	0.520	1.87
1	1x64	0.174	0.160	0.721	0.558	0.461
1	16x1	0.179	0.198	3.70	0.216	1.58
1	16x16	0.125	1.78	3.37	0.354	1.12
1	64x1	0.172	0.163	1.27	0.215	0.569
1	64x64	0.106	1.65	4.76	0.319	1.46
128	1x1	39.7	35.7	110.2	26.4	16.4
128	1x16	9.91	10.4	14.7	0.530	5.58
128	1x64	5.58	0.624	8.20	0.572	0.504
128	16x1	8.88	14.9	14.3	0.210	1.53
128	16x16	18.8	21.4	21.8	3.62	2.87
128	64x1	5.92	0.198	9.31	0.219	0.436
128	64x64	16.9	8.48	18.7	2.29	3.71
1024	1x1	52.9	63.0	238.0	6.50	40.1
1024	1x16	27.5	12.9	69.1	0.545	4.44
1024	1x64	20.4	12.8	31.2	0.521	0.492
1024	16x1	24.2	17.5	31.9	0.243	1.43
1024	16x16	43.9	43.1	66.7	10.9	3.17
1024	64x1	19.0	5.96	25.6	0.233	0.502
1024	64x64	36.0	16.1	49.4	7.37	4.24
unbounded	1x1	134.4	143.7	126.9	9.47	137.7
unbounded	1x16	0.500	36.9	43.3	0.539	20.7
unbounded	1x64	0.445	18.0	38.4	0.536	0.411
unbounded	16x1	12.0	23.3	22.7	43.2	13.6
unbounded	16x16	9.76	50.7	44.0	18.2	3.88
unbounded	64x1	13.3	26.1	34.8	43.3	7.98
unbounded	64x64	9.36	46.6	44.5	15.3	3.68

Unreliable, sample spread over 2x on a 1×1 cell where the rest of the matrix holds within a few percent: fibre at capacity unbounded (2.8x, 48.4 to 137.5), flume at capacity unbounded (2.3x, 6.41 to 15.0). Those medians do not support a comparison.

sync, batched (512 items per call)

Capacity	P×C	fibre	flume
128	1x1	135.2 (3.4x)	20.2 (0.8x)
128	1x16	12.1 (1.2x)	0.474 (0.9x)
128	1x64	12.1 (2.2x)	0.501 (0.9x)
128	16x1	15.6 (1.8x)	3.55 (16.9x)
128	16x16	29.5 (1.6x)	4.83 (1.3x)
128	64x1	8.65 (1.5x)	1.39 (6.3x)
128	64x64	25.1 (1.5x)	3.59 (1.6x)
1024	1x1	216.5 (4.1x)	72.4 (11.1x)
1024	1x16	63.6 (2.3x)	0.500 (0.9x)
1024	1x64	38.7 (1.9x)	0.487 (0.9x)
1024	16x1	65.1 (2.7x)	18.0 (74.1x)
1024	16x16	117.5 (2.7x)	11.7 (1.1x)
1024	64x1	39.3 (2.1x)	6.34 (27.2x)
1024	64x64	92.3 (2.6x)	8.34 (1.1x)
unbounded	1x1	225.6 (1.7x)	106.8 (11.3x)
unbounded	1x16	101.4 (202.8x)	0.474 (0.9x)
unbounded	1x64	91.1 (204.7x)	0.478 (0.9x)
unbounded	16x1	203.2 (16.9x)	49.4 (1.1x)
unbounded	16x16	157.3 (16.1x)	12.8 (0.7x)
unbounded	64x1	202.2 (15.2x)	57.3 (1.3x)
unbounded	64x64	151.2 (16.1x)	14.1 (0.9x)

Native batch: fibre none, flume receive only.

async

Capacity	P×C	fibre	async-channel	crossfire	flume	kanal
rendezvous	1x1	25.4	-	-	7.11	21.3
rendezvous	1x16	7.05	-	-	1.45	6.24
rendezvous	1x64	3.95	-	-	0.461	0.981
rendezvous	16x1	6.63	-	-	0.929	6.14
rendezvous	16x16	5.36	-	-	3.05	9.45
rendezvous	64x1	4.12	-	-	0.586	0.978
rendezvous	64x64	4.77	-	-	2.85	13.1
1	1x1	10.0	3.88	6.56	7.16	28.2
1	1x16	7.80	1.71	0.682	1.50	9.80
1	1x64	6.84	1.76	0.576	0.472	0.972
1	16x1	7.86	1.75	0.682	0.899	16.6
1	16x16	6.85	0.554	1.12	2.65	13.9
1	64x1	6.59	1.73	0.566	0.600	10.8
1	64x64	5.10	0.446	1.05	2.92	11.6
128	1x1	57.0	24.2	72.3	38.1	76.3
128	1x16	20.3	3.49	13.1	5.26	14.2
128	1x64	10.3	1.06	9.43	0.981	2.08
128	16x1	18.5	3.52	14.7	3.52	22.3
128	16x16	23.7	7.15	8.59	13.1	45.8
128	64x1	7.11	1.24	9.38	0.983	30.6
128	64x64	21.3	7.01	7.08	13.1	9.65
1024	1x1	58.8	21.2	85.1	23.6	81.2
1024	1x16	21.7	7.48	16.3	6.89	6.44
1024	1x64	15.0	2.28	14.2	1.67	2.93
1024	16x1	20.4	7.39	17.7	5.12	3.64
1024	16x16	30.2	9.51	14.7	16.3	7.66
1024	64x1	10.4	3.95	15.6	1.43	9.16
1024	64x64	25.7	9.65	15.8	16.1	20.8
unbounded	1x1	74.6	28.9	101.2	50.2	100.5
unbounded	1x16	29.0	17.1	29.2	7.41	6.58
unbounded	1x64	23.6	11.7	29.6	1.65	4.31
unbounded	16x1	13.0	10.3	41.2	28.2	19.3
unbounded	16x16	9.44	8.33	31.1	15.0	7.96
unbounded	64x1	11.9	9.87	20.4	25.7	23.3
unbounded	64x64	8.44	9.04	29.6	16.4	10.4

Unreliable, sample spread over 2x on a 1×1 cell where the rest of the matrix holds within a few percent: crossfire at capacity 1 (3.3x, 2.33 to 7.66). Those medians do not support a comparison.

async, batched (512 items per call)

Capacity	P×C	fibre	flume
128	1x1	155.4 (2.7x)	54.6 (1.4x)
128	1x16	76.3 (3.8x)	5.44 (1.0x)
128	1x64	29.0 (2.8x)	0.744 (0.8x)
128	16x1	58.5 (3.2x)	7.40 (2.1x)
128	16x16	74.8 (3.2x)	16.6 (1.3x)
128	64x1	21.5 (3.0x)	4.58 (4.7x)
128	64x64	72.0 (3.4x)	16.1 (1.2x)
1024	1x1	213.5 (3.6x)	57.1 (2.4x)
1024	1x16	130.8 (6.0x)	7.78 (1.1x)
1024	1x64	88.1 (5.9x)	1.02 (0.6x)
1024	16x1	115.1 (5.6x)	18.1 (3.5x)
1024	16x16	168.2 (5.6x)	26.2 (1.6x)
1024	64x1	70.4 (6.8x)	15.4 (10.7x)
1024	64x64	164.8 (6.4x)	24.9 (1.5x)
unbounded	1x1	182.9 (2.5x)	87.6 (1.7x)
unbounded	1x16	91.3 (3.1x)	7.95 (1.1x)
unbounded	1x64	93.2 (3.9x)	1.06 (0.6x)
unbounded	16x1	201.7 (15.5x)	39.1 (1.4x)
unbounded	16x16	125.2 (13.3x)	17.9 (1.2x)
unbounded	64x1	203.1 (17.1x)	38.3 (1.5x)
unbounded	64x64	118.2 (14.0x)	26.3 (1.6x)

Native batch: fibre none, flume receive only.