tasty-bench

Featherlight benchmark framework (only one file!) for performance measurement with API mimicking criterion and gauge. A prominent feature is built-in comparison against previous runs and between benchmarks.

• How lightweight is it?
• How is it possible?
• How to switch?
• How to write a benchmark?
• How to read results?
• Wall-clock time vs. CPU time
• Statistical model
• Memory usage
• Combining tests and benchmarks
• Troubleshooting
• Isolating interfering benchmarks
• Comparison against baseline
• Comparison between benchmarks
• Plotting results
• Build flags
• Command-line options
• Custom command-line options

How lightweight is it?

There is only one source file Test.Tasty.Bench and no non-boot dependencies except tasty. So if you already depend on tasty for a test suite, there is nothing else to install.

Compare this to criterion (10+ modules, 50+ dependencies) and gauge (40+ modules, depends on basement and vector). A build on a clean machine is up to 16x faster than criterion and up to 4x faster than gauge. A build without dependencies is up to 6x faster than criterion and up to 8x faster than gauge.

tasty-bench is a native Haskell library and works everywhere, where GHC does. We support a full range of architectures (i386, amd64, armhf, arm64, ppc64le, s390x) and operating systems (Linux, Windows, macOS, FreeBSD, OpenBSD, NetBSD), plus any GHC from 7.0 to 9.6.

How is it possible?

Our benchmarks are literally regular tasty tests, so we can leverage all existing machinery for command-line options, resource management, structuring, listing and filtering benchmarks, running and reporting results. It also means that tasty-bench can be used in conjunction with other tasty ingredients.

Unlike criterion and gauge we use a very simple statistical model described below. This is arguably a questionable choice, but it works pretty well in practice. A rare developer is sufficiently well-versed in probability theory to make sense and use of all numbers generated by criterion.

How to switch?

Cabal mixins allow to taste tasty-bench instead of criterion or gauge without changing a single line of code:

cabal-version: 2.0

benchmark foo
  ...
  build-depends:
    tasty-bench
  mixins:
    tasty-bench (Test.Tasty.Bench as Criterion, Test.Tasty.Bench as Criterion.Main, Test.Tasty.Bench as Gauge, Test.Tasty.Bench as Gauge.Main)

This works vice versa as well: if you use tasty-bench, but at some point need a more comprehensive statistical analysis, it is easy to switch temporarily back to criterion.

How to write a benchmark?

Benchmarks are declared in a separate section of cabal file:

cabal-version:   2.0
name:            bench-fibo
version:         0.0
build-type:      Simple
synopsis:        Example of a benchmark

benchmark bench-fibo
  main-is:       BenchFibo.hs
  type:          exitcode-stdio-1.0
  build-depends: base, tasty-bench
  ghc-options:   "-with-rtsopts=-A32m"
  if impl(ghc >= 8.6)
    ghc-options: -fproc-alignment=64

And here is BenchFibo.hs:

import Test.Tasty.Bench

fibo :: Int -> Integer
fibo n = if n < 2 then toInteger n else fibo (n - 1) + fibo (n - 2)

main :: IO ()
main = defaultMain
  [ bgroup "fibonacci numbers"
    [ bench "fifth"     $ nf fibo  5
    , bench "tenth"     $ nf fibo 10
    , bench "twentieth" $ nf fibo 20
    ]
  ]

Since tasty-bench provides an API compatible with criterion, one can refer to its documentation for more examples.

How to read results?

Running the example above (cabal bench or stack bench) results in the following output:

All
  fibonacci numbers
    fifth:     OK (2.13s)
       63 ns ± 3.4 ns
    tenth:     OK (1.71s)
      809 ns ±  73 ns
    twentieth: OK (3.39s)
      104 μs ± 4.9 μs

All 3 tests passed (7.25s)

The output says that, for instance, the first benchmark was repeatedly executed for 2.13 seconds (wall-clock time), its predicted mean CPU time was 63 nanoseconds and means of individual samples do not often diverge from it further than ±3.4 nanoseconds (double standard deviation). Take standard deviation numbers with a grain of salt; there are lies, damned lies, and statistics.

Wall-clock time vs. CPU time

What time are we talking about? Both criterion and gauge by default report wall-clock time, which is affected by any other application which runs concurrently. Ideally benchmarks are executed on a dedicated server without any other load, but — let's face the truth — most of developers run benchmarks on a laptop with a hundred other services and a window manager, and watch videos while waiting for benchmarks to finish. That's the cause of a notorious "variance introduced by outliers: 88% (severely inflated)" warning.

To alleviate this issue tasty-bench measures CPU time by getCPUTime instead of wall-clock time by default. It does not provide a perfect isolation from other processes (e. g., if CPU cache is spoiled by others, populating data back from RAM is your burden), but is a bit more stable.

Caveat: this means that for multithreaded algorithms tasty-bench reports total elapsed CPU time across all cores, while criterion and gauge print maximum of core's wall-clock time. It also means that by default tasty-bench does not measure time spent out of process, e. g., calls to other executables. To work around this limitation use --time-mode command-line option or set it locally via TimeMode option.

Statistical model

Here is a procedure used by tasty-bench to measure execution time:

1. Set n ← 1.
2. Measure execution time tₙ of n iterations and execution time t₂ₙ of 2n iterations.
3. Find t which minimizes deviation of (nt, 2nt) from (tₙ, t₂ₙ), namely t ← (tₙ + 2t₂ₙ) / 5n.
4. If deviation is small enough (see --stdev below) or time is running out soon (see --timeout below), return t as a mean execution time.
5. Otherwise set n ← 2n and jump back to Step 2.

This is roughly similar to the linear regression approach which criterion takes, but we fit only two last points. This allows us to simplify away all heavy-weight statistical analysis. More importantly, earlier measurements, which are presumably shorter and noisier, do not affect overall result. This is in contrast to criterion, which fits all measurements and is biased to use more data points corresponding to shorter runs (it employs n ← 1.05n progression).

Mean time and its deviation does not say much about the distribution of individual timings. E. g., imagine a computation which (according to a coarse system timer) takes either 0 ms or 1 ms with equal probability. While one would be able to establish that its mean time is 0.5 ms with a very small deviation, this does not imply that individual measurements are anywhere near 0.5 ms. Even assuming an infinite precision of a system timer, the distribution of individual times is not known to be normal.

Obligatory disclaimer: statistics is a tricky matter, there is no one-size-fits-all approach. In the absence of a good theory simplistic approaches are as (un)sound as obscure ones. Those who seek statistical soundness should rather collect raw data and process it themselves using a proper statistical toolbox. Data reported by tasty-bench is only of indicative and comparative significance.

Memory usage

Configuring RTS to collect GC statistics (e. g., via cabal bench --benchmark-options '+RTS -T' or stack bench --ba '+RTS -T') enables tasty-bench to estimate and report memory usage:

All
  fibonacci numbers
    fifth:     OK (2.13s)
       63 ns ± 3.4 ns, 223 B  allocated,   0 B  copied, 2.0 MB peak memory
    tenth:     OK (1.71s)
      809 ns ±  73 ns, 2.3 KB allocated,   0 B  copied, 4.0 MB peak memory
    twentieth: OK (3.39s)
      104 μs ± 4.9 μs, 277 KB allocated,  59 B  copied, 5.0 MB peak memory

All 3 tests passed (7.25s)

This data is reported as per RTSStats fields: allocated_bytes, copied_bytes and max_mem_in_use_bytes.

Combining tests and benchmarks

When optimizing an existing function, it is important to check that its observable behavior remains unchanged. One can rebuild both tests and benchmarks after each change, but it would be more convenient to run sanity checks within benchmark itself. Since our benchmarks are compatible with tasty tests, we can easily do so.

Imagine you come up with a faster function myFibo to generate Fibonacci numbers:

import Test.Tasty.Bench
import Test.Tasty.QuickCheck -- from tasty-quickcheck package

fibo :: Int -> Integer
fibo n = if n < 2 then toInteger n else fibo (n - 1) + fibo (n - 2)

myFibo :: Int -> Integer
myFibo n = if n < 3 then toInteger n else myFibo (n - 1) + myFibo (n - 2)

main :: IO ()
main = Test.Tasty.Bench.defaultMain -- not Test.Tasty.defaultMain
  [ bench "fibo   20" $ nf fibo   20
  , bench "myFibo 20" $ nf myFibo 20
  , testProperty "myFibo = fibo" $ \n -> fibo n === myFibo n
  ]

This outputs:

All
  fibo   20:     OK (3.02s)
    104 μs ± 4.9 μs
  myFibo 20:     OK (1.99s)
     71 μs ± 5.3 μs
  myFibo = fibo: FAIL
    *** Failed! Falsified (after 5 tests and 1 shrink):
    2
    1 /= 2
    Use --quickcheck-replay=927711 to reproduce.

1 out of 3 tests failed (5.03s)

We see that myFibo is indeed significantly faster than fibo, but unfortunately does not do the same thing. One should probably look for another way to speed up generation of Fibonacci numbers.

Troubleshooting

• If benchmarks take too long, set --timeout to limit execution time of individual benchmarks, and tasty-bench will do its best to fit into a given time frame. Without --timeout we rerun benchmarks until achieving a target precision set by --stdev, which in a noisy environment of a modern laptop with GUI may take a lot of time.

  While criterion runs each benchmark at least for 5 seconds, tasty-bench is happy to conclude earlier, if it does not compromise the quality of results. In our experiments tasty-bench suites tend to finish earlier, even if some individual benchmarks take longer than with criterion.

  A common source of noisiness is garbage collection. Setting a larger allocation area (nursery) is often a good idea, either via cabal bench --benchmark-options '+RTS -A32m' or stack bench --ba '+RTS -A32m'. Alternatively bake it into cabal file as ghc-options: "-with-rtsopts=-A32m".

  For GHC ≥ 8.10 consider switching benchmarks to a non-moving garbage collector, because it decreases GC pauses and corresponding noise: +RTS --nonmoving-gc.

• Never compile benchmarks with -fstatic-argument-transformation, because it breaks a trick we use to force GHC into reevaluation of the same function application over and over again.

• If benchmark results look malformed like below, make sure that you are invoking Test.Tasty.Bench.defaultMain and not Test.Tasty.defaultMain (the difference is consoleBenchReporter vs. consoleTestReporter):

  All
    fibo 20:       OK (1.46s)
      Response {respEstimate = Estimate {estMean = Measurement {measTime = 87496728, measAllocs = 0, measCopied = 0}, estStdev = 694487}, respIfSlower = FailIfSlower Infinity, respIfFaster = FailIfFaster Infinity}
  

• If benchmarks fail with an error message

  Unhandled resource. Probably a bug in the runner you're using.
  

  or

  Unexpected state of the resource (NotCreated) in getResource. Report as a tasty bug.
  

  this is likely caused by env or envWithCleanup affecting benchmarks structure. You can use env to read test data from IO, but not to read benchmark names or affect their hierarchy in other way. This is a fundamental restriction of tasty to list and filter benchmarks without launching missiles.

• If benchmarks fail with Test dependencies form a loop or Test dependencies have cycles, this is likely because of bcompare, which compares a benchmark with itself. Locating a benchmark in a global environment may be tricky, please refer to tasty documentation for details and consider using locateBenchmark.

Isolating interfering benchmarks

One difficulty of benchmarking in Haskell is that it is hard to isolate benchmarks so that they do not interfere. Changing the order of benchmarks or skipping some of them has an effect on heap's layout and thus affects garbage collection. This issue is well attested in both criterion and gauge.

Usually (but not always) skipping some benchmarks speeds up remaining ones. That's because once a benchmark allocated heap which for some reason was not promptly released afterwards (e. g., it forced a top-level thunk in an underlying library), all further benchmarks are slowed down by garbage collector processing this additional amount of live data over and over again.

There are several mitigation strategies. First of all, giving garbage collector more breathing space by +RTS -A32m (or more) is often good enough.

Further, avoid using top-level bindings to store large test data. Once such thunks are forced, they remain allocated forever, which affects detrimentally subsequent unrelated benchmarks. Treat them as external data, supplied via env: instead of

largeData :: String
largeData = replicate 1000000 'a'

main :: IO ()
main = defaultMain
  [ bench "large" $ nf length largeData, ... ]

use

import Control.DeepSeq (force)
import Control.Exception (evaluate)

main :: IO ()
main = defaultMain
  [ env (evaluate (force (replicate 1000000 'a'))) $ \largeData ->
    bench "large" $ nf length largeData, ... ]

Finally, as an ultimate measure to reduce interference between benchmarks, one can run each of them in a separate process. We do not quite recommend this approach, but if you are desperate, here is how:

cabal run -v0 all:benches -- -l | sed -e 's/[\"]/\\\\\\&/g' | while read -r name; do cabal run -v0 all:benches -- -p '$0 == "'"$name"'"'; done

This assumes that there is a single benchmark suite in the project and that benchmark names do not contain newlines.

Comparison against baseline

One can compare benchmark results against an earlier run in an automatic way.

When using this feature, it's especially important to compile benchmarks with ghc-options: -fproc-alignment=64, otherwise results could be skewed by intermittent changes in cache-line alignment.

Firstly, run tasty-bench with --csv FILE key to dump results to FILE in CSV format (it could be a good idea to set smaller --stdev, if possible):

Name,Mean (ps),2*Stdev (ps)
All.fibonacci numbers.fifth,48453,4060
All.fibonacci numbers.tenth,637152,46744
All.fibonacci numbers.twentieth,81369531,3342646

Now modify implementation and rerun benchmarks with --baseline FILE key. This produces a report as follows:

All
  fibonacci numbers
    fifth:     OK (0.44s)
       53 ns ± 2.7 ns,  8% more than baseline
    tenth:     OK (0.33s)
      641 ns ±  59 ns,       same as baseline
    twentieth: OK (0.36s)
       77 μs ± 6.4 μs,  5% less than baseline

All 3 tests passed (1.50s)

You can also fail benchmarks, which deviate too far from baseline, using --fail-if-slower and --fail-if-faster options. For example, setting both of them to 6 will fail the first benchmark above (because it is more than 6% slower), but the last one still succeeds (even while it is measurably faster than baseline, deviation is less than 6%). Consider also using --hide-successes to show only problematic benchmarks, or even tasty-rerun package to focus on rerunning failing items only.

If you wish to compare two CSV reports non-interactively, here is a handy awk incantation:

awk 'BEGIN{FS=",";OFS=",";print "Name,Old,New,Ratio"}FNR==1{trueNF=NF;next}NF<trueNF{print "Benchmark names should not contain newlines";exit 1}FNR==NR{oldTime=$(NF-trueNF+2);NF-=trueNF-1;a[$0]=oldTime;next}{newTime=$(NF-trueNF+2);NF-=trueNF-1;print $0,a[$0],newTime,newTime/a[$0];gs+=log(newTime/a[$0]);gc++}END{if(gc>0)print "Geometric mean,,",exp(gs/gc)}' old.csv new.csv

A larger shell snippet to compare two git commits can be found in compare_benches.sh.

Note that columns in CSV report are different from what criterion or gauge would produce. If names do not contain commas, missing columns can be faked this way:

awk 'BEGIN{FS=",";OFS=",";print "Name,Mean,MeanLB,MeanUB,Stddev,StddevLB,StddevUB"}NR==1{trueNF=NF;next}NF<trueNF{print $0;next}{mean=$(NF-trueNF+2);stddev=$(NF-trueNF+3);NF-=trueNF-1;print $0,mean/1e12,mean/1e12,mean/1e12,stddev/2e12,stddev/2e12,stddev/2e12}'

To fake gauge in --csvraw mode use

awk 'BEGIN{FS=",";OFS=",";print "name,iters,time,cycles,cpuTime,utime,stime,maxrss,minflt,majflt,nvcsw,nivcsw,allocated,numGcs,bytesCopied,mutatorWallSeconds,mutatorCpuSeconds,gcWallSeconds,gcCpuSeconds"}NR==1{trueNF=NF;next}NF<trueNF{print $0;next}{mean=$(NF-trueNF+2);fourth=$(NF-trueNF+4);fifth=$(NF-trueNF+5);sixth=$(NF-trueNF+6);NF-=trueNF-1;print $0,1,mean/1e12,0,mean/1e12,mean/1e12,0,sixth+0,0,0,0,0,fourth+0,0,fifth+0,0,0,0,0}'

Comparison between benchmarks

You can also compare benchmarks to each other without any external tools, all in the comfort of your terminal.

import Test.Tasty.Bench

fibo :: Int -> Integer
fibo n = if n < 2 then toInteger n else fibo (n - 1) + fibo (n - 2)

main :: IO ()
main = defaultMain
  [ bgroup "fibonacci numbers"
    [ bcompare "tenth"  $ bench "fifth"     $ nf fibo  5
    ,                     bench "tenth"     $ nf fibo 10
    , bcompare "tenth"  $ bench "twentieth" $ nf fibo 20
    ]
  ]

This produces a report, comparing mean times of fifth and twentieth to tenth:

All
  fibonacci numbers
    fifth:     OK (16.56s)
      121 ns ± 2.6 ns, 0.08x
    tenth:     OK (6.84s)
      1.6 μs ±  31 ns
    twentieth: OK (6.96s)
      203 μs ± 4.1 μs, 128.36x

To locate a baseline benchmark in a larger suite use locateBenchmark.

One can leverage comparisons between benchmarks to implement portable performance tests, expressing properties like "this algorithm must be at least twice faster than that one" or "this operation should not be more than thrice slower than that". This can be achieved with bcompareWithin, which takes an acceptable interval of performance as an argument.

Plotting results

Users can dump results into CSV with --csv FILE and plot them using gnuplot or other software. But for convenience there is also a built-in quick-and-dirty SVG plotting feature, which can be invoked by passing --svg FILE. Here is a sample of its output:

Build flags

Build flags are a brittle subject and users do not normally need to touch them.

• If you find yourself in an environment, where tasty is not available and you have access to boot packages only, you can still use tasty-bench! Just copy Test/Tasty/Bench.hs to your project (imagine it like a header-only C library). It will provide you with functions to build Benchmarkable and run them manually via measureCpuTime. This mode of operation can be also configured by disabling Cabal flag tasty.

• If results are amiss or oscillate wildly and adjusting --timeout and --stdev does not help, you may be interested to investigate individual timings of successive runs by enabling Cabal flag debug. This will pipe raw data into stderr.

Command-line options

Use --help to list command-line options.

• -p, --pattern

  This is a standard tasty option, which allows filtering benchmarks by a pattern or awk expression. Please refer to tasty documentation for details.

• -t, --timeout

  This is a standard tasty option, setting timeout for individual benchmarks in seconds. Use it when benchmarks tend to take too long: tasty-bench will make an effort to report results (even if of subpar quality) before timeout. Setting timeout too tight (insufficient for at least three iterations) will result in a benchmark failure. One can adjust it locally for a group of benchmarks, e. g., localOption (mkTimeout 100000000) for 100 seconds.

• --stdev

  Target relative standard deviation of measurements in percents (5% by default). Large values correspond to fast and loose benchmarks, and small ones to long and precise. It can also be adjusted locally for a group of benchmarks, e. g., localOption (RelStDev 0.02). If benchmarking takes far too long, consider setting --timeout, which will interrupt benchmarks, potentially before reaching the target deviation.

• --csv

  File to write results in CSV format.

• --baseline

  File to read baseline results in CSV format (as produced by --csv).

• --fail-if-slower, --fail-if-faster

  Upper bounds of acceptable slow down / speed up in percents. If a benchmark is unacceptably slower / faster than baseline (see --baseline), it will be reported as failed. Can be used in conjunction with a standard tasty option --hide-successes to show only problematic benchmarks. Both options can be adjusted locally for a group of benchmarks, e. g., localOption (FailIfSlower 0.10).

• --svg

  File to plot results in SVG format.

• --time-mode

  Whether to measure CPU time (cpu, default) or wall-clock time (wall).

• +RTS -T

  Estimate and report memory usage.

Custom command-line options

As usual with tasty, it is easy to extend benchmarks with custom command-line options. Here is an example:

import Data.Proxy
import Test.Tasty.Bench
import Test.Tasty.Ingredients.Basic
import Test.Tasty.Options
import Test.Tasty.Runners

newtype RandomSeed = RandomSeed Int

instance IsOption RandomSeed where
  defaultValue = RandomSeed 42
  parseValue = fmap RandomSeed . safeRead
  optionName = pure "seed"
  optionHelp = pure "Random seed used in benchmarks"

main :: IO ()
main = do
  let customOpts  = [Option (Proxy :: Proxy RandomSeed)]
      ingredients = includingOptions customOpts : benchIngredients
  opts <- parseOptions ingredients benchmarks
  let RandomSeed seed = lookupOption opts
  defaultMainWithIngredients ingredients benchmarks

benchmarks :: Benchmark
benchmarks = bgroup "All" []