bent

bent

Bent automates downloading, compiling, and running Go tests and benchmarks from various Github repositories. The benchmarks and tests retrieved are pinned at a particular version, which avoids that source of variation and also avoids the problem of running unknown code.

Older versions of bent defaulted to running benchmarks in a container, however that is not the default now and it is no longer well tested.

Installation:

go install golang.org/x/benchmarks/cmd/bent@latest

Depends on burntsushi/toml.

Install rsync for slightly improved copy performance.

Initial usage :

go install golang.org/x/benchmarks/cmd/bent@latest
mkdir scratch
cd scratch
bent -I
cp configurations-sample.toml configurations.toml
nano configurations.toml # or use your favorite editor
bent -v -N 10 # will run default set of ~50 benchmarks 10 times using supplied configuration(s)

The output binaries are placed in subdirectory testbin, various benchmark results (from building, run, and others requested) are placed in subdirectory bench, and the binaries are also incorporated into Docker containers if Docker is used. Each benchmark and configuration has a shortname, and the generated binaries combine these shortnames, for example gonum_mat_Tip and gonum_mat_Go1.9. Benchmark files are prefixed with a run timestamp, and grouped by configuration, with various suffixes for the various benchmarks. Run benchmarks appears in files with suffix .stdout. Others are more obviously named, with suffixes .build, .benchsize, and .benchdwarf.

Flags for your use:

Suite, Benchmark and Configuration files

Suites, benchmarks and configurations appear in toml format, since that is somewhat more human-friendly than JSON and in particular allows comments. Values in these files can refer to environment variables; PATH, USER, HOME, SHELL, BENT*, and GO*. Unless it is explicitly supplied, ROOT is defined to be $PWD. For iterated builds and benchmarks BENT_I is set to the iteration count, and for runs BENT_BINARY is the file name (excluding path).

The Suite file defines the shortnames for benchmarks, where they are obtained, and what version, plus any always-necessary flags:

[[Suites]]
  Name = "gonum_mat"
  Repo = "gonum.org/v1/gonum/mat"
  BuildFlags = ["-tags", "safe"]
  Version = "@v0.9.3"

 [[Suites]]
  Name = "toml"
  Repo = "github.com/BurntSushi/toml"
  Version = "@v1.3.2"
  ExtraFiles = ["_example"]

Here, gonum_mat is checked out at version 0.9.3 and is always build with -tags safe. And to run the toml benchmarks, the contents of the _example directory are also required.

A sample benchmark entry:

[[Benchmarks]]
  Name = "gonum_topo"
  Tests = "Test"
  Benchmarks = "Benchmark(TarjanSCCGnp_1000_half|TarjanSCCGnp_10_tenth)"
  # NotSandboxed = true # uncomment if cannot be run in a Docker container
  # Disabled = true # uncomment to disable benchmark

Here, Name is a short name, and Tests and Benchmarks are the regular expressions for go test specifying which tests or benchmarks to run.

A sample configuration entry with all the options supplied:

[[Configurations]]
  Name = "Go-preempt"
  Root = "$HOME/work/go/"
 # Optional flags below
  BuildFlags = ["-gccgoflags=all=-O3 -static-libgo","-tags=noasm"] # for Gollvm
  AfterBuild = ["benchsize", "benchdwarf"]
  GcFlags = "-d=ssa/insert_resched_checks/on"
  LdFlags = "-randlayout=$BENT_I"
  GcEnv = ["GOMAXPROCS=1","GOGC=200"]
  RunFlags = ["-test.short"]
  RunEnv = ["GOGC=1000"]
  RunWrapper = ["cpuprofile"]
  Disabled = false

The Gc... attributes apply to the test or benchmark compilation, the Run... attributes apply to the test or benchmark run. A RunWrapper command receives the entire command line as arguments, plus the environment variable BENT_BINARY set to the filename (excluding path) of the binary being run (for example, "uuid_Tip") and BENT_I set to the run number for this binary. One useful example is cpuprofile:

#!/bin/bash
# Run args as command, but run cpuprofile and then pprof to capture test cpuprofile output
if [ -n ${BENT_PGO} ] ; then
    pf="${BENT_PGO}/${BENT_BENCH}.prof"
    mkdir -p ${BENT_PGO}
else
    pf="${BENT_PROFILES}/${BENT_BINARY}_${BENT_I}.prof"
    mkdir -p ${BENT_PROFILES}
fi
"$@" -test.cpuprofile="$pf"
echo cpuprofile in "$pf"
if [[ x`which pprof` == x"" ]] ; then
    go tool pprof -text -flat -nodecount=20 "$pf"
else
    pprof -text -flat -nodecount=20 "$pf"
fi

When both configuration and benchmark wrappers are used the configuration wrapper runs the benchmark wrapper runs the actual benchmark, i.e.

ConfigWrapper ConfigArg BenchWrapper BenchArg ActualBenchmark

The Disabled attribute for both benchmarks and configurations removes them from normal use, but leaves them accessible to explicit request with -b or -c.

Special configurations

Bent includes sample configurations to support PGO-optimized benchmarks and randomized link order to normalize away branch alignment artifacts. These may need editing to reference local paths before use.

For PGO optimization:

# First generate profiles
bent -C configurations-pgo.toml -c pgo-generate
# Use profiles for 10 benchmark runs
bent -C configurations-pgo.toml -N 10

For randomized link order:

# Randomized benchmarking -R is incompatible with -a or -N
bent -C configurations-random.toml -R 10

Scripts

Bent now comes with several shell scripts to automate common uses. These all run using perflock if it is available, and default to different numbers of builds (usually 1) and benchmark runs (usually 15) which can be overridden at invocation.

cmpcl.sh refs/changes/<nn>/<cl>/<patch> [options]

This checks out a particular version of a CL, and its immediate predecessor, and benchmarks the change. The refs/changes/<nn>/<cl>/<patch> parameter is the same one that appears as a Gerrit download option for the CL. The default is to build once, benchmark 15 times. The results are also uploaded with benchsave to perf.golang.org.

cmpjob.sh <branch-or-tag> <branch-or-tag> [options]

This checks out two particular tag or branches, and benchmarks the difference. This can be helpful when binary-searching a performance regression. The default is to build once, benchmark 15 times. The results are also uploaded with benchsave to perf.golang.org.

cronjob.sh [options]

This checks out the current development tip and the most recent release (e.g. 1.14) and benchmarks their difference. This can be helpful for nightly performance monitoring. The default is to build 25 times and benchmark 25 times. The results are also uploaded with benchsave to perf.golang.org. The script also contains glue to tweet the results, but by default this will silently do nothing.

cmpcl-phase.sh refs/changes/<nn>/<cl>/<patch> [options]

This checks out a particular version of a CL, and its immediate predecessor, compiles each once with the ssa phase timing flag turned on, does not run benchmarks, and feeds the log (with all the embedded phase timings) to phase-times to help spot any bad performance trends in the new CL. The resulting CSVs can be imported into a spreadsheet and graphed (select the "Test" sheet and scroll down below the vast table of numbers, there is a pretty chart).