refactor(ext/node): Use an internal DataView in Buffer #17815

aapoalas · 2023-02-18T15:11:29Z

Refactor the node Buffer class polyfill to take less code lines by using an internal DataView for reading. This improves performance in some places (reading signed bigints and floating point numbers, writing bigints and floating point numbers) and keeps the same performance roughly the same in others at the cost of increased creation time.

Measurements

Creation

Creation of the internal DataView requires quite a bit of extra work, making the creation of Buffer significantly slower, especially so for empty buffers.

It turns out creating an empty Uint8Array is optimized somehow in that it doesn't truly create an ArrayBuffer for the Uint8Array. Accessing the buffer property triggers the creation which then takes more than 400 nanoseconds. This is the difference between creation of an empty buffer in main vs in this PR.

This PR:

benchmark                 time (avg)        iter/s             (min … max)       p75       p99      p995
-------------------------------------------------------------------------- -----------------------------
empty creation           543.29 ns/iter   1,840,641.8 (341.72 ns … 719.75 ns) 593.96 ns 713.97 ns 719.75 ns
8 creation               712.03 ns/iter   1,404,431.2 (516.36 ns … 937.88 ns) 770.02 ns 937.88 ns 937.88 ns
64 creation               754.8 ns/iter   1,324,858.1       (458.7 ns … 1 µs) 838.74 ns      1 µs      1 µs
1024 * 1000 creation     292.61 µs/iter       3,417.5     (7.05 µs … 3.89 ms) 399.34 µs   3.01 ms   3.41 ms

Main:

benchmark                 time (avg)        iter/s             (min … max)       p75       p99      p995
-------------------------------------------------------------------------- -----------------------------
empty creation            94.88 ns/iter  10,540,038.9  (88.85 ns … 108.85 ns)   95.6 ns  98.02 ns 107.27 ns
8 creation               504.84 ns/iter   1,980,817.6 (315.56 ns … 788.94 ns) 695.35 ns 787.98 ns 788.94 ns
64 creation              562.65 ns/iter   1,777,307.9 (326.82 ns … 848.31 ns) 771.65 ns 848.31 ns 848.31 ns
1024 * 1000 creation     293.49 µs/iter       3,407.2    (17.37 µs … 3.84 ms) 389.35 µs    2.9 ms   3.01 ms

Reading integers

Reading integers is equivalent in performance or just the tiniest bit slower, less than a nanosecond of a difference.

This PR:

benchmark                           time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------ -----------------------------
read u8 non-zero offset              9.77 ns/iter 102,315,090.7     (9.6 ns … 28.11 ns)   9.69 ns  10.24 ns  11.23 ns
read i8 non-zero offset              10.4 ns/iter  96,111,809.3     (9.99 ns … 34.6 ns)  10.05 ns  17.74 ns  18.25 ns
read u16 LE non-zero offset         10.08 ns/iter  99,181,207.7    (9.99 ns … 30.17 ns)  10.05 ns   10.8 ns  11.32 ns
read u16 BE non-zero offset         10.32 ns/iter  96,865,482.3    (9.96 ns … 30.64 ns)  10.06 ns   16.4 ns  16.96 ns
read i16 LE non-zero offset         10.29 ns/iter  97,154,731.7      (10 ns … 31.94 ns)  10.08 ns  15.96 ns  16.49 ns
read i16 BE non-zero offset         10.24 ns/iter  97,630,253.9   (10.01 ns … 16.01 ns)  10.16 ns  15.48 ns  15.48 ns
read u24 LE non-zero offset         10.12 ns/iter  98,790,422.8    (9.77 ns … 19.03 ns)   9.95 ns  15.91 ns  16.44 ns
read u24 BE non-zero offset          9.97 ns/iter 100,327,698.2    (9.77 ns … 17.58 ns)   9.95 ns  10.56 ns   12.1 ns
read i24 LE non-zero offset          9.96 ns/iter 100,407,851.3    (9.77 ns … 16.55 ns)   9.97 ns  10.51 ns  10.68 ns
read i24 BE non-zero offset            10 ns/iter 100,011,343.1    (9.77 ns … 26.28 ns)   9.99 ns  10.64 ns  11.53 ns
read u32 LE non-zero offset          9.84 ns/iter 101,672,354.3    (9.77 ns … 12.51 ns)   9.83 ns  10.49 ns  10.87 ns
read u32 BE non-zero offset          9.85 ns/iter 101,528,044.6    (9.77 ns … 16.01 ns)   9.81 ns  10.56 ns   11.3 ns
read i32 LE non-zero offset          9.85 ns/iter 101,574,064.1     (9.77 ns … 18.6 ns)   9.81 ns  11.28 ns  11.71 ns
read i32 BE non-zero offset          9.91 ns/iter 100,916,939.4    (9.77 ns … 16.99 ns)   9.83 ns  15.48 ns  15.48 ns
read u40 LE non-zero offset         10.02 ns/iter  99,832,294.5    (9.77 ns … 16.47 ns)   9.96 ns  15.83 ns  15.87 ns
read u40 BE non-zero offset         10.26 ns/iter  97,474,628.4    (9.77 ns … 18.66 ns)   9.95 ns  15.99 ns  16.46 ns
read i40 LE non-zero offset          9.96 ns/iter 100,402,452.2    (9.77 ns … 18.59 ns)   9.95 ns  10.76 ns  11.57 ns
read i40 BE non-zero offset         10.14 ns/iter  98,659,108.9    (9.77 ns … 17.56 ns)   9.96 ns  15.96 ns  15.97 ns
read u48 LE non-zero offset          9.96 ns/iter 100,391,646.8    (9.77 ns … 16.75 ns)   9.95 ns   11.2 ns  11.87 ns
read u48 BE non-zero offset         10.04 ns/iter  99,601,062.0    (9.77 ns … 16.49 ns)   9.96 ns  15.96 ns  15.96 ns
read i48 LE non-zero offset         10.04 ns/iter  99,560,280.3    (9.77 ns … 17.77 ns)   9.96 ns  15.96 ns  15.96 ns
read i48 BE non-zero offset          9.96 ns/iter 100,426,793.7    (9.77 ns … 17.35 ns)   9.96 ns  10.57 ns  10.87 ns

Main:

benchmark                           time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------ -----------------------------
read u8 non-zero offset              9.75 ns/iter 102,593,084.9    (9.63 ns … 28.92 ns)   9.67 ns   10.8 ns  11.57 ns
read i8 non-zero offset              9.79 ns/iter 102,148,046.3    (9.59 ns … 29.82 ns)   9.68 ns  10.71 ns  14.77 ns
read u16 LE non-zero offset          9.71 ns/iter 103,008,946.4    (9.59 ns … 29.55 ns)   9.67 ns  10.41 ns  10.92 ns
read u16 BE non-zero offset          9.25 ns/iter 108,136,507.4    (9.05 ns … 16.37 ns)   9.25 ns   10.1 ns  10.68 ns
read i16 LE non-zero offset          9.74 ns/iter 102,624,042.1    (9.06 ns … 30.55 ns)   9.68 ns  10.97 ns  11.94 ns
read i16 BE non-zero offset           9.9 ns/iter 100,962,055.0    (9.67 ns … 15.34 ns)   9.79 ns  14.29 ns  14.29 ns
read u24 LE non-zero offset         10.01 ns/iter  99,928,791.1    (9.85 ns … 13.94 ns)  10.06 ns  10.55 ns  10.62 ns
read u24 BE non-zero offset          9.84 ns/iter 101,620,145.7    (9.59 ns … 16.74 ns)   9.77 ns  16.07 ns  16.14 ns
read i24 LE non-zero offset          9.73 ns/iter 102,736,462.4    (9.59 ns … 11.91 ns)   9.77 ns   10.3 ns  10.38 ns
read i24 BE non-zero offset          9.73 ns/iter 102,737,084.6       (9.53 ns … 14 ns)   9.77 ns   10.3 ns  10.42 ns
read u32 LE non-zero offset           9.8 ns/iter 102,053,625.7    (9.06 ns … 15.11 ns)   9.24 ns  14.29 ns  14.61 ns
read u32 BE non-zero offset          9.26 ns/iter 107,977,878.2    (9.06 ns … 10.53 ns)   9.25 ns   9.74 ns   9.77 ns
read i32 LE non-zero offset          9.24 ns/iter 108,194,503.6    (9.06 ns … 10.21 ns)   9.25 ns   9.74 ns   9.77 ns
read i32 BE non-zero offset          9.31 ns/iter 107,369,182.7    (9.06 ns … 16.48 ns)   9.26 ns     11 ns  12.53 ns
read u40 LE non-zero offset           9.8 ns/iter 102,074,231.4     (9.5 ns … 20.65 ns)   9.77 ns   13.5 ns  15.91 ns
read u40 BE non-zero offset          9.83 ns/iter 101,732,018.5    (9.58 ns … 28.68 ns)   9.77 ns   12.9 ns  14.44 ns
read i40 LE non-zero offset           9.8 ns/iter 102,054,738.3    (9.53 ns … 22.68 ns)   9.77 ns  12.61 ns  13.99 ns
read i40 BE non-zero offset         10.16 ns/iter  98,414,577.5    (9.53 ns … 35.51 ns)   9.77 ns  15.46 ns  15.88 ns
read u48 LE non-zero offset          9.81 ns/iter 101,965,387.7    (9.53 ns … 16.37 ns)   9.77 ns  11.74 ns  13.61 ns
read u48 BE non-zero offset          9.86 ns/iter 101,442,731.4    (9.53 ns … 17.92 ns)   9.77 ns  13.43 ns  14.84 ns
read i48 LE non-zero offset          9.66 ns/iter 103,484,172.8    (9.41 ns … 17.15 ns)   9.57 ns  12.53 ns  14.29 ns
read i48 BE non-zero offset           9.8 ns/iter 102,023,504.7    (9.53 ns … 16.61 ns)   9.77 ns  12.63 ns  14.38 ns

Reading floating point numbers

Reading floating points is faster by 2-3 nanoseconds (15-20%), bringing the speed in line with integer reads.

This PR:

benchmark                           time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------ -----------------------------
read float LE non-zero offset      10.15 ns/iter  98,492,448.8    (9.98 ns … 25.43 ns)  10.08 ns  11.76 ns   14.7 ns
read float BE non-zero offset      10.12 ns/iter  98,857,245.8   (10.01 ns … 27.07 ns)   10.1 ns  11.73 ns  12.41 ns
read double LE non-zero offset       9.98 ns/iter 100,227,692.1    (9.77 ns … 17.69 ns)   9.83 ns  15.48 ns  15.83 ns
read double BE non-zero offset      10.23 ns/iter  97,724,139.7    (9.77 ns … 21.27 ns)   9.84 ns  17.25 ns  17.58 ns

Main:

benchmark                           time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------ -----------------------------
read float LE non-zero offset      12.32 ns/iter  81,186,069.2    (11.75 ns … 30.7 ns)  12.39 ns  17.28 ns  18.19 ns
read float BE non-zero offset      12.32 ns/iter  81,173,186.0   (11.74 ns … 30.96 ns)  12.39 ns  13.32 ns  13.96 ns
read double LE non-zero offset      12.64 ns/iter  79,089,114.7   (12.39 ns … 20.62 ns)   12.6 ns  17.74 ns  18.29 ns
read double BE non-zero offset       12.5 ns/iter  79,996,955.4   (12.39 ns … 20.59 ns)  12.39 ns  14.99 ns  16.37 ns

Reading BigInts

Reading signed 64-bit integers is significantly faster, about 33% or so. Reading unsigned integers is significantly slower by more than 100%.

This PR:

benchmark                           time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------ -----------------------------
read u64 LE non-zero offset         25.03 ns/iter  39,947,827.6   (22.63 ns … 34.35 ns)  24.57 ns  31.28 ns  31.99 ns
read u64 BE non-zero offset         22.95 ns/iter  43,568,064.9   (20.81 ns … 33.38 ns)  23.24 ns  29.64 ns  29.85 ns
read i64 LE non-zero offset         21.37 ns/iter  46,790,864.8    (20.5 ns … 33.54 ns)  20.78 ns  27.85 ns  28.65 ns
read i64 BE non-zero offset         20.51 ns/iter  48,761,472.4   (19.78 ns … 57.56 ns)  19.84 ns  27.07 ns  27.58 ns

Main:

benchmark                           time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------ -----------------------------
read u64 LE non-zero offset          9.59 ns/iter 104,324,029.6    (9.06 ns … 17.76 ns)   9.29 ns  16.49 ns  16.88 ns
read u64 BE non-zero offset          9.36 ns/iter 106,848,018.2    (9.06 ns … 16.08 ns)   9.32 ns  11.79 ns  14.03 ns
read i64 LE non-zero offset          32.3 ns/iter  30,962,568.5   (31.44 ns … 47.85 ns)  31.98 ns  40.05 ns   40.8 ns
read i64 BE non-zero offset         32.43 ns/iter  30,832,610.6   (31.68 ns … 46.09 ns)  32.07 ns  39.61 ns     40 ns

Writing values

Writing values is mostly equivalent in speed, but writing BigInts and floating point numbers is faster by 10-30%.

This PR:

benchmark                            time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------- -----------------------------
write u8 non-zero offset             10.27 ns/iter  97,337,490.9   (10.01 ns … 29.43 ns)  10.25 ns  11.19 ns  11.87 ns
write i8 non-zero offset             10.27 ns/iter  97,416,669.4   (10.07 ns … 29.55 ns)  10.22 ns  11.02 ns  12.93 ns
write u16 LE non-zero offset         10.45 ns/iter  95,704,782.3   (10.35 ns … 14.56 ns)  10.48 ns     11 ns  11.18 ns
write u16 BE non-zero offset         10.91 ns/iter  91,638,842.8    (10.25 ns … 31.7 ns)  10.91 ns  11.47 ns  11.81 ns
write i16 LE non-zero offset         11.12 ns/iter  89,895,929.8    (10.56 ns … 21.7 ns)  10.92 ns  16.33 ns  17.18 ns
write i16 BE non-zero offset         11.04 ns/iter  90,544,662.0   (10.72 ns … 23.08 ns)  10.96 ns   15.5 ns  17.14 ns
write u24 LE non-zero offset         10.74 ns/iter  93,087,180.9   (10.49 ns … 17.18 ns)  10.72 ns  11.98 ns  13.23 ns
write u24 BE non-zero offset         10.74 ns/iter  93,126,589.8   (10.48 ns … 16.56 ns)  10.76 ns  12.44 ns   13.5 ns
write i24 LE non-zero offset         10.74 ns/iter  93,111,024.0   (10.52 ns … 18.52 ns)  10.72 ns  11.86 ns  12.49 ns
write i24 BE non-zero offset         10.81 ns/iter  92,493,301.9   (10.48 ns … 21.72 ns)  10.76 ns  13.38 ns  17.93 ns
write u32 LE non-zero offset          10.4 ns/iter  96,159,238.9   (10.25 ns … 17.32 ns)  10.39 ns  12.47 ns  16.39 ns
write u32 BE non-zero offset         10.47 ns/iter  95,469,436.4    (10.25 ns … 17.3 ns)  10.46 ns   11.9 ns  12.37 ns
write i32 LE non-zero offset         10.36 ns/iter  96,523,279.9   (10.25 ns … 14.03 ns)  10.39 ns  11.47 ns  12.22 ns
write i32 BE non-zero offset         10.44 ns/iter  95,787,841.4   (10.25 ns … 15.36 ns)  10.46 ns  11.31 ns  12.84 ns
write u40 LE non-zero offset          16.4 ns/iter  60,989,837.4   (16.32 ns … 23.22 ns)  16.32 ns  18.12 ns  18.86 ns
write u40 BE non-zero offset         16.91 ns/iter  59,119,700.8   (16.32 ns … 31.57 ns)  17.39 ns  24.83 ns  27.91 ns
write i40 LE non-zero offset         18.16 ns/iter  55,080,096.4    (18.1 ns … 21.69 ns)  18.11 ns  19.01 ns  19.44 ns
write i40 BE non-zero offset         18.65 ns/iter  53,623,209.6    (18.1 ns … 23.69 ns)  18.82 ns  20.25 ns  20.25 ns
write u48 LE non-zero offset         16.52 ns/iter  60,545,392.7   (16.32 ns … 32.78 ns)  16.32 ns  24.15 ns   25.5 ns
write u48 BE non-zero offset         16.37 ns/iter  61,100,467.9   (16.32 ns … 20.65 ns)  16.32 ns  17.09 ns  17.66 ns
write i48 LE non-zero offset         19.34 ns/iter  51,693,197.8   (18.82 ns … 24.46 ns)  20.25 ns  20.78 ns  21.16 ns
write i48 BE non-zero offset         18.87 ns/iter  53,000,896.8   (18.82 ns … 22.35 ns)  18.82 ns  19.62 ns  19.99 ns
write u64 LE non-zero offset        123.74 ns/iter   8,081,304.0 (109.67 ns … 140.39 ns) 127.06 ns 137.01 ns 138.87 ns
write u64 BE non-zero offset        123.27 ns/iter   8,112,561.4  (109.37 ns … 139.7 ns) 126.26 ns  135.5 ns 136.67 ns
write i64 LE non-zero offset         202.3 ns/iter   4,943,194.0 (194.01 ns … 226.05 ns) 206.02 ns 218.03 ns 219.62 ns
write i64 BE non-zero offset        203.23 ns/iter   4,920,459.3  (192.9 ns … 220.72 ns)  207.5 ns 218.17 ns  218.2 ns
write float LE non-zero offset        9.83 ns/iter 101,747,626.5    (9.77 ns … 13.62 ns)   9.85 ns  10.41 ns  10.86 ns
write float BE non-zero offset       10.01 ns/iter  99,948,359.7    (9.77 ns … 14.59 ns)  10.01 ns  10.55 ns  10.74 ns
write double LE non-zero offset      10.06 ns/iter  99,381,501.6    (9.85 ns … 19.18 ns)   9.97 ns  15.63 ns  16.07 ns
write double BE non-zero offset      10.05 ns/iter  99,494,774.1    (9.95 ns … 13.34 ns)  10.09 ns  10.59 ns  10.78 ns

Main:

benchmark                            time (avg)        iter/s             (min … max)       p75       p99      p995
------------------------------------------------------------------------------------- -----------------------------
write u8 non-zero offset             10.34 ns/iter  96,695,054.4    (9.97 ns … 41.91 ns)  10.18 ns  17.37 ns  18.28 ns
write i8 non-zero offset             10.21 ns/iter  97,963,814.8    (9.95 ns … 34.21 ns)  10.15 ns  12.13 ns  15.38 ns
write u16 LE non-zero offset         11.01 ns/iter  90,837,591.3   (10.66 ns … 28.85 ns)  11.04 ns   12.6 ns  13.64 ns
write u16 BE non-zero offset         11.06 ns/iter  90,416,177.5   (10.72 ns … 32.32 ns)  11.04 ns  11.66 ns  25.02 ns
write i16 LE non-zero offset         10.95 ns/iter  91,322,430.7   (10.74 ns … 20.01 ns)  10.95 ns  11.87 ns  12.36 ns
write i16 BE non-zero offset         10.74 ns/iter  93,141,189.0   (10.31 ns … 19.98 ns)  10.55 ns  17.48 ns  17.76 ns
write u24 LE non-zero offset         10.65 ns/iter  93,911,183.4   (10.49 ns … 16.68 ns)  10.63 ns  11.88 ns  12.42 ns
write u24 BE non-zero offset         10.65 ns/iter  93,867,321.8   (10.52 ns … 16.68 ns)  10.65 ns  11.31 ns  11.72 ns
write i24 LE non-zero offset         10.67 ns/iter  93,680,913.6   (10.54 ns … 13.26 ns)  10.75 ns   11.2 ns  11.28 ns
write i24 BE non-zero offset         10.66 ns/iter  93,831,171.4   (10.49 ns … 13.11 ns)  10.69 ns   11.3 ns  11.66 ns
write u32 LE non-zero offset         10.76 ns/iter  92,947,006.7   (10.59 ns … 17.33 ns)  10.72 ns  11.67 ns   12.5 ns
write u32 BE non-zero offset         10.75 ns/iter  92,983,545.6       (10.6 ns … 18 ns)  10.72 ns  11.53 ns  12.53 ns
write i32 LE non-zero offset         10.74 ns/iter  93,113,607.9   (10.56 ns … 17.89 ns)  10.72 ns  11.31 ns  11.66 ns
write i32 BE non-zero offset         10.74 ns/iter  93,109,369.7   (10.48 ns … 16.77 ns)  10.72 ns  11.36 ns   11.8 ns
write u40 LE non-zero offset         19.62 ns/iter  50,963,332.3   (17.39 ns … 26.05 ns)  19.48 ns  22.44 ns  22.44 ns
write u40 BE non-zero offset         19.41 ns/iter  51,511,903.1   (17.39 ns … 26.92 ns)  19.46 ns  22.44 ns  22.47 ns
write i40 LE non-zero offset         18.41 ns/iter  54,319,022.7   (18.34 ns … 28.79 ns)  18.34 ns  19.54 ns  19.95 ns
write i40 BE non-zero offset          18.4 ns/iter  54,358,750.5   (17.15 ns … 25.91 ns)  18.34 ns  19.43 ns  19.95 ns
write u48 LE non-zero offset         19.27 ns/iter  51,882,890.4   (18.02 ns … 26.63 ns)  19.18 ns  21.88 ns  21.91 ns
write u48 BE non-zero offset          19.4 ns/iter  51,540,337.5   (18.02 ns … 29.05 ns)  19.18 ns     25 ns  25.05 ns
write i48 LE non-zero offset          18.7 ns/iter  53,464,774.9   (17.41 ns … 31.61 ns)  18.58 ns  21.91 ns  23.83 ns
write i48 BE non-zero offset         28.23 ns/iter  35,422,951.2    (28.1 ns … 39.73 ns)  28.11 ns  30.43 ns  30.64 ns
write u64 LE non-zero offset        156.17 ns/iter   6,403,144.5 (147.57 ns … 175.35 ns) 160.62 ns 166.04 ns 167.47 ns
write u64 BE non-zero offset        152.79 ns/iter   6,544,968.0 (139.48 ns … 201.68 ns) 157.12 ns 173.33 ns  198.8 ns
write i64 LE non-zero offset        258.36 ns/iter   3,870,540.4 (244.96 ns … 379.12 ns) 259.76 ns 368.95 ns 369.44 ns
write i64 BE non-zero offset        258.05 ns/iter   3,875,267.2 (241.41 ns … 321.28 ns) 261.87 ns 268.66 ns 268.88 ns
write float LE non-zero offset       12.53 ns/iter  79,823,725.9    (12.15 ns … 28.6 ns)  12.51 ns  13.09 ns  13.28 ns
write float BE non-zero offset       12.52 ns/iter  79,841,979.0   (12.32 ns … 16.22 ns)  12.52 ns  13.25 ns  13.74 ns
write double LE non-zero offset      13.68 ns/iter  73,106,632.6   (13.44 ns … 35.19 ns)  13.69 ns  14.28 ns  14.53 ns
write double BE non-zero offset      13.73 ns/iter  72,841,178.5   (13.46 ns … 21.24 ns)  13.69 ns  15.55 ns  18.34 ns

V8 bugs spawned from this work

DataView#getBigUint64 has lackluster performance: https://bugs.chromium.org/p/v8/issues/detail?id=14263

It should not be the case that manually reading the 8 bytes of a 64-bit unsigned integer one by one and adding them together with appropriate bitshifts or multiplications, including a conversion of upper and lower 4 bytes to BigInt and a final BigInt-bitshift and addition, is faster than simply having V8 read the 64-bit unsigned integer and create a BigInt out of that. This seems like a bug in the engine.

Extending Uint8Array for a class base (or presumably any other TypedArray at least) leads to all of its methods losing performance: https://bugs.chromium.org/p/v8/issues/detail?id=14265

If Buffer were a class with an internal #buffer and #view then the performance of all its methods would be 10-30% better based on my testing. This applies even to methods that do not use this at all. This likewise seems like a bug, or possibly a limitation, of the engine.

bartlomieju · 2023-02-20T13:05:19Z

Maybe let's wait on #17827 to land before merging this one.

bartlomieju · 2023-02-20T16:38:13Z

@aapoalas it appears that some tests related to buffer started failing with these changes.

Extracted from #17815 Optimise Buffer's string operations, most significantly when dealing with ASCII and UTF-16. Base64 and HEX encodings are affected to much lesser degrees. ## Performance ### String length 15 With very small strings we're at break-even or sometimes even lose a tad bit of performance from creating a `DataView` that ends up not paying for itself. **This PR:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 1.15 µs/iter 871,388.6 (728.78 ns … 1.56 µs) 1.23 µs 1.56 µs 1.56 µs Buffer.from base64 string 1.63 µs/iter 612,790.9 (1.31 µs … 1.96 µs) 1.77 µs 1.96 µs 1.96 µs Buffer.from utf16 string 1.41 µs/iter 707,396.3 (915.24 ns … 1.93 µs) 1.61 µs 1.93 µs 1.93 µs Buffer.from hex string 1.87 µs/iter 535,357.9 (1.56 µs … 2.19 µs) 2 µs 2.19 µs 2.19 µs Buffer.toString ascii string 154.58 ns/iter 6,469,162.8 (149.69 ns … 198 ns) 154.51 ns 182.89 ns 191.91 ns Buffer.toString base64 string 161.65 ns/iter 6,186,189.6 (150.91 ns … 181.15 ns) 165.18 ns 171.87 ns 174.94 ns Buffer.toString utf16 string 292.74 ns/iter 3,415,959.8 (285.43 ns … 312.47 ns) 295.25 ns 310.47 ns 312.47 ns Buffer.toString hex string 89.61 ns/iter 11,159,315.6 (81.09 ns … 123.77 ns) 91.09 ns 113.62 ns 119.28 ns ``` **Main:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 1.26 µs/iter 794,875.8 (1.07 µs … 1.46 µs) 1.31 µs 1.46 µs 1.46 µs Buffer.from base64 string 1.65 µs/iter 607,853.3 (1.38 µs … 2.01 µs) 1.69 µs 2.01 µs 2.01 µs Buffer.from utf16 string 1.34 µs/iter 744,894.6 (1.09 µs … 1.55 µs) 1.45 µs 1.55 µs 1.55 µs Buffer.from hex string 2.01 µs/iter 496,345.8 (1.54 µs … 2.6 µs) 2.26 µs 2.6 µs 2.6 µs Buffer.toString ascii string 150.16 ns/iter 6,659,630.5 (144.99 ns … 166.68 ns) 152.4 ns 157.26 ns 159.14 ns Buffer.toString base64 string 164.73 ns/iter 6,070,692.0 (158.77 ns … 185.63 ns) 168.48 ns 175.74 ns 176.68 ns Buffer.toString utf16 string 150.61 ns/iter 6,639,864.0 (148.2 ns … 168.29 ns) 150.93 ns 157.21 ns 168.15 ns Buffer.toString hex string 94.21 ns/iter 10,614,972.9 (86.21 ns … 98.75 ns) 95.43 ns 97.99 ns 98.21 ns ``` ### String length 1500 With moderate lengths we already see great upsides for `Buffer.from()` with ASCII and UTF-16. **This PR:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 5.79 µs/iter 172,562.6 (4.72 µs … 4.71 ms) 5.04 µs 10.3 µs 11.67 µs Buffer.from base64 string 5.08 µs/iter 196,678.9 (4.97 µs … 5.76 µs) 5.08 µs 5.76 µs 5.76 µs Buffer.from utf16 string 9.68 µs/iter 103,316.5 (7.14 µs … 3.44 ms) 10.32 µs 13.42 µs 15.21 µs Buffer.from hex string 53.7 µs/iter 18,620.2 (49.37 µs … 2.2 ms) 54.74 µs 72.2 µs 81.07 µs Buffer.toString ascii string 6.63 µs/iter 150,761.3 (5.59 µs … 1.11 ms) 6.08 µs 15.68 µs 24.77 µs Buffer.toString base64 string 460.57 ns/iter 2,171,224.4 (448.33 ns … 511.73 ns) 465.05 ns 495.54 ns 511.73 ns Buffer.toString utf16 string 6.52 µs/iter 153,287.0 (6.47 µs … 6.66 µs) 6.53 µs 6.66 µs 6.66 µs Buffer.toString hex string 3.68 µs/iter 271,965.4 (3.64 µs … 3.82 µs) 3.68 µs 3.82 µs 3.82 µs ``` **Main:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 11.46 µs/iter 87,298.1 (8.53 µs … 834.1 µs) 9.61 µs 83.31 µs 87.3 µs Buffer.from base64 string 5.4 µs/iter 185,027.8 (5.07 µs … 7.49 µs) 5.44 µs 7.49 µs 7.49 µs Buffer.from utf16 string 20.3 µs/iter 49,270.8 (13.55 µs … 649.11 µs) 18.8 µs 113.93 µs 125.17 µs Buffer.from hex string 52.03 µs/iter 19,218.9 (48.74 µs … 2.59 ms) 52.84 µs 67.05 µs 73.56 µs Buffer.toString ascii string 6.46 µs/iter 154,822.5 (6.32 µs … 6.69 µs) 6.52 µs 6.69 µs 6.69 µs Buffer.toString base64 string 440.19 ns/iter 2,271,764.6 (427 ns … 490.77 ns) 444.74 ns 484.64 ns 490.77 ns Buffer.toString utf16 string 6.89 µs/iter 145,106.7 (6.81 µs … 7.24 µs) 6.91 µs 7.24 µs 7.24 µs Buffer.toString hex string 3.66 µs/iter 273,456.5 (3.6 µs … 4.02 µs) 3.64 µs 4.02 µs 4.02 µs ``` ### String length 2^20 With massive lengths we the difference in ASCII and UTF-16 parsing performance is enormous. **This PR:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 4.1 ms/iter 243.7 (2.64 ms … 6.74 ms) 4.43 ms 6.26 ms 6.74 ms Buffer.from base64 string 3.74 ms/iter 267.6 (2.91 ms … 4.92 ms) 3.96 ms 4.31 ms 4.92 ms Buffer.from utf16 string 7.72 ms/iter 129.5 (5.91 ms … 11.03 ms) 7.97 ms 11.03 ms 11.03 ms Buffer.from hex string 35.72 ms/iter 28.0 (34.71 ms … 38.42 ms) 35.93 ms 38.42 ms 38.42 ms Buffer.toString ascii string 78.92 ms/iter 12.7 (42.72 ms … 94.13 ms) 91.64 ms 94.13 ms 94.13 ms Buffer.toString base64 string 833.62 µs/iter 1,199.6 (638.05 µs … 5.97 ms) 826.86 µs 2.45 ms 2.48 ms Buffer.toString utf16 string 79.35 ms/iter 12.6 (69.72 ms … 88.9 ms) 86.66 ms 88.9 ms 88.9 ms Buffer.toString hex string 31.04 ms/iter 32.2 (4.3 ms … 46.9 ms) 37.21 ms 46.9 ms 46.9 ms ``` **Main:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 18.66 ms/iter 53.6 (15.61 ms … 23.26 ms) 20.62 ms 23.26 ms 23.26 ms Buffer.from base64 string 4.7 ms/iter 212.9 (2.94 ms … 9.07 ms) 4.65 ms 9.06 ms 9.07 ms Buffer.from utf16 string 33.49 ms/iter 29.9 (31.24 ms … 35.67 ms) 34.08 ms 35.67 ms 35.67 ms Buffer.from hex string 39.38 ms/iter 25.4 (38.66 ms … 42.36 ms) 39.58 ms 42.36 ms 42.36 ms Buffer.toString ascii string 77.68 ms/iter 12.9 (67.46 ms … 95.68 ms) 84.71 ms 95.68 ms 95.68 ms Buffer.toString base64 string 825.53 µs/iter 1,211.3 (655.38 µs … 6.69 ms) 816.62 µs 3.07 ms 3.13 ms Buffer.toString utf16 string 76.54 ms/iter 13.1 (66.9 ms … 85.26 ms) 83.63 ms 85.26 ms 85.26 ms Buffer.toString hex string 38.56 ms/iter 25.9 (33.83 ms … 46.56 ms) 45.33 ms 46.56 ms 46.56 ms ```

Extracted from denoland#17815 Optimise Buffer's string operations, most significantly when dealing with ASCII and UTF-16. Base64 and HEX encodings are affected to much lesser degrees. ## Performance ### String length 15 With very small strings we're at break-even or sometimes even lose a tad bit of performance from creating a `DataView` that ends up not paying for itself. **This PR:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 1.15 µs/iter 871,388.6 (728.78 ns … 1.56 µs) 1.23 µs 1.56 µs 1.56 µs Buffer.from base64 string 1.63 µs/iter 612,790.9 (1.31 µs … 1.96 µs) 1.77 µs 1.96 µs 1.96 µs Buffer.from utf16 string 1.41 µs/iter 707,396.3 (915.24 ns … 1.93 µs) 1.61 µs 1.93 µs 1.93 µs Buffer.from hex string 1.87 µs/iter 535,357.9 (1.56 µs … 2.19 µs) 2 µs 2.19 µs 2.19 µs Buffer.toString ascii string 154.58 ns/iter 6,469,162.8 (149.69 ns … 198 ns) 154.51 ns 182.89 ns 191.91 ns Buffer.toString base64 string 161.65 ns/iter 6,186,189.6 (150.91 ns … 181.15 ns) 165.18 ns 171.87 ns 174.94 ns Buffer.toString utf16 string 292.74 ns/iter 3,415,959.8 (285.43 ns … 312.47 ns) 295.25 ns 310.47 ns 312.47 ns Buffer.toString hex string 89.61 ns/iter 11,159,315.6 (81.09 ns … 123.77 ns) 91.09 ns 113.62 ns 119.28 ns ``` **Main:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 1.26 µs/iter 794,875.8 (1.07 µs … 1.46 µs) 1.31 µs 1.46 µs 1.46 µs Buffer.from base64 string 1.65 µs/iter 607,853.3 (1.38 µs … 2.01 µs) 1.69 µs 2.01 µs 2.01 µs Buffer.from utf16 string 1.34 µs/iter 744,894.6 (1.09 µs … 1.55 µs) 1.45 µs 1.55 µs 1.55 µs Buffer.from hex string 2.01 µs/iter 496,345.8 (1.54 µs … 2.6 µs) 2.26 µs 2.6 µs 2.6 µs Buffer.toString ascii string 150.16 ns/iter 6,659,630.5 (144.99 ns … 166.68 ns) 152.4 ns 157.26 ns 159.14 ns Buffer.toString base64 string 164.73 ns/iter 6,070,692.0 (158.77 ns … 185.63 ns) 168.48 ns 175.74 ns 176.68 ns Buffer.toString utf16 string 150.61 ns/iter 6,639,864.0 (148.2 ns … 168.29 ns) 150.93 ns 157.21 ns 168.15 ns Buffer.toString hex string 94.21 ns/iter 10,614,972.9 (86.21 ns … 98.75 ns) 95.43 ns 97.99 ns 98.21 ns ``` ### String length 1500 With moderate lengths we already see great upsides for `Buffer.from()` with ASCII and UTF-16. **This PR:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 5.79 µs/iter 172,562.6 (4.72 µs … 4.71 ms) 5.04 µs 10.3 µs 11.67 µs Buffer.from base64 string 5.08 µs/iter 196,678.9 (4.97 µs … 5.76 µs) 5.08 µs 5.76 µs 5.76 µs Buffer.from utf16 string 9.68 µs/iter 103,316.5 (7.14 µs … 3.44 ms) 10.32 µs 13.42 µs 15.21 µs Buffer.from hex string 53.7 µs/iter 18,620.2 (49.37 µs … 2.2 ms) 54.74 µs 72.2 µs 81.07 µs Buffer.toString ascii string 6.63 µs/iter 150,761.3 (5.59 µs … 1.11 ms) 6.08 µs 15.68 µs 24.77 µs Buffer.toString base64 string 460.57 ns/iter 2,171,224.4 (448.33 ns … 511.73 ns) 465.05 ns 495.54 ns 511.73 ns Buffer.toString utf16 string 6.52 µs/iter 153,287.0 (6.47 µs … 6.66 µs) 6.53 µs 6.66 µs 6.66 µs Buffer.toString hex string 3.68 µs/iter 271,965.4 (3.64 µs … 3.82 µs) 3.68 µs 3.82 µs 3.82 µs ``` **Main:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 11.46 µs/iter 87,298.1 (8.53 µs … 834.1 µs) 9.61 µs 83.31 µs 87.3 µs Buffer.from base64 string 5.4 µs/iter 185,027.8 (5.07 µs … 7.49 µs) 5.44 µs 7.49 µs 7.49 µs Buffer.from utf16 string 20.3 µs/iter 49,270.8 (13.55 µs … 649.11 µs) 18.8 µs 113.93 µs 125.17 µs Buffer.from hex string 52.03 µs/iter 19,218.9 (48.74 µs … 2.59 ms) 52.84 µs 67.05 µs 73.56 µs Buffer.toString ascii string 6.46 µs/iter 154,822.5 (6.32 µs … 6.69 µs) 6.52 µs 6.69 µs 6.69 µs Buffer.toString base64 string 440.19 ns/iter 2,271,764.6 (427 ns … 490.77 ns) 444.74 ns 484.64 ns 490.77 ns Buffer.toString utf16 string 6.89 µs/iter 145,106.7 (6.81 µs … 7.24 µs) 6.91 µs 7.24 µs 7.24 µs Buffer.toString hex string 3.66 µs/iter 273,456.5 (3.6 µs … 4.02 µs) 3.64 µs 4.02 µs 4.02 µs ``` ### String length 2^20 With massive lengths we the difference in ASCII and UTF-16 parsing performance is enormous. **This PR:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 4.1 ms/iter 243.7 (2.64 ms … 6.74 ms) 4.43 ms 6.26 ms 6.74 ms Buffer.from base64 string 3.74 ms/iter 267.6 (2.91 ms … 4.92 ms) 3.96 ms 4.31 ms 4.92 ms Buffer.from utf16 string 7.72 ms/iter 129.5 (5.91 ms … 11.03 ms) 7.97 ms 11.03 ms 11.03 ms Buffer.from hex string 35.72 ms/iter 28.0 (34.71 ms … 38.42 ms) 35.93 ms 38.42 ms 38.42 ms Buffer.toString ascii string 78.92 ms/iter 12.7 (42.72 ms … 94.13 ms) 91.64 ms 94.13 ms 94.13 ms Buffer.toString base64 string 833.62 µs/iter 1,199.6 (638.05 µs … 5.97 ms) 826.86 µs 2.45 ms 2.48 ms Buffer.toString utf16 string 79.35 ms/iter 12.6 (69.72 ms … 88.9 ms) 86.66 ms 88.9 ms 88.9 ms Buffer.toString hex string 31.04 ms/iter 32.2 (4.3 ms … 46.9 ms) 37.21 ms 46.9 ms 46.9 ms ``` **Main:** ``` benchmark time (avg) iter/s (min … max) p75 p99 p995 -------------------------------------------------------------------------------------------------------------------- ----------------------------- Buffer.from ascii string 18.66 ms/iter 53.6 (15.61 ms … 23.26 ms) 20.62 ms 23.26 ms 23.26 ms Buffer.from base64 string 4.7 ms/iter 212.9 (2.94 ms … 9.07 ms) 4.65 ms 9.06 ms 9.07 ms Buffer.from utf16 string 33.49 ms/iter 29.9 (31.24 ms … 35.67 ms) 34.08 ms 35.67 ms 35.67 ms Buffer.from hex string 39.38 ms/iter 25.4 (38.66 ms … 42.36 ms) 39.58 ms 42.36 ms 42.36 ms Buffer.toString ascii string 77.68 ms/iter 12.9 (67.46 ms … 95.68 ms) 84.71 ms 95.68 ms 95.68 ms Buffer.toString base64 string 825.53 µs/iter 1,211.3 (655.38 µs … 6.69 ms) 816.62 µs 3.07 ms 3.13 ms Buffer.toString utf16 string 76.54 ms/iter 13.1 (66.9 ms … 85.26 ms) 83.63 ms 85.26 ms 85.26 ms Buffer.toString hex string 38.56 ms/iter 25.9 (33.83 ms … 46.56 ms) 45.33 ms 46.56 ms 46.56 ms ```

This way Buffer polyfill doesn't need to reimplement the methods for reading data from an ArrayBuffer backing store. Instead the DataView APIs are used that then internally call V8 implementations of these algorithms. This reduces the amount of code needed for the Buffer polyfill significantly. It also provides performance benefits especially for reading floating point numbers.

aapoalas · 2024-11-03T21:04:15Z

I have decided that this is not a good idea: Adding something like 80 bytes to an already fairly large static size of Buffer AKA Uint8Array plus the internal ArrayBuffer, for what is a fairly small performance upside (and even some downsides, especially in the more necessary parts).

The proper way to improve the performance of Buffer read and write APIs would be to implement them in Rust.

anonrig mentioned this pull request Feb 19, 2023

Buffer.(read|write)(U)Int(LE|BE)(8|16|32) nodejs/performance#2

Open

aapoalas force-pushed the exp/ext-node/buffer-optimisation branch from 0762e00 to 76bd16b Compare February 19, 2023 16:13

aapoalas marked this pull request as ready for review February 19, 2023 18:11

bartlomieju requested a review from littledivy February 20, 2023 13:04

bartlomieju mentioned this pull request Feb 20, 2023

Rewrite some of JavaScript API in "ext/node" to Rust #17809

Closed

6 tasks

aapoalas force-pushed the exp/ext-node/buffer-optimisation branch from 13be2d7 to e74c58a Compare March 19, 2023 19:53

aapoalas changed the title ~~Experiment: Try optimise Node Buffer polyfill~~ refactor(ext/node): Try optimise Node Buffer polyfill Mar 19, 2023

aapoalas force-pushed the exp/ext-node/buffer-optimisation branch from ca6ebc1 to 53d1be6 Compare July 30, 2023 13:17

aapoalas force-pushed the exp/ext-node/buffer-optimisation branch from efd0922 to 29f34a8 Compare August 8, 2023 18:18

aapoalas mentioned this pull request Aug 14, 2023

perf(ext/node): Optimise Buffer string operations #20158

Merged

aapoalas force-pushed the exp/ext-node/buffer-optimisation branch from cada329 to 030fe87 Compare October 28, 2023 15:42

Optimise creation of empty Buffer

fcf06f3

aapoalas changed the title ~~refactor(ext/node): Try optimise Node Buffer polyfill~~ refactor(ext/node): Use an internal DataView in Buffer Oct 28, 2023

aapoalas closed this Nov 3, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

refactor(ext/node): Use an internal DataView in Buffer #17815

refactor(ext/node): Use an internal DataView in Buffer #17815

aapoalas commented Feb 18, 2023 •

edited

Loading

bartlomieju commented Feb 20, 2023

bartlomieju commented Feb 20, 2023

aapoalas commented Nov 3, 2024

refactor(ext/node): Use an internal DataView in Buffer #17815

refactor(ext/node): Use an internal DataView in Buffer #17815

Conversation

aapoalas commented Feb 18, 2023 • edited Loading

Measurements

Creation

Reading integers

Reading floating point numbers

Reading BigInts

Writing values

V8 bugs spawned from this work

bartlomieju commented Feb 20, 2023

bartlomieju commented Feb 20, 2023

aapoalas commented Nov 3, 2024

aapoalas commented Feb 18, 2023 •

edited

Loading