Overview

The goal is to provide a benchmark suite, testing something representative of real-world situations. This suite also includes some unrealistic microbenchmarks - comparing the results of these is fairly pointless, however they can still be useful to profile, to find optimization opportunities that may carry over to a real site.

This script configures and runs nginx, siege, and PHP5/PHP7/HHVM over FastCGI, over a TCP socket. Configuration is as close to identical as possible.

The script will run 300 warmup requests, then as many requests as possible in 1 minute. Statistics are only collected for the second set of data.

Results

We don't have anything to share yet - we want to standardize and document how the interpreters are built/installed first.

Usage

As a regular user:

hhvm composer.phar install # see https://getcomposer.org/download/
hhvm perf.php --wordpress --php5=/path/to/bin/php-cgi # also works with php7
hhvm perf.php --wordpress --php=/path/to/bin/php-fpm # also works with php7
hhvm perf.php --wordpress --hhvm=/path/to/hhvm

Running with --hhvm gives some additional server-side statistics. It is usual for HHVM to report more requests than siege - some frameworks do call-back requests to the current webserver.

:heavy_exclamation_mark: If you run with a debug build you may hit timeouts and other issues.

Batch Usage

If you want to run multiple combinations:

hhvm composer.phar install # see https://getcomposer.org/download
hhvm batch-run.php < batch-run.json > batch-run-out.json

See batch-run.json.example to get an idea of how to create batch-run.json.

Requirements

On Ubuntu you can run scripts/setup.sh. This should provision your machine with everything you need to begin running the benchmarks.

This installs:

• composer
• nginx
• siege (versions 2.x, or 3.1.x or 4.0.3)
• unzip
• A mysql server on 127.0.0.1:3306
• hhvm

I've been using the current versions available from yum on Centos 6.3. HHVM is required as this is written in Hack.

Siege 3.0.x is not supported; as of writing, all 3.0.x releases have bugs that make it unable to correctly make the benchmark requests. 4.0.0, 4.0.1, 4.0.2 all automatically request resources on pages, and should not be used for benchmarking.

The Targets

Toys

Unrealistic microbenchmarks. We do not care about these results - they're only here to give a simple, quick target to test that the script works correctly.

'hello, world' is useful for profiling request handling.

WordPress

• Data comes from installing the demo-data-creator plugin (included) on a fresh install of WordPress, and clicking 'generate data' in the admin panel a bunch of times.
• DISABLE_WP_CRON is set to true to disable the auto-update and other requests to rpc.pingomatic.com and wordpress.org.
  • auto-updating is not suitable for a like-to-like benchmark system like this
  • WP_CRON increases noise, as it makes the benchmark results include the time taken by external sites
  • serious deployments should trigger the WP_CRON jobs via cron or similar:
  • an administrator should be aware of all production changes, especially code version changes
  • scheduled maintainance should not make an unfortunate end-user request take significantly more time

• URLs file is based on traffic to hhvm.com - request ratios are:

  100: even spread over long tail of posts 50: WP front page. This number is an estimate - we get ~ 90 to /, ~ 1 to /blog/. Assuming most wordpress sites don't have our magic front page, so taking a value roughly in the middle. 40: RSS feed 5: Some popular post 5: Some other popular post 3: Some other not quite as popular post

The long tail was generated with:

<?php
  for ($i = 0; $i <= 52; ++$i) {
  printf("http://localhost:__HTTP_PORT__/?p=%d\n", mt_rand(1,52));
}

Ordering of the URLs file is courtesy of the unix 'shuf' command.

Drupal 7

Aims to be realistic. Demo data is from devel-generate, provided by the devel module. Default values were used, except:

• Users were spread over the last year, rather than the last week
• New main menus and navigation menus were created
• New articles and pages were created, with up to 30 comments per content, spread over the last year instead of the last week

As well as the database dump, the static files generated by the above process (user images, images embedded in content) have also been included.

Drupal 8

As above, aims to be realistic. Demo data is from the devel_generate module and default values were used, except:

• Users spread over the last year
• One new menu was created and the navigation menu replaced with 15 items each
• New articles and pages, up to 30 comments per node, spread over the last year

The structure is similar to the Drupal 7 target, except for:

• The settings.php.gz has been replaced by a settings.tar.gz which includes a settings.php, services.yml, and setup.php file.
• The setup.php file is used to pre-populate the Twig template cache so that Repo Authoritative mode can be used.
• A Drupal 8-compatible version of Drush 8 is included with a matching vendor directory. This is used to run the above setup script.

SugarCRM

The upstream installation script provides an option to create demonstration data - this was used to create the database dump included here.

There are two unrealistic microbenchmarks:

• just the login page - the page with the username/password form. Added to confirm a reported issue.
• just the logged-in home page. Added to be a little more realistic than rendering the form, however we have no idea what a realistic request distribution would look like

Laravel

Unrealistic microbenchmark: just the 'You have arrived' page from an empty installation.

Laravel 4 and 5 are both available.

Magento

• Data is official Magento sample data, however because of its original size we have replaced all images with a compressed HHVM logo and removed all mp3 files.
• After importing sample data we use the Magento console installer to do the installation for us.
• URLs are a variety of different pages:
  • Homepage
  • Category page
  • CMS page
  • Quicksearch
  • Advanced search
  • Simple product
  • Product with options
  • Product reviews

MediaWiki

The main page is the Barack Obama page from Wikipedia; this is based on the Wikimedia Foundation using it as a benchmark, and finding it fairly representative of Wikipedia. A few other pages (HHVM, talk, edit) are also loaded to provide a slightly more rounded workload.

Profiling

Perf

perf.php can keep the suite running indefinitely:

hhvm perf.php --i-am-not-benchmarking --no-time-limit --wordpress --hhvm=$HHVM_BIN

You can then attach 'perf' or another profiler to the running HHVM or php-cgi process, once the 'benchmarking' phase has started.

There is also a script to run perf for you at the apropriate moment:

hhvm perf.php --i-am-not-benchmarking --wordpress --hhvm=$HHVM_BIN --exec-after-warmup="./scripts/perf.sh -e cycles"

This will record 25 seconds of samples. To see where most time is spent you can dive into the data using perf, or use the perf rollup script as follows:

sudo perf script -F comm,ip,sym | hhvm -vEval.EnableHipHopSyntax=true <HHVM SRC>/hphp/tools/perf-rollup.php

In order to have all the symbols from the the translation cache you may need to change the owner of /tmp/perf-.map to root.

TC-print

TC-print will use data from perf to determine the hotest functions and translations. TC-print supports a number of built in perf counters. To capture all relevant counters, run the benchmark as follows: NOTE: perf.sh uses sudo, so look for the password prompt or disable it.

# Just cycles
hhvm perf.php --i-am-not-benchmarking --mediawiki --hhvm=$HHVM_BIN --exec-after-warmup="./scripts/perf.sh -e cycles" --tcprint

# All supported perf event types (Intel)
hhvm perf.php --i-am-not-benchmarking --mediawiki --hhvm=$HHVM_BIN --exec-after-warmup="./scripts/perf.sh -e cycles,branch-misses,L1-icache-misses,L1-dcache-misses,cache-misses,LLC-store-misses,iTLB-misses,dTLB-misses" --tcprint

# All supported perf event types (ARM doesn't have LLC-store-misses)
hhvm perf.php --i-am-not-benchmarking --mediawiki --hhvm=$HHVM_BIN --exec-after-warmup="./scripts/perf.sh -e cycles,branch-misses,L1-icache-misses,L1-dcache-misses,cache-misses,iTLB-misses,dTLB-misses" --tcprint

In order to have all the symbols from the the translation cache you may need to change the owner of /tmp/perf-.map to root.

We process the perf data before passing it along to tc-print

sudo perf script -f -F hw:comm,event,ip,sym | <HHVM SRC>/hphp/tools/perf-script-raw.py > processed-perf.data

If perf script is displaying additional fields, then re-run with -F <-field>,...

sudo perf script -f -F -tid,-pid,-time,-cpu,-period -F hw:comm,event,ip,sym | <HHVM SRC>/hphp/tools/perf-script-raw.py > processed-perf.data

tc-print is only built if the appropriate disassembly tools are available. On x86 this is LibXed. Consider building hhvm using:

cmake . -DLibXed_INCLUDE_DIR=<path to xed include> -DLibXed_LIBRARY=<path to libxed.a>

Use tc-print with the generated perf.data:

<HHVM SRC>/hphp/tools/tc-print/tc-print -c /tmp/<TMP DIR FOR BENCHMARK RUN>/conf.hdf -p processed-perf.data

Contributing

Please see CONTRIBUTING.md for details.