Home

Awesome

Streaming Benchmarks

Hackage Gitter chat Build Status Windows Build status

This package provides micro-benchmarks to measure and compare the performance of various streaming implementations in Haskell.

We have taken due to care to make sure that we are benchmarking correctly and fairly. See the notes on correct benchmarking.

DISCLAIMER: This package is a result of benchmarking effort done during the development of streamly by the authors of streamly.

Benchmarks

The benchmark names are obvious, some of them are described below. Single operation benchmarks:

NameDescription
drainJust discards all the elements in the stream
drop-alldrops all element using the drop operation
lastextract the last element of the stream
foldsum all the numbers in the stream
mapincrements each number in the stream by 1
take-allUse take to retain all the elements in the stream
filter-evenKeep even numbers, discard odd
scanscan the stream using + operation
mapMtransform the stream using a monadic action
zipcombines corresponding elements of the two streams together

Composite operation benchmarks:

NameDescription
map x 4perform map operation 4 times
take-maptake followed by a map

For more details on how each benchmark is implemented see this benchmark file.

Each benchmark is run in a separate process to avoid any effects of GC interference and sharing across benchmarks.

Benchmark Results

Below we present some results comparing streamly with other streaming implementations. Due care has been taken to keep the comparisons fair. We have optimized each library's code to the best of our knowledge, please point out if you find any measurement issues.

Reproducing benchmark results

Commands to reproduce the benchmark results are provided in each section below. But before you run those commands you need to build the reporting tool once using the following command. Note that this command works with only ghc-8.8.4 or lower. However, after building this tool you can run the benchmarks with any GHC version.

$ bin/bench.sh --with-compiler ghc-8.8.4 --no-measure

Nix users can use --use-nix option. It uses an older version of nixpkgs that contains the required dependencies. For example:

$ bin/bench.sh --use-nix --quick

Streamly vs Haskell Lists

Streamly, when used with Identity monad, is almost the same as Haskell lists (in the base package). See this for more details.

The following table compares the timing of several operations for streamly with lists using a one million element stream. For brevity only those operations where the performance of the two packages differ by more than 10% are shown in the table below. The last column shows how many times slower list is compared to streamly.

Benchmarkstreamly(μs)list(μs)list/streamly
drop-map x 4375.0976925.32205.08
filter-drop x 4382.0354848.54143.57
drop-scan x 4795.8176716.7996.40
filter-scan x 4795.6044559.1556.01
scan-map x 41192.1948838.2240.97
take-map x 41500.9960126.5840.06
filter-take x 41502.0148766.8732.47
take-drop x 41499.6241720.0327.82
take-scan x 41874.9451283.3027.35
drop-one x 4375.338993.8723.96
dropWhile-false x 4374.618957.7923.91
dropWhile-false374.838670.0523.13
drop-one390.778681.8522.22
dropWhile-true571.6012237.4821.41
drop-all562.948262.3814.68
take-all624.83564.341/1.11
scan x 4795.83385.851/2.06
appendR[10000]360.75126.951/2.84
concatMap34957.711124.851/31.08

To reproduce these results use the following commands:

$ bin/bench.sh --benchmarks "StreamlyPure List" --compare --diff-style absolute --diff-cutoff-percent 10 --quick
$ bin/bench.sh --benchmarks "StreamlyPure List" --compare --diff-style multiples --diff-cutoff-percent 10 --no-measure

Streamly vs Streaming

The following table compares the timing of several operations for streamly with streaming using a million element stream.

Benchmarkstreamly(μs)streaming(μs)streaming/streamly
appendR[10000]326.561301176.693984.54
mapM x 4374.42223591.08597.17
filter-map x 4381.07194903.88511.47
filter-scan x 4795.66233527.90293.50
filter-all-in x 4375.40102629.64273.38
filter-drop x 4387.1599096.98255.96
map x 4386.4994944.87245.66
drop-map x 4375.6289669.37238.73
scan x 4797.00166332.40208.70
scan-map x 41194.30238804.48199.95
filter-even x 4396.3777865.47196.45
drop-scan x 4796.98156063.52195.82
takeWhile-true x 4562.4990183.53160.33
scan375.2447520.57126.64
filter-take x 41498.55189635.34126.55
mapM388.1046689.61120.30
take-map x 41500.71178954.50119.25
zip656.6566689.73101.56
take-scan x 42380.35241675.75101.53
filter-all-in375.9733590.1489.34
map375.0233081.1388.21
filter-even393.2630458.4677.45
filter-all-out382.8726826.2170.07
take-all x 41499.71101332.5367.57
take-drop x 41498.5398281.9965.59
takeWhile-true562.6231863.2556.63
foldl'388.2218503.1547.66
drop-all562.0825200.3244.83
take-all768.6533247.9743.26
dropWhile-true564.8724431.5043.25
last385.5315240.8539.53
dropWhile-false374.8314566.7038.86
drop-one374.8014565.0138.86
drop-one x 4375.8814448.6738.44
dropWhile-false x 4390.1214619.4237.47
drain375.0613702.2936.53
toList117708.83201444.811.71

To reproduce these results use the following commands:

$ bin/bench.sh --benchmarks "Streamly Streaming" --compare --diff-style absolute --diff-cutoff-percent 10 --quick
$ bin/bench.sh --benchmarks "Streamly Streaming" --compare --diff-style multiples --diff-cutoff-percent 10 --no-measure

Streamly vs Pipes

The following table compares the timing of several operations for streamly with pipes using a million element stream.

Benchmarkstreamly(μs)pipes(μs)pipes/streamly
appendR[10000]327.90901135.922748.21
mapM x 4375.20407184.391085.23
filter-map x 4381.52366759.70961.31
drop-map x 4375.48281296.82749.16
filter-all-in x 4375.60222331.68591.93
filter-drop x 4387.44222830.71575.14
drop-scan x 4797.23336737.89422.39
filter-even x 4389.87152688.91391.64
filter-scan x 4797.38309733.91388.44
drop-one x 4375.48139851.13372.46
map x 4386.56136289.32352.57
dropWhile-false x 4390.72137395.44351.65
scan-map x 41194.38381286.88319.23
takeWhile-true x 4562.86165143.23293.40
scan x 4796.68222986.17279.90
mapM388.1995576.97246.21
filter-all-in375.2171297.42190.02
take-map x 41502.76275887.24183.59
scan374.8165549.13174.89
take-drop x 41503.43256448.45170.58
filter-even390.2966183.72169.57
filter-all-out376.9959074.54156.70
drop-one375.1958395.24155.64
dropWhile-false375.3558223.03155.12
map375.0557736.43153.94
filter-take x 41503.00227925.71151.65
take-scan x 42455.91354284.33144.26
zip657.0786011.93130.90
takeWhile-true564.1461390.21108.82
take-all x 41502.32139730.7093.01
dropWhile-true564.0349227.1987.28
drop-all562.0546505.3782.74
take-all824.0960511.3473.43
drain375.2926390.5970.32
foldl'397.3419064.0547.98
last387.1117364.4444.86
toList117257.09207405.941.77

To reproduce these results use the following commands:

$ bin/bench.sh --benchmarks "Streamly Pipes" --compare --diff-style absolute --diff-cutoff-percent 10 --quick
$ bin/bench.sh --benchmarks "Streamly Pipes" --compare --diff-style multiples --diff-cutoff-percent 10 --no-measure

Streamly vs Conduit

The following table compares the timing of several operations for streamly with conduit using a million element stream.

Benchmarkstreamly(μs)conduit(μs)conduit/streamly
mapM x 4375.46297002.31791.04
filter-map x 4380.79267543.81702.60
drop-map x 4375.66232307.84618.39
filter-drop x 4386.05235029.15608.81
filter-scan x 4796.56306556.67384.85
drop-scan x 4797.19300789.06377.31
zip657.29210069.05319.60
filter-all-in x 4375.24118506.68315.82
scan-map x 41194.67360671.18301.90
map x 4387.00113497.14293.27
drop-one x 4375.49101842.95271.23
dropWhile-false x 4389.44102051.22262.04
scan x 4796.72190479.35239.08
takeWhile-true x 4564.58114459.57202.73
filter-even x 4391.7672369.30184.73
filter-take x 41502.04267921.27178.37
take-map x 41502.88238875.95158.95
take-drop x 41500.34232606.19155.04
take-scan x 42443.83309738.86126.74
mapM389.1541897.48107.66
scan375.4038137.85101.59
take-all x 41502.32110682.7473.67
filter-all-in375.3126024.2169.34
dropWhile-false375.1025307.1367.47
map375.1823088.0961.54
drop-one375.4322020.6558.65
filter-even392.2821504.2854.82
takeWhile-true562.7929012.6851.55
filter-all-out378.7615736.0541.55
drop-all562.8919916.4835.38
foldl'388.8812499.0332.14
dropWhile-true564.4317983.3531.86
take-all784.6724425.3631.13
last385.7510974.8428.45
drain375.184272.1511.39
appendR[10000]326.931207.883.69
toList116441.26199138.091.71

To reproduce these results use the following commands:

$ bin/bench.sh --benchmarks "Streamly Conduit" --compare --diff-style absolute --diff-cutoff-percent 10 --quick
$ bin/bench.sh --benchmarks "Streamly Conduit" --compare --diff-style multiples --diff-cutoff-percent 10 --no-measure

Stack and heap utilization

To report heap utilization by individual benchmarks you can include maxrss in the --fields option.

To know about stack and heap utilization by the libraries you can also take a look at the RTS heap and stack limits used to run the benchmarks of various libraries in bench-config.sh.

Comparing other libraries

This package supports many streaming libraries. Use the following command to see all available benchmarks:

$ ./bench.sh --help

You can then select the libraries you want to compare:

$ ./bench.sh --benchmarks "streaming,pipes" --measure

Adding New Libraries

It is trivial to add a new package. This is how a benchmark file for a streaming package looks like. Pull requests are welcome, we will be happy to help, just join the gitter chat and ask!