Benchmark of BLAKE3 used for evaluation as a filesystem checksum, published here:
| Hash | Total cycles | Cycles/iteration | Perf vs BLAKE3 | Perf vs BLAKE2b |
|---|---|---|---|---|
| BLAKE3 (AVX2) | 111260245256 | 11126 | 1.0 | 0.876 (-13%) |
| BLAKE2b (AVX2) | 127009487092 | 12700 | 1.141 (+14%) | 1.0 |
| BLAKE2b (AVX) | 166426785907 | 16642 | 1.496 (+50%) | 1.310 (+31%) |
| BLAKE2b (ref) | 225053579540 | 22505 | 2.022 (+102%) | 1.772 (+77%) |
Block size is 4KiB, number of iterations: 1000000.
The performance improvement +14% against optimized version of BLAKE2b is smaller than expected (above 70%). To understand where this comes from, this repository hosts sources that can be used for testing and evaluation in other environments.
For continuity and calibration with older results, the tool also measures empty loop, plain memcpy and crc32c/xxhash/blake2b.
Reference host:
- Intel(R) Xeon(R) CPU E5-1620 v3 @ 3.50GHz, 16G of memory, AVX2 capable
Build:
$ make hash-speedtest
$ hash-speedtest [iterations]
Verify that published results and what's measured by current sources match. The BLAKE3 sources are from 0.1.1.
By default, 100K iterations are run, 1M is there to see if the overhead makes
any difference. There are 3 targets compiled with various -O flags.
Conclusion: published and current numbers match, -O flags don't bring
significant improvement.
Reference results with 100.000 iterations:
$ ./hash-speedtest.O1
Block size: 4096
Iterations: 100000
NULL-NOP: cycles: 49731562, c/i 497
NULL-MEMCPY: cycles: 66853283, c/i 668
CRC32C: cycles: 181899845, c/i 1818
XXHASH: cycles: 137766217, c/i 1377
BLAKE2b: cycles: 2274532357, c/i 22745
BLAKE3: cycles: 2400285391, c/i 24002
BLAKE3-SSE41: cycles: 1167069441, c/i 11670
BLAKE3-AVX2: cycles: 1167974091, c/i 11679
$ ./hash-speedtest.O2
Block size: 4096
Iterations: 100000
NULL-NOP: cycles: 38871962, c/i 388
NULL-MEMCPY: cycles: 60576411, c/i 605
CRC32C: cycles: 185520410, c/i 1855
XXHASH: cycles: 140206405, c/i 1402
BLAKE2b: cycles: 1871446270, c/i 18714
BLAKE3-port: cycles: 2047875279, c/i 20478
BLAKE3-SSE41: cycles: 1051016715, c/i 10510
BLAKE3-AVX2: cycles: 1052129371, c/i 10521
$ ./hash-speedtest.O3
Block size: 4096
Iterations: 100000
NULL-NOP: cycles: 46861642, c/i 468
NULL-MEMCPY: cycles: 58843572, c/i 588
CRC32C: cycles: 183644437, c/i 1836
XXHASH: cycles: 137361306, c/i 1373
BLAKE2b: cycles: 1859611165, c/i 18596
BLAKE3-port: cycles: 2044300312, c/i 20443
BLAKE3-SSE41: cycles: 1049436859, c/i 10494
BLAKE3-AVX2: cycles: 1051208264, c/i 10512
Reference results with 1.000.000 iterations:
$ ./hash-speedtest.O1 1000000
Block size: 4096
Iterations: 1000000
NULL-NOP: cycles: 311069562, c/i 311
NULL-MEMCPY: cycles: 600739577, c/i 600
CRC32C: cycles: 1744911994, c/i 1744
XXHASH: cycles: 1327265526, c/i 1327
BLAKE2b: cycles: 22500254356, c/i 22500
BLAKE3: cycles: 23999837826, c/i 23999
BLAKE3-SSE41: cycles: 11672269851, c/i 11672
BLAKE3-AVX2: cycles: 11677903076, c/i 11677
$ ./hash-speedtest.O2 1000000
Block size: 4096
Iterations: 1000000
NULL-NOP: cycles: 318622473, c/i 318
NULL-MEMCPY: cycles: 585150709, c/i 585
CRC32C: cycles: 1782021818, c/i 1782
XXHASH: cycles: 1332179797, c/i 1332
BLAKE2b: cycles: 18637886508, c/i 18637
BLAKE3-port: cycles: 20486804111, c/i 20486
BLAKE3-SSE41: cycles: 10511936730, c/i 10511
BLAKE3-AVX2: cycles: 10516193170, c/i 10516
$ ./hash-speedtest.O3 1000000
Block size: 4096
Iterations: 1000000
NULL-NOP: cycles: 312371265, c/i 312
NULL-MEMCPY: cycles: 596678141, c/i 596
CRC32C: cycles: 1796960915, c/i 1796
XXHASH: cycles: 1341290405, c/i 1341
BLAKE2b: cycles: 18554328509, c/i 18554
BLAKE3-port: cycles: 20471714973, c/i 20471
BLAKE3-SSE41: cycles: 10498995789, c/i 10498
BLAKE3-AVX2: cycles: 10504252272, c/i 10504
Sources: