Performance of 4.0.0-alpha versions

[bogovicj]

This issue tracks benchmark runs for in-development 4.0.0-alpha versions.

[bogovicj]

Using N5ReadBenchmark:

few iterations

When running few iterations (8+8), master performs similarly to 3.5.1

3.5.1: 1.636 ± 0.059 ms/op

commit

Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Xmx4g -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 8 iterations, 500 ms each
# Measurement: 8 iterations, 500 ms each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench

# Run progress: 0.00% complete, ETA 00:00:08
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: basePath = /tmp/n5-test-6442921815361840385
2.256 ms/op
# Warmup Iteration   2: 2.015 ms/op
# Warmup Iteration   3: 2.150 ms/op
# Warmup Iteration   4: 1.802 ms/op
# Warmup Iteration   5: 2.042 ms/op
# Warmup Iteration   6: 1.702 ms/op
# Warmup Iteration   7: 1.624 ms/op
# Warmup Iteration   8: 1.691 ms/op
Iteration   1: 1.621 ms/op
Iteration   2: 1.692 ms/op
Iteration   3: 1.651 ms/op
Iteration   4: 1.642 ms/op
Iteration   5: 1.594 ms/op
Iteration   6: 1.649 ms/op
Iteration   7: 1.602 ms/op
Iteration   8: 1.636 ms/op


Result "org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench":
  1.636 ±(99.9%) 0.059 ms/op [Average]
  (min, avg, max) = (1.594, 1.636, 1.692), stdev = 0.031
  CI (99.9%): [1.577, 1.695] (assumes normal distribution)


# Run complete. Total time: 00:00:08

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt    8  1.636 ± 0.059  ms/op

master: 1.664 ± 0.161 ms/op

commit

Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Xmx4g -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 8 iterations, 500 ms each
# Measurement: 8 iterations, 500 ms each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench

# Run progress: 0.00% complete, ETA 00:00:08
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: basePath = /tmp/n5-test-4737034783181087096
2.896 ms/op
# Warmup Iteration   2: 2.445 ms/op
# Warmup Iteration   3: 1.953 ms/op
# Warmup Iteration   4: 1.672 ms/op
# Warmup Iteration   5: 1.842 ms/op
# Warmup Iteration   6: 1.754 ms/op
# Warmup Iteration   7: 1.676 ms/op
# Warmup Iteration   8: 1.592 ms/op
Iteration   1: 1.643 ms/op
Iteration   2: 1.575 ms/op
Iteration   3: 1.676 ms/op
Iteration   4: 1.618 ms/op
Iteration   5: 1.839 ms/op
Iteration   6: 1.580 ms/op
Iteration   7: 1.693 ms/op
Iteration   8: 1.687 ms/op


Result "org.janelia.saalfFor a small number of iterations, master performs similarly to 3.5.1eldlab.n5.N5ReadBenchmark.bench":
  1.664 ±(99.9%) 0.161 ms/op [Average]
  (min, avg, max) = (1.575, 1.664, 1.839), stdev = 0.084
  CI (99.9%): [1.503, 1.825] (assumes normal distribution)


# Run complete. Total time: 00:00:08

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt    8  1.664 ± 0.161  ms/op

many iterations

When running many iterations (10+40), 3.5.1 and master perform similarly for the first ~10 iterations,
after which master becomes much slower.

3.5.1: 1.649 ± 0.022 ms/op

commit

Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Xmx4g -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 10 iterations, 500 ms each
# Measurement: 40 iterations, 500 ms each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench

# Run progress: 0.00% complete, ETA 00:00:25
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: basePath = /tmp/n5-test-3311287752458391042
2.348 ms/op
# Warmup Iteration   2: 1.912 ms/op
# Warmup Iteration   3: 2.086 ms/op
# Warmup Iteration   4: 1.793 ms/op
# Warmup Iteration   5: 1.969 ms/op
# Warmup Iteration   6: 1.828 ms/op
# Warmup Iteration   7: 1.642 ms/op
# Warmup Iteration   8: 1.640 ms/op
# Warmup Iteration   9: 1.634 ms/op
# Warmup Iteration  10: 1.661 ms/op
Iteration   1: 1.674 ms/op
Iteration   2: 1.617 ms/op
Iteration   3: 1.655 ms/op
Iteration   4: 1.619 ms/op
Iteration   5: 1.653 ms/op
Iteration   6: 1.608 ms/op
Iteration   7: 1.688 ms/op
Iteration   8: 1.699 ms/op
Iteration   9: 1.706 ms/op
Iteration  10: 1.710 ms/op
Iteration  11: 1.651 ms/op
Iteration  12: 1.636 ms/op
Iteration  13: 1.646 ms/op
Iteration  14: 1.629 ms/op
Iteration  15: 1.624 ms/op
Iteration  16: 1.631 ms/op
Iteration  17: 1.603 ms/op
Iteration  18: 1.642 ms/op
Iteration  19: 1.628 ms/op
Iteration  20: 1.653 ms/op
Iteration  21: 1.636 ms/op
Iteration  22: 1.652 ms/op
Iteration  23: 1.627 ms/op
Iteration  24: 1.645 ms/op
Iteration  25: 1.612 ms/op
Iteration  26: 1.672 ms/op
Iteration  27: 1.610 ms/op
Iteration  28: 1.633 ms/op
Iteration  29: 1.591 ms/op
Iteration  30: 1.645 ms/op
Iteration  31: 1.623 ms/op
Iteration  32: 1.631 ms/op
Iteration  33: 1.618 ms/op
Iteration  34: 1.671 ms/op
Iteration  35: 1.643 ms/op
Iteration  36: 1.644 ms/op
Iteration  37: 1.609 ms/op
Iteration  38: 1.665 ms/op
Iteration  39: 1.786 ms/op
Iteration  40: 1.759 ms/op


Result "org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench":
  1.649 ±(99.9%) 0.022 ms/op [Average]
  (min, avg, max) = (1.591, 1.649, 1.786), stdev = 0.040
  CI (99.9%): [1.626, 1.671] (assumes normal distribution)


# Run complete. Total time: 00:00:25

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  1.649 ± 0.022  ms/op

master: 3.662 ± 0.648 ms/op

commit

Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Xmx4g -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 10 iterations, 500 ms each
# Measurement: 40 iterations, 500 ms each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench

# Run progress: 0.00% complete, ETA 00:00:25
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: basePath = /tmp/n5-test-2291493108050986111
2.755 ms/op
# Warmup Iteration   2: 2.320 ms/op
# Warmup Iteration   3: 1.941 ms/op
# Warmup Iteration   4: 1.641 ms/op
# Warmup Iteration   5: 1.783 ms/op
# Warmup Iteration   6: 1.867 ms/op
# Warmup Iteration   7: 1.676 ms/op
# Warmup Iteration   8: 1.647 ms/op
# Warmup Iteration   9: 1.585 ms/op
# Warmup Iteration  10: 1.668 ms/op
Iteration   1: 1.591 ms/op
Iteration   2: 1.622 ms/op
Iteration   3: 1.595 ms/op
Iteration   4: 1.652 ms/op
Iteration   5: 1.584 ms/op
Iteration   6: 1.733 ms/op
Iteration   7: 1.754 ms/op
Iteration   8: 1.718 ms/op
Iteration   9: 1.662 ms/op
Iteration  10: 2.110 ms/op
Iteration  11: 4.309 ms/op
Iteration  12: 4.345 ms/op
Iteration  13: 4.285 ms/op
Iteration  14: 4.246 ms/op
Iteration  15: 4.252 ms/op
Iteration  16: 4.275 ms/op
Iteration  17: 4.204 ms/op
Iteration  18: 4.319 ms/op
Iteration  19: 4.216 ms/op
Iteration  20: 4.364 ms/op
Iteration  21: 4.271 ms/op
Iteration  22: 4.302 ms/op
Iteration  23: 4.210 ms/op
Iteration  24: 4.336 ms/op
Iteration  25: 4.386 ms/op
Iteration  26: 4.338 ms/op
Iteration  27: 4.250 ms/op
Iteration  28: 4.223 ms/op
Iteration  29: 4.266 ms/op
Iteration  30: 4.293 ms/op
Iteration  31: 4.248 ms/op
Iteration  32: 4.309 ms/op
Iteration  33: 4.233 ms/op
Iteration  34: 4.344 ms/op
Iteration  35: 4.306 ms/op
Iteration  36: 4.374 ms/op
Iteration  37: 4.779 ms/op
Iteration  38: 4.461 ms/op
Iteration  39: 4.307 ms/op
Iteration  40: 4.388 ms/op


Result "org.janelia.saalfeldlab.n5.N5ReadBenchmark.bench":
  3.662 ±(99.9%) 0.648 ms/op [Average]
  (min, avg, max) = (1.584, 3.662, 4.779), stdev = 1.152
  CI (99.9%): [3.014, 4.309] (assumes normal distribution)


# Run complete. Total time: 00:00:25

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  3.662 ± 0.648  ms/op

[bogovicj]

Using N5BlockWriteBenchmarks:

N5BlockWriteBenchmarks.readBenchmark 64 int32 5 3 avgt 50 7914.344 ± 814.143 us/op
N5BlockWriteBenchmarks.writeBenchmark 64 int32 5 3 avgt 50 119873.194 ± 2568.508 us/op

3.5.1:
read 7914.344 ± 814.143
write 119873.194 ± 2568.508

c489d16

Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 5 iterations, 100 us each
# Measurement: 50 iterations, 100 us each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronizPicked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 5 iterations, 100 us each
# Measurement: 50 iterations, 100 us each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.readBenchmark
# Parameters: (blockDim = 64, dataType = int32, numBlocks = 5, numDimensions = 3)

# Run progress: 0.00% complete, ETA 00:00:00
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: 28550.743 us/op
# Warmup Iteration   2: 11664.386 us/op
# Warmup Iteration   3: 11687.640 us/op
# Warmup Iteration   4: 11237.961 us/op
# Warmup Iteration   5: 12630.948 us/op
Iteration   1: 10017.150 us/op
Iteration   2: 10722.065 us/op
Iteration   3: 14334.821 us/op
Iteration   4: 10627.114 us/op
Iteration   5: 8767.769 us/op
Iteration   6: 8230.047 us/op
Iteration   7: 8496.656 us/op
Iteration   8: 8294.610 us/op
Iteration   9: 8457.519 us/op
Iteration  10: 7832.730 us/op
Iteration  11: 8165.068 us/op
Iteration  12: 7763.605 us/op
Iteration  13: 8198.116 us/op
Iteration  14: 7841.566 us/op
Iteration  15: 10926.036 us/op
Iteration  16: 8566.758 us/op
Iteration  17: 8018.546 us/op
Iteration  18: 8233.195 us/op
Iteration  19: 8234.010 us/op
Iteration  20: 8096.515 us/op
Iteration  21: 8759.201 us/op
Iteration  22: 7292.076 us/op
Iteration  23: 7634.624 us/op
Iteration  24: 7481.190 us/op
Iteration  25: 6291.297 us/op
Iteration  26: 6408.259 us/op
Iteration  27: 10715.391 us/op
Iteration  28: 6460.904 us/op
Iteration  29: 6497.484 us/op
Iteration  30: 6494.460 us/op
Iteration  31: 5827.277 us/op
Iteration  32: 8203.576 us/op
Iteration  33: 6369.055 us/op
Iteration  34: 6322.379 us/op
Iteration  35: 6387.056 us/op
Iteration  36: 6739.568 us/op
Iteration  37: 6918.264 us/op
Iteration  38: 6607.027 us/op
Iteration  39: 10890.462 us/op
Iteration  40: 8073.648 us/op
Iteration  41: 6381.008 us/op
Iteration  42: 6376.766 us/op
Iteration  43: 6455.863 us/op
Iteration  44: 6251.868 us/op
Iteration  45: 8580.951 us/op
Iteration  46: 7615.118 us/op
Iteration  47: 8004.969 us/op
Iteration  48: 6589.877 us/op
Iteration  49: 6718.254 us/op
Iteration  50: 6545.450 us/op


Result "org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.readBenchmark":
  7914.344 ±(99.9%) 814.143 us/op [Average]
  (min, avg, max) = (5827.277, 7914.344, 14334.821), stdev = 1644.609
  CI (99.9%): [7100.201, 8728.487] (assumes normal distribution)


# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 5 iterations, 100 us each
# Measurement: 50 iterations, 100 us each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.writeBenchmark
# Parameters: (blockDim = 64, dataType = int32, numBlocks = 5, numDimensions = 3)

# Run progress: 50.00% complete, ETA 00:00:00
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: 119366.337 us/op
# Warmup Iteration   2: 117822.970 us/op
# Warmup Iteration   3: 118025.981 us/op
# Warmup Iteration   4: 116887.904 us/op
# Warmup Iteration   5: 115171.634 us/op
Iteration   1: 116216.068 us/op
Iteration   2: 119787.264 us/op
Iteration   3: 120589.805 us/op
Iteration   4: 123721.815 us/op
Iteration   5: 121177.101 us/op
Iteration   6: 119928.284 us/op
Iteration   7: 119760.192 us/op
Iteration   8: 119669.134 us/op
Iteration   9: 120834.378 us/op
Iteration  10: 116947.111 us/op
Iteration  11: 114893.216 us/op
Iteration  12: 121920.507 us/op
Iteration  13: 114944.982 us/op
Iteration  14: 116023.019 us/op
Iteration  15: 116602.438 us/op
Iteration  16: 116049.354 us/op
Iteration  17: 116583.192 us/op
Iteration  18: 115863.753 us/op
Iteration  19: 116371.015 us/op
Iteration  20: 111100.083 us/op
Iteration  21: 116146.972 us/op
Iteration  22: 115845.897 us/op
Iteration  23: 116224.137 us/op
Iteration  24: 112285.514 us/op
Iteration  25: 115696.805 us/op
Iteration  26: 118143.386 us/op
Iteration  27: 129333.074 us/op
Iteration  28: 128275.055 us/op
Iteration  29: 131589.722 us/op
Iteration  30: 133210.604 us/op
Iteration  31: 134228.844 us/op
Iteration  32: 123076.994 us/op
Iteration  33: 120230.026 us/op
Iteration  34: 118423.199 us/op
Iteration  35: 121634.797 us/op
Iteration  36: 116859.867 us/op
Iteration  37: 115993.893 us/op
Iteration  38: 115958.798 us/op
Iteration  39: 115978.166 us/op
Iteration  40: 124803.153 us/op
Iteration  41: 132794.988 us/op
Iteration  42: 119339.946 us/op
Iteration  43: 117999.797 us/op
Iteration  44: 120948.696 us/op
Iteration  45: 119910.770 us/op
Iteration  46: 119582.842 us/op
Iteration  47: 119711.508 us/op7914.344 ±  814.143
Iteration  48: 120014.651 us/op
Iteration  49: 120348.524 us/op
Iteration  50: 120086.350 us/op


Result "org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.writeBenchmark":
  119873.194 ±(99.9%) 2568.508 us/op [Average]
  (min, avg, max) = (111100.083, 119873.194, 134228.844), stdev = 5188.513
  CI (99.9%): [117304.686, 122441.702] (assumes normal distribution)


# Run complete. Total time: 00:00:07

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark                              (blockDim)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt       Score      Error  Units
N5BlockWriteBenchmarks.readBenchmark           64       int32            5                3  avgt   50    7914.344 ±  814.143  us/op
N5BlockWriteBenchmarks.writeBenchmark          64       int32            5                3  avgt   50  119873.194 ± 2568.508  us/op
e iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.readBenchmark
# Parameters: (blockDim = 64, dataType = int32, numBlocks = 5, numDimensions = 3)

# Run progress: 0.00% complete, ETA 00:00:00
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: 28550.743 us/op
# Warmup Iteration   2: 11664.386 us/op
# Warmup Iteration   3: 11687.640 us/op
# Warmup Iteration   4: 11237.961 us/op
# Warmup Iteration   5: 12630.948 us/op
Iteration   1: 10017.150 us/op
Iteration   2: 10722.065 us/op
Iteration   3: 14334.821 us/op
Iteration   4: 10627.114 us/op
Iteration   5: 8767.769 us/op
Iteration   6: 8230.047 us/op
Iteration   7: 8496.656 us/op
Iteration   8: 8294.610 us/op
Iteration   9: 8457.519 us/op
Iteration  10: 7832.730 us/op
Iteration  11: 8165.068 us/op
Iteration  12: 7763.605 us/op
Iteration  13: 8198.116 us/op
Iteration  14: 7841.566 us/op
Iteration  15: 10926.036 us/op
Iteration  16: 8566.758 us/op
Iteration  17: 8018.546 us/op
Iteration  18: 8233.195 us/op
Iteration  19: 8234.010 us/op
Iteration  20: 8096.515 us/op
Iteration  21: 8759.201 us/op
Iteration  22: 7292.076 us/op
Iteration  23: 7634.624 us/op
Iteration  24: 7481.190 us/op
Iteration  25: 6291.297 us/op
Iteration  26: 6408.259 us/op
Iteration  27: 10715.391 us/op
Iteration  28: 6460.904 us/op
Iteration  29: 6497.484 us/op
Iteration  30: 6494.460 us/op
Iteration  31: 5827.277 us/op
Iteration  32: 8203.576 us/op
Iteration  33: 6369.055 us/op
Iteration  34: 6322.379 us/op
Iteration  35: 6387.056 us/op
Iteration  36: 6739.568 us/op
Iteration  37: 6918.264 us/op
Iteration  38: 6607.027 us/op
Iteration  39: 10890.462 us/op
Iteration  40: 8073.648 us/op
Iteration  41: 6381.008 us/op
Iteration  42: 6376.766 us/op
Iteration  43: 6455.863 us/op
Iteration  44: 6251.868 us/op
Iteration  45: 8580.951 us/op
Iteration  46: 7615.118 us/op
Iteration  47: 8004.969 us/op
Iteration  48: 6589.877 us/op
Iteration  49: 6718.254 us/op
Iteration  50: 6545.450 us/op


Result "org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.readBenchmark":
  7914.344 ±(99.9%) 814.143 us/op [Average]
  (min, avg, max) = (5827.277, 7914.344, 14334.821), stdev = 1644.609
  CI (99.9%): [7100.201, 8728.487] (assumes normal distribution)


# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 5 iterations, 100 us each
# Measurement: 50 iterations, 100 us each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.writeBenchmark
# Parameters: (blockDim = 64, dataType = int32, numBlocks = 5, numDimensions = 3)

# Run progress: 50.00% complete, ETA 00:00:00
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: 119366.337 us/op
# Warmup Iteration   2: 117822.970 us/op
# Warmup Iteration   3: 118025.981 us/op
# Warmup Iteration   4: 116887.904 us/op
# Warmup Iteration   5: 115171.634 us/op
Iteration   1: 116216.068 us/op
Iteration   2: 119787.264 us/op
Iteration   3: 120589.805 us/op
Iteration   4: 123721.815 us/op
Iteration   5: 121177.101 us/op
Iteration   6: 119928.284 us/op
Iteration   7: 119760.192 us/op
Iteration   8: 119669.134 us/op
Iteration   9: 120834.378 us/op
Iteration  10: 116947.111 us/op
Iteration  11: 114893.216 us/op
Iteration  12: 121920.507 us/op
Iteration  13: 114944.982 us/op
Iteration  14: 116023.019 us/op
Iteration  15: 116602.438 us/op
Iteration  16: 116049.354 us/op
Iteration  17: 116583.192 us/op
Iteration  18: 115863.753 us/op
Iteration  19: 116371.015 us/op
Iteration  20: 111100.083 us/op
Iteration  21: 116146.972 us/op
Iteration  22: 115845.897 us/op
Iteration  23: 116224.137 us/op
Iteration  24: 112285.514 us/op
Iteration  25: 115696.805 us/op
Iteration  26: 118143.386 us/op
Iteration  27: 129333.074 us/op
Iteration  28: 128275.055 us/op
Iteration  29: 131589.722 us/op
Iteration  30: 133210.604 us/op
Iteration  31: 134228.844 us/op
Iteration  32: 123076.994 us/op
Iteration  33: 120230.026 us/op
Iteration  34: 118423.199 us/op
Iteration  35: 121634.797 us/op
Iteration  36: 116859.867 us/op
Iteration  37: 115993.893 us/op
Iteration  38: 115958.798 us/op
Iteration  39: 115978.166 us/op
Iteration  40: 124803.153 us/op
Iteration  41: 132794.988 us/op
Iteration  42: 119339.946 us/op
Iteration  43: 117999.797 us/op
Iteration  44: 120948.696 us/op
Iteration  45: 119910.770 us/op
Iteration  46: 119582.842 us/op
Iteration  47: 119711.508 us/op
Iteration  48: 120014.651 us/op
Iteration  49: 120348.524 us/op
Iteration  50: 120086.350 us/op


Result "org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.writeBenchmark":
  119873.194 ±(99.9%) 2568.508 us/op [Average]
  (min, avg, max) = (111100.083, 119873.194, 134228.844), stdev = 5188.513
  CI (99.9%): [117304.686, 122441.702] (assumes normal distribution)


# Run complete. Total time: 00:00:07

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark                              (blockDim)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt       Score      Error  Units
N5BlockWriteBenchmarks.readBenchmark           64       int32            5                3  avgt   50    7914.344 ±  814.143  us/op
N5BlockWriteBenchmarks.writeBenchmark          64       int32            5                3  avgt   50  119873.194 ± 2568.508  us/op

master:
read 7110.020 ± 902.128
write 121271.255 ± 1566.285

3955cdf

Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 5 iterations, 100 us each
# Measurement: 50 iterations, 100 us each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.readBenchmark
# Parameters: (blockDim = 64, dataType = int32, numBlocks = 5, numDimensions = 3)

# Run progress: 0.00% complete, ETA 00:00:00
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: [ECLIPSE HELPER] Indexing annotations...
26850.287 us/op
# Warmup Iteration   2: 12965.602 us/op
# Warmup Iteration   3: 12518.038 us/op
# Warmup Iteration   4: 17008.911 us/op
# Warmup Iteration   5: 8265.489 us/op
Iteration   1: 6643.999 us/op
Iteration   2: 6786.188 us/op
Iteration   3: 7511.223 us/op
Iteration   4: 7172.687 us/op
Iteration   5: 10245.697 us/op
Iteration   6: 7318.876 us/op
Iteration   7: 7341.323 us/op
Iteration   8: 6994.743 us/op
Iteration   9: 6733.962 us/op
Iteration  10: 6479.576 us/op
Iteration  11: 9644.747 us/op
Iteration  12: 6827.640 us/op
Iteration  13: 6264.978 us/op
Iteration  14: 6404.042 us/op
Iteration  15: 6053.296 us/op
Iteration  16: 6688.914 us/op
Iteration  17: 10370.109 us/op
Iteration  18: 6596.318 us/op
Iteration  19: 6598.664 us/op
Iteration  20: 6299.206 us/op
Iteration  21: 6864.555 us/op
Iteration  22: 6610.050 us/op
Iteration  23: 6331.227 us/op
Iteration  24: 8341.899 us/op
Iteration  25: 10207.501 us/op
Iteration  26: 11339.584 us/op
Iteration  27: 11486.667 us/op
Iteration  28: 10384.424 us/op
Iteration  29: 12973.865 us/op
Iteration  30: 5986.063 us/op
Iteration  31: 5745.543 us/op
Iteration  32: 5939.738 us/op
Iteration  33: 5604.034 us/op
Iteration  34: 5327.453 us/op
Iteration  35: 6258.831 us/op
Iteration  36: 5539.149 us/op
Iteration  37: 5647.355 us/op
Iteration  38: 5655.985 us/op
Iteration  39: 5331.257 us/op
Iteration  40: 5979.051 us/op
Iteration  41: 8386.762 us/op
Iteration  42: 5625.211 us/op
Iteration  43: 6247.875 us/op
Iteration  44: 5803.755 us/op
Iteration  45: 6052.463 us/op
Iteration  46: 6999.908 us/op
Iteration  47: 7131.391 us/op
Iteration  48: 6086.675 us/op
Iteration  49: 5465.469 us/op
Iteration  50: 5171.074 us/op


Result "org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.readBenchmark":
  7110.020 ±(99.9%) 902.128 us/op [Average]
  (min, avg, max) = (5171.074, 7110.020, 12973.865), stdev = 1822.343
  CI (99.9%): [6207.892, 8012.148] (assumes normal distribution)


# JMH version: 1.37
# VM version: JDK 1.8.0_452, OpenJDK 64-Bit Server VM, 25.452-b09
# VM invoker: /usr/lib/jvm/java-8-openjdk-amd64/jre/bin/java
# VM options: -Dfile.encoding=UTF-8 -Dstdout.encoding=UTF-8 -Dstderr.encoding=UTF-8 -Dsun.awt.disablegrab=true
# Blackhole mode: full + dont-inline hint (auto-detected, use -Djmh.blackhole.autoDetect=false to disable)
# Warmup: 5 iterations, 100 us each
# Measurement: 50 iterations, 100 us each
# Timeout: 10 min per iteration
# Threads: 1 thread, will synchronize iterations
# Benchmark mode: Average time, time/op
# Benchmark: org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.writeBenchmark
# Parameters: (blockDim = 64, dataType = int32, numBlocks = 5, numDimensions = 3)

# Run progress: 50.00% complete, ETA 00:00:00
# Fork: 1 of 1
Picked up _JAVA_OPTIONS: -Dsun.awt.disablegrab=true
# Warmup Iteration   1: 120216.516 us/op
# Warmup Iteration   2: 120882.323 us/op
# Warmup Iteration   3: 120181.092 us/op
# Warmup Iteration   4: 119338.797 us/op
# Warmup Iteration   5: 119548.607 us/op
Iteration   1: 122623.901 us/op
Iteration   2: 123741.415 us/op
Iteration   3: 123498.058 us/op
Iteration   4: 121795.297 us/op
Iteration   5: 121519.145 us/op
Iteration   6: 121611.835 us/op
Iteration   7: 122047.117 us/op
Iteration   8: 124125.736 us/op
Iteration   9: 121953.370 us/op
Iteration  10: 118494.684 us/op
Iteration  11: 126038.735 us/op
Iteration  12: 115523.571 us/op
Iteration  13: 117837.172 us/op
Iteration  14: 117676.918 us/op
Iteration  15: 118850.414 us/op
Iteration  16: 119120.472 us/op
Iteration  17: 119486.516 us/op
Iteration  18: 118790.739 us/op
Iteration  19: 118905.992 us/op
Iteration  20: 116115.143 us/op
Iteration  21: 116449.306 us/op
Iteration  22: 117838.652 us/op
Iteration  23: 129840.186 us/op
Iteration  24: 122090.555 us/op
Iteration  25: 122247.408 us/op
Iteration  26: 132514.196 us/op
Iteration  27: 125581.000 us/op
Iteration  28: 121714.437 us/op
Iteration  29: 123134.168 us/op
Iteration  30: 122908.932 us/op
Iteration  31: 121314.847 us/op
Iteration  32: 119919.673 us/op
Iteration  33: 119198.114 us/op
Iteration  34: 119320.789 us/op
Iteration  35: 122909.545 us/op
Iteration  36: 118906.984 us/op
Iteration  37: 118843.099 us/op
Iteration  38: 117988.288 us/op
Iteration  39: 120743.139 us/op
Iteration  40: 123442.415 us/op
Iteration  41: 121212.205 us/op
Iteration  42: 122386.138 us/op
Iteration  43: 122354.943 us/op
Iteration  44: 123052.000 us/op
Iteration  45: 123988.861 us/op
Iteration  46: 122935.223 us/op
Iteration  47: 122235.807 us/op
Iteration  48: 120460.078 us/op
Iteration  49: 119306.714 us/op
Iteration  50: 118968.819 us/op


Result "org.janelia.saalfeldlab.n5.benchmarks.N5BlockWriteBenchmarks.writeBenchmark":
  121271.255 ±(99.9%) 1566.285 us/op [Average]
  (min, avg, max) = (115523.571, 121271.255, 132514.196), stdev = 3163.974
  CI (99.9%): [119704.970, 122837.540] (assumes normal distribution)


# Run complete. Total time: 00:00:08

REMEMBER: The numbers below are just data. To gain reusable insights, you need to follow up on
why the numbers are the way they are. Use profilers (see -prof, -lprof), design factorial
experiments, perform baseline and negative tests that provide experimental control, make sure
the benchmarking environment is safe on JVM/OS/HW level, ask for reviews from the domain experts.
Do not assume the numbers tell you what you want them to tell.

Benchmark                              (blockDim)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt       Score      Error  Units
N5BlockWriteBenchmarks.readBenchmark           64       int32            5                3  avgt   50    7110.020 ±  902.128  us/op
N5BlockWriteBenchmarks.writeBenchmark          64       int32            5                3  avgt   50  121271.255 ± 1566.285  us/op

[tpietzsch]

Java version differences

I ran N5ReadBenchmark with (10+40) iterations with different JVM versions

When running many iterations (10+40), 3.5.1 and master perform similarly for the first ~10 iterations,
after which master becomes much slower.

This happened only with 1.8. Still no idea why.
Big picture is still that master is consistently slower than 3.5.1:

Java 8

# VM version: JDK 1.8.0_362, OpenJDK 64-Bit Server VM, 25.362-b09
(master)
Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  3.284 ± 0.331  ms/op

(3.5.1)
Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  0.951 ± 0.008  ms/op

Java 11

# VM version: JDK 11.0.15, OpenJDK 64-Bit Server VM, 11.0.15+10-LTS
(master)
Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  0,687 ± 0,060  ms/op

(3.5.1)
Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  0,435 ± 0,004  ms/op

Java 21

# VM version: JDK 21, OpenJDK 64-Bit Server VM, 21+35
(master)
Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  0,552 ± 0,004  ms/op

(3.5.1)
Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  0,322 ± 0,001  ms/op

[tpietzsch]

Possible causes

If we materialize an InputStream that we don’t know the length of, we use
org.apache.commons.io.IOUtils.toByteArray(...)

If we know the length, then we allocate an appropriately sized byte[] and use
new DataInputStream(is).readFully(…) to read from the stream.

n5/src/main/java/org/janelia/saalfeldlab/n5/readdata/AbstractInputStreamReadData.java

Lines 48 to 61 in 50ae586

	if (length >= 0) {
	data = new byte[length];
	try( InputStream is = inputStream()) {
	new DataInputStream(is).readFully(data);
	} catch (IOException e) {
	throw new N5IOException(e);
	}
	} else {
	try( InputStream is = inputStream()) {
	data = IOUtils.toByteArray(is);
	} catch (IOException e) {
	throw new N5IOException(e);
	}
	}

I changed this line:

n5/src/main/java/org/janelia/saalfeldlab/n5/codec/N5Codecs.java

Line 167 in 3955cdf

final ReadData decodeData = compression.decode(ReadData.from(in));

to final ReadData decodeData = compression.decode(ReadData.from(in, 8 * numElements)); for experimentation.
This only works because we know it is RawCompression and we know it is double[] in the N5ReadBenchmark.
But, just for the point of demonstration, it triggers that other branch (new DataInputStream(is).readFully(…)), and suddenly master is faster than 3.5.1 (java 21)

Benchmark              Mode  Cnt  Score   Error  Units
N5ReadBenchmark.bench  avgt   40  0,266 ± 0,007  ms/op

Things to investigate:

• Replace IOUtils.toByteArray(…) with something faster…
• Putting some effort into building ReadData that know their length (even when not "materialized") could be worth it too
• Also, try to postpone materializing as long as possible
• With sharding, the compressed lengths are known exactly through the index. We should take advantage of that.

[bogovicj]

I've made this script
https://gist.github.com/bogovicj/4a1c4c620aaf795718114eaaa49dfd9b

That runs BlockReadWriteBenchmarks.readBenchmark in n5-universe for a set of commits of this repo.

Early conclusion:
The optimizations in https://github.com/saalfeldlab/n5/tree/refactor/blockCodecsReadData provide meaningful improvements for zstd

On my machine using java 8:

Benchmark                               (blockDim)  (compressionType)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt     Score      Error  Units
BlockReadWriteBenchmarks.readBenchmark          64               zstd       int32            5                3  avgt  500  6333.230 ± 2169.187  us/op
b  3201e6a08bcf42f6d2d1f529c6e5dc949b251d05  materialize, avoid IOUtils
--
Benchmark                               (blockDim)  (compressionType)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt     Score      Error  Units
BlockReadWriteBenchmarks.readBenchmark          64               zstd       int32            5                3  avgt  500  7855.012 ± 2036.625  us/op
a  bfa73fb2962025b8d2f32573776e0ce2a9bfe1aa  materialize
--
Benchmark                               (blockDim)  (compressionType)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt     Score      Error  Units
BlockReadWriteBenchmarks.readBenchmark          64               zstd       int32            5                3  avgt  500  7799.787 ± 1979.056  us/op
master  0b6da3de9993e99c696633f8c4b64c898b54b184  master
--

on @cmhulbert 's mac using java 25.0.1

Benchmark                               (blockDim)  (compressionType)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt     Score     Error  Units
BlockReadWriteBenchmarks.readBenchmark          64               zstd       int32            5                3  avgt  500  3720.983 ± 566.958  us/op
b  3201e6a08bcf42f6d2d1f529c6e5dc949b251d05  materialize, avoid IOUtils
--
Benchmark                               (blockDim)  (compressionType)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt     Score     Error  Units
BlockReadWriteBenchmarks.readBenchmark          64               zstd       int32            5                3  avgt  500  3443.320 ± 214.785  us/op
a  bfa73fb2962025b8d2f32573776e0ce2a9bfe1aa  materialize
--
Benchmark                               (blockDim)  (compressionType)  (dataType)  (numBlocks)  (numDimensions)  Mode  Cnt     Score     Error  Units
BlockReadWriteBenchmarks.readBenchmark          64               zstd       int32            5                3  avgt  500  4167.976 ± 285.401  us/op
master  0b6da3de9993e99c696633f8c4b64c898b54b184  master

and another, related benchmark

BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000   631.498 ±  27.722  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000   707.485 ±  28.705  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1087.022 ±  37.387  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000   774.293 ± 114.296  us/op
b  3201e6a08bcf42f6d2d1f529c6e5dc949b251d05  materialize, avoid IOUtils
#####
--
BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000   606.606 ± 28.887  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000  1109.485 ± 49.658  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1079.877 ± 36.433  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000  1108.086 ± 87.585  us/op
a  bfa73fb2962025b8d2f32573776e0ce2a9bfe1aa  materialize
#####
--
BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000   950.597 ± 47.297  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000  1048.057 ± 49.625  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1330.334 ± 55.317  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000  1007.342 ± 82.946  us/op
master  0b6da3de9993e99c696633f8c4b64c898b54b184  master
#####

[bogovicj]

Running n5-benchmarks now also shows that (on my system with Java 8), the optimizations show improvements over both the latest alpha release, and the n5-3.5.1 release

#####
3.5.1    n5-3.5.1
BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000  1000.545 ±  25.443  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000  1112.481 ±  27.952  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1333.930 ±  64.724  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000  1191.348 ± 110.917  us/op

#####
b  3201e6a08bcf42f6d2d1f529c6e5dc949b251d05  materialize, avoid IOUtils
BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000   638.721 ±  27.695  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000   765.036 ±  27.884  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1185.277 ±  40.118  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000   852.571 ± 138.698  us/op

#####
a  bfa73fb2962025b8d2f32573776e0ce2a9bfe1aa  materialize
BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000   641.506 ± 28.654  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000  1153.628 ± 53.120  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1181.396 ± 38.522  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000  1237.269 ± 96.916  us/op

#####
master  0b6da3de9993e99c696633f8c4b64c898b54b184  master
BlockCodecBenchmarks.decodeBenchmark          64                raw       int32      random                3  avgt  1000   933.558 ± 46.553  us/op
BlockCodecBenchmarks.decodeBenchmark          64               gzip       int32      random                3  avgt  1000  1064.279 ± 44.638  us/op
BlockCodecBenchmarks.decodeBenchmark          64              blosc       int32      random                3  avgt  1000  1482.013 ± 58.523  us/op
BlockCodecBenchmarks.decodeBenchmark          64               zstd       int32      random                3  avgt  1000  1122.895 ± 90.548  us/op