2.7.4 Performance Statistics

[james-pre]

v2.7.4 significantly improved VFS cache performance and added new benchmarks.

This release includes the addition of benchmarks using zbench, which are located in tests/perf. The new benchmarks use the InMemory backend.

Please note MB/s numbers are average throughput.

Read with createReadStream, cold cache (writeFile prep)

The vnode is evicted before the read, so the cache is repopulated from the backend.

configuration	v2.7.3	v2.7.4	speedup
4K ×2000	73.65 ms	68.22 ms	1.08×
256K ×200	162.8 ms	38.64 ms	4.21×
8M ×8	1.062 s	43.87 ms	24.2×
8M ×8, hwm 1M	1.181 s	20.75 ms	56.9×
(aggregate)	78.56 MB/s	1136 MB/s	14.5×

Read with createReadStream, warm cache (stream prep)

configuration	v2.7.3	v2.7.4	speedup
4K ×2000	83.6 ms	81.38 ms	1.03×
256K ×200	42.16 ms	20.41 ms	2.07×
8M ×8	45.57 ms	9.996 ms	4.56×
8M ×8, hwm 1M	36.23 ms	4.961 ms	7.3×
(aggregate)	938.7 MB/s	1669 MB/s	1.78×

Write with createWriteStream

configuration	v2.7.3	v2.7.4	speedup
4K ×2000	882 ms	753.7 ms	1.17×
256K ×200	67.31 ms	64.99 ms	1.04×
8M ×8	153.8 ms	47.09 ms	3.27×
8M ×8, hwm 1M	64.32 ms	27.95 ms	2.3×
(aggregate)	166.9 MB/s	218 MB/s	1.31×

Read across many cached regions

Per-read cost. On v2.7.3 this grows with the region count; afterwards it is flat.

configuration	v2.7.3	v2.7.4	speedup
128 regions	1.573 µs/read	1.483 µs/read	1.06×
512 regions	2.363 µs/read	1.366 µs/read	1.73×
2048 regions	6.342 µs/read	1.684 µs/read	3.77×
8192 regions	41.08 µs/read	1.792 µs/read	22.9×
(aggregate)	12.84 µs/read	1.581 µs/read	8.12×

Scattered writes into one file

configuration	v2.7.3	v2.7.4	speedup
512 writes	13.15 µs/write	4.506 µs/write	2.92×
2048 writes	24.64 µs/write	3.057 µs/write	8.06×
8192 writes	83.17 µs/write	2.566 µs/write	32.4×
(aggregate)	68.69 µs/write	2.752 µs/write	25×

Append 4 KB at a time

The change is from quadratic to constant time appends, so the speedup is also quadratic.

configuration	v2.7.3	v2.7.4	speedup
500 appends	23.38 µs/write	6.077 µs/write	3.85×
2000 appends	82.41 µs/write	4.242 µs/write	19.4×
8000 appends	596.5 µs/write	3.858 µs/write	155×
(aggregate)	471.3 µs/write	4.037 µs/write	117×