Erlang Performance & Benchmarking Suite.
Simple way to say "this code is faster than that one". See CLI reference
and detailed API reference for erlperf
and erlperf_job
modules.
Build (tested with OTP 23, 24, 25):
$ rebar3 as prod escriptize
Beware of the shell escaping your code in an unpredictable way!
- Run a single process iterating
rand:uniform()
in a tight loop for 3 seconds, printing average iterations per second (~17 millions) and an average time to run a single iteration (57 ns).
$ ./erlperf 'rand:uniform().'
Code || QPS Time
rand:uniform(). 1 17266 Ki 57 ns
- Run four processes doing this same concurrently.
$ ./erlperf 'rand:uniform().' -c 4
Code || QPS Time
rand:uniform(). 4 53893 Ki 74 ns
- Benchmark
rand:uniform()
vscrypto:strong_rand_bytes/1
for 10 seconds, adding an extra second to warm up the algorithms.
$ ./erlperf 'rand:uniform().' 'crypto:strong_rand_bytes(2).' --samples 10 --warmup 1
Code || Samples Avg StdDev Median P99 Iteration Rel
rand:uniform(). 1 10 16611 Ki 0.20% 16614 Ki 16664 Ki 60 ns 100%
crypto:strong_rand_bytes(2). 1 10 1804 Ki 0.79% 1797 Ki 1829 Ki 554 ns 11%
- Run a function passing the state into the next iteration. This code demonstrates performance difference
between
rand:uniform_s/1
with state passed explicitly, andrand:uniform/1
reading state from the process dictionary.
$ ./erlperf 'r(_, S) -> {_, N} = rand:uniform_s(S), N.' --init_runner 'rand:seed(exsss).' \
'r() -> rand:uniform().'
Code || QPS Time Rel
r(_, S) -> {_, N} = rand:uniform_s(S), N. 1 26180 Ki 38 ns 100%
r() -> rand:uniform(). 1 16958 Ki 58 ns 65%
- Estimate
code:is_loaded/1
concurrency characteristics. This function is implemented asgen_server:call
, and all calculations are done in a single process. It is still possible to squeeze a bit more from a single process by putting work into the queue from multiple runners.
$ ./erlperf 'code:is_loaded(local_udp).' --init 'code:ensure_loaded(local_udp).' --squeeze
Code || QPS Time
code:is_loaded(local_udp). 4 969 Ki 4129 ns
$ ./erlperf 'persistent_term:put(atom, "string").' -q
Code || QPS Time
persistent_term:put(atom, "string"). 1 8882 Ki 112 ns
- Start a server (
pg
scope in this example), use it in benchmark, and shut down after.
$ ./erlperf 'pg:join(scope, group, self()), pg:leave(scope, group, self()).' \
--init 'pg:start_link(scope).' --done 'gen_server:stop(scope).'
Code || QPS Time
pg:join(scope, group, self()), pg:leave(scope, group, self()). 1 336 Ki 2976 ns
- Run the same code with different arguments, returned from
init_runner
function. Note the trick of adding extra spaces in the source code to know which code is where.
$ ./erlperf 'runner(X) -> timer:sleep(X).' --init_runner '1.' \
' runner(X) -> timer:sleep(X).' --init_runner '2.'
Code || QPS Time Rel
runner(X) -> timer:sleep(X). 1 500 2001 us 100%
runner(X) -> timer:sleep(X). 1 333 3001 us 67%
- Determine how many times a process can join/leave pg2 group on a single node (requires OTP 23 or older, as pg2 is removed in later versions).
$ ./erlperf 'ok = pg2:join(g, self()), ok = pg2:leave(g, self()).' --init 'pg2:create(g).'
Code || QPS Time
ok = pg2:join(g, self()), ok = pg2:leave(g, self()). 1 64021 15619 ns
- Compare
pg
withpg2
running in a 3-node cluster. Note the-i
argument spawning an isolated extra Erlang VM for each benchmark.
./erlperf 'ok = pg2:join(g, self()), ok = pg2:leave(g, self()).' --init 'pg2:create(g).' \
'ok = pg:join(g, self()), ok = pg:leave(g, self()).' --init 'pg:start(pg).' -i
Code || QPS Time Rel
ok = pg:join(g, self()), ok = pg:leave(g, self()). 1 241 Ki 4147 ns 100%
ok = pg2:join(g, self()), ok = pg2:leave(g, self()). 1 1415 707 us 0%
- Watch the progress of your test running (
-v
option) with extra information: scheduler utilisation, dirty CPU & IO schedulers, number of running processes, ports, ETS tables, and memory consumption. Last column is the job throughput. When there are multiple jobs, multiple columns are printed. Test will continue until adding 8 more workers (-t 8
) does not increase total throughput.
$ ./erlperf 'rand:uniform().' -q -v -t 8
YYYY-MM-DDTHH:MM:SS-oo:oo Sched DCPU DIO Procs Ports ETS Mem Total Mem Proc Mem Bin Mem ETS <0.84.0>
2023-01-22T11:02:51-08:00 6.12 0.00 0.20 46 2 21 24737 Kb 4703 Kb 191 Kb 471 Kb 14798 Ki
2023-01-22T11:02:52-08:00 6.31 0.00 0.00 46 2 21 25105 Kb 5565 Kb 218 Kb 472 Kb 16720 Ki
2023-01-22T11:02:53-08:00 6.26 0.00 0.00 46 2 21 25501 Kb 5427 Kb 218 Kb 472 Kb 16715 Ki
<...>
2023-01-22T11:03:37-08:00 100.00 0.00 0.00 61 2 21 25874 Kb 5696 Kb 221 Kb 472 Kb 55235 Ki
2023-01-22T11:03:38-08:00 100.00 0.00 0.00 61 2 21 25955 Kb 5565 Kb 218 Kb 472 Kb 55139 Ki
Code || QPS Time
rand:uniform(). 8 61547 Ki 130 ns
Running benchmark is called a job, see erlperf_job
for detailed description.
Every job has a controller process, responsible for starting and stopping worker
processes, or workers. Worker processes execute runner function in a tight
loop, incrementing iteration counter.
Benchmark runs either for a specified amount of time (sample duration in continuous mode), or until requested number of iterations is made (timed mode). Resulting sample is total number of iterations for all workers, or elapsed time it took in timed mode.
The process repeats until the specified amount of samples is collected, producing a report (see details below).
For comparison convenience, basic reports contain QPS - historical metric from the original implementation (designed for network service throughput assessment). It is approximate amount of runner iterations per sample_duration achieved by all workers of the job. Given that default duration is 1 second, QPS is a good proxy for the total job throughput.
Single worker performance can be estimated using time metric. It can also be considered as function latency - how long it takes on average to execute a single iteration of a runner.
A benchmark may define following functions:
- runner: code that is executed in the tight loop
- init (optional): executed once when the job starts
- done (optional): executed once when the job is about to stop
- init_runner (optional): executed on every worker process startup
See erlperf_job
for the detailed reference and ways to define a function (callable).
Note that different ways to call a function have different performance characteristics:
$ ./erlperf '{rand, uniform, []}' 'rand:uniform().' -l 10M
Code || QPS Time Rel
rand:uniform(). 1 18519 Ki 54 ns 100%
{rand,uniform,[]} 1 16667 Ki 60 ns 90%
This difference may get more pronounced depending on ERTS version and runner code:
(erlperf@max-au)7> erlperf:benchmark([
#{runner => "runner(X) -> is_float(X).", init_runner=>"2."},
#{runner => {erlang, is_float, [2]}},
#{runner => fun (X) -> is_float(X) end, init_runner => "2."}],
#{}, undefined).
[105824351,66424280,5057372]
It is caused by the ERTS: running compiled code (first variant) with OTP 25 is
two times faster than applying a function, and 20 times faster than repeatedly
calling anonymous fun
. Use the same invocation method to get a relevant result.
Absolute benchmarking overhead may be significant for very fast functions taking just a few nanoseconds. Use timed mode for such occasions.
See erlperf
module documentation and command line reference for all available options.
Benchmarking is done by counting number of runner iterations done over a specified period of time (sample_duration).
Two examples below demonstrate the effect caused by changing sample_duration.
First run takes 20 samples (-s 20
) with 100 ms duration. Second invocation
takes the same 20 sample, but with 200 ms duration (-d 200
). Note that all metrics,
except a single iteration time, doubled.
$ ./erlperf 'rand:uniform().' -d 100 -s 20
Code || Samples Avg StdDev Median P99 Iteration
rand:uniform(). 1 20 1647 Ki 0.39% 1648 Ki 1660 Ki 60 ns
$ ./erlperf 'rand:uniform().' -d 200 -s 20
Code || Samples Avg StdDev Median P99 Iteration
rand:uniform(). 1 20 3354 Ki 0.16% 3354 Ki 3368 Ki 59 ns
In this mode runner code is executed for sample_duration iterations for every sample.
Report contains average/median/p99 time it takes to produce a single sample. In the
example below, it takes an average of 554 ms to make 10 million calls to rand:uniform()
.
$ ./erlperf 'rand:uniform().' 'rand:uniform(1000).' -l 10M -s 20
Code || Samples Avg StdDev Median P99 Iteration Rel
rand:uniform(1000). 1 20 554 ms 0.37% 554 ms 563 ms 55 ns 100%
rand:uniform(). 1 20 560 ms 0.60% 560 ms 564 ms 55 ns 99%
Effectively, this example runs following code: loop(0) -> ok; loop(Count) -> rand:uniform(), loop(Count - 1).
Timed mode has slightly less overhead compared to continuous mode.
Timed mode does not support --concurrency
setting, using only
one process. However, it does support comparison run with multiple concurrent jobs.
In this mode erlperf
performs multiple continuous benchmarks with
increasing concurrency. The test concludes when increasing worker
count does not result in increase of the total throughput. Report
contains statistics of the most successful run.
This mode can also be used to detect bottlenecks, e.g. lock contention, single
gen_server
processes, or VM-wide shared resources (persistent_term
s).
Example (with maximum concurrency limited to 50):
$ ./erlperf '{code, is_loaded, [local_udp]}' -w 1 --max 50 -q
Code || QPS Time
{code,is_loaded,[local_udp]} 6 1665 Ki 3604 ns
Same in the Erlang shell:
> erlperf:run({code, is_loaded, [local_udp]}, #{warmup => 1}, #{max => 50}).
{1676758,6}
In this example, 6 concurrent processes were able to squeeze 1676758 calls per second
for code:is_loaded(local_udp)
. In current OTP version code:is_loaded
is implemented
as a gen_server:call
to a single process (code_server
), that limits potential
performance.
See erlperf_job
for the detailed description of different benchmarking modes.
Historically erlperf
had only the basic reporting available for command line
usage. Since 2.2 it is possible to request additional information.
This is the default report form when less than 10 samples were collected.
Use -r basic
to force basic reports with 10 and more samples.
Basic report contains following columns:
- Code: Erlang code supplied to the benchmark
- ||: how many concurrent processes were running. In the timed mode, it is always 1. In the concurrency estimation mode, the number that achieved the highest total throughput (QPS)
- QPS: average number of runner code iterations (throughput). Measure per single sample_duration in the continuous mode. In the timed mode, calculated with the assumption that sample_duration is 1 second
- Time: single runner iteration time
- Rel: relative performance of this code, compared to others. Printed only when more than one runner is specified.
When 10 or more samples were collected, this mode is the default. Use -r extended
to force printing this report for smaller sample sets.
Note that average, deviation, median and 99th percentile are calculated for the sample_duration.
If you requested 20 samples of 100 ms in the continuous mode, these fields will contain iteration
count per 100 ms. If you requested 10 million iterations (-l 10M
), extended report for timed mode
displays average time it takes to do 10M iterations. Single iteration time is printed as Iteration.
Code, concurrency, and relative performance fields have the same meaning as in basic report. In addition, following columns are printed:
- Samples: how many samples were collected (useful when requesting continuous test with standard deviation requirement)
- Avg: same as QPS for continuous mode, but in the timed mode, average sample time
- StdDev: standard deviation from average
- Median: median value, in the continuous mode, median estimated throughput, in the timed mode - time to complete the requested iterations
- Iteration: single runner iteration time
- P99: 99th percentile
This mode must be explicitly specified with -r full
.
Contains everything that extended report has. Includes extra information about the system used for benchmarking - OS type, CPU and Erlang VM characteristics.
erlperf
can be used to measure performance of your application running in production, or code that is stored
on disk.
Use -pa
argument to add an extra code path. Example:
$ ./erlperf 'argparse:parse([], #{}).' -pa _build/test/lib/argparse/ebin
Code || QPS Time
argparse:parse([], #{}). 1 955 Ki 1047 ns
If you need to add multiple released applications, supply ERL_LIBS
environment variable instead:
$ ERL_LIBS="_build/test/lib" erlperf 'argparse:parse([], #{}).'
Code || QPS Time
argparse:parse([], #{}). 1 735 Ki 1361 ns
It is possible to use erlperf
to benchmark an application running in production.
Add erlperf
as a dependency, and use remote shell:
# connect a remote shell to the production node
erl -remsh production@max-au
(production@max-au)3> erlperf:run(timer, sleep, [1]).
488
You can run a job continuously, to examine performance gains or losses while doing hot code reload. This process is designed to help during development and testing stages, allowing to quickly notice performance regressions.
Example source code:
-module(mymod).
-export([do/1]).
do(Arg) -> timer:sleep(Arg).
Example below assumes you have erlperf
application started (e.g. in a rebar3 shell
)
% start a logger that prints VM monitoring information
> {ok, Logger} = erlperf_file_log:start_link().
{ok,<0.235.0>}
% start a job that will continuously benchmark mymod:do(),
% with initial concurrency 2.
> JobPid = erlperf:start(#{init_runner => "rand:uniform(10).",
runner => "runner(Arg) -> mymod:do(Arg)."}, 2).
{ok,<0.291.0>}
% increase concurrency to 4
> erlperf_job:set_concurrency(JobPid, 4).
ok.
% watch your job performance
% modify your application code,
% set do(Arg) -> timer:sleep(2*Arg), do hot code reload
> c(mymod).
{module, mymod}.
% see that after hot code reload throughput halved!
ERTS cannot guarantee precise timing when there is severe lock contention happening, and scheduler utilisation is 100%. This often happens with ETS:
$ ./erlperf -c 50 'ets:insert(ac_tab, {1, 2}).' -d 100 -s 50
Timer accuracy problem detected, results may be inaccurate
Code || Samples Avg StdDev Median P99 Iteration
ets:insert(ac_tab, {1, 2}). 50 50 6079 82.27% 5497 40313 823 us
Running 50 concurrent processes trying to overwrite the very same key of an ETS
table leads to lock contention on a shared resource (ETS table/bucket lock). erlperf
may detect this issue and switch to a busy wait loop for precise timing. This may
result in lowered throughput and other metrics skew. erlperf
does not attempt to
pinpoint the source of contention, it is up to user to figure that out. It's recommended
to use lock-counting emulator, or Linux perf
utility to troubleshoot VM-level issues.
These features are not fully supported. APIs may change in the future erlperf
releases.
It's possible to run a job on a separate node in the cluster. See
erlperf_cluster_monitor
for additional details.
% watch the entire cluster (printed to console)
(node1@host)> {ok, _} = erlperf_history:start_link().
{ok,<0.213.0>}
(node1@host)> {ok, ClusterLogger} = erlperf_cluster_monitor:start_link(group_leader(), 1000, [node, sched_util, jobs]).
{ok, <0.216.0>}
% also log cluster-wide reports to file (jobs & sched_util)
(node1@host)> {ok, FileLogger} = erlperf_cluster_monitor:start_link("/tmp/cluster", 1000, [time, node, sched_util, jobs]).
{ok, <0.223.0>}
% run the benchmarking process in a different node of your cluster
(node1@host)> rpc:call('node2@host', erlperf, run, [#{runner => {rand, uniform, []}}]).