For people who know me it's not a secret that PHP performance is my main
responsibility and passion at Zend. Actually, starting from PHP 5.0 we
already made 6 times speedup on synthetic benchmarks and about 2 times
speedup on real-life applications. We endlessly made improvements in PHP
engine and OPCache. However, by PHP 5.5 release we weren’t be able to make
any serious progress, and among other things started to experiment with
memory managers, JIT technologies and other potential ideas.
I spent a significant amount of time experimenting with JIT, and even
created a PoC of transparent LLVM based JIT compiler embedded into OPCache.
The results on bench.php was just amazing – (0.219 seconds against 2.175 – *10
times speedup of PHP 5.5*), but on real-life apps we got just few percent
speedup. This made us look much deeper into some of the runtime
characteristics and what was truly the bottleneck to making more
substantial progress. It was clear the VM is already highly optimized, but
works with data structures that require endless memory allocation,
deallocation and reference counting. Typical real-life PHP application
spends about 20% of the CPU time in memory manager, 10% doing hash tables
operations, 30% in internal functions and only 30% in VM. Of course, we
tried to JIT only VM code and in most cases it had to perform the same
memory allocations. So we decided to change focus and work on the big
bottlenecks. The idea was to change our data types to minimize heap
allocations. This was a very difficult decision because we had to start
with a huge refactoring, and we had no idea whether it’s going to have any
impact or not.
Now I'm glad to present you a result of our recent four month work. It's a
refactoring of the PHP engine that significantly improves performance,
memory usage and builds a foundation for a lot more future performance
improvements incl. JIT. I'll avoid technical details (more details will be
published at *http://wiki.php.net/phpng <http://wiki.php.net/phpng>*), but
in few words - we changed the basement trying to keep most of the building
unchanged. Right now the new engine already makes *10-30% speedup of
php*not only on benchmarks but on real-life applications as well!
*Some benchmarks we ran so far:*
Wordpress 3.6 – 20.0% gain (253 vs 211 req/sec)
Drupal 6.1 – 11.7% gain (1770 vs 1585 req/sec
Qdig – 15.3% gain (555 vs 482 req/sec)
ZF test app – 30.5% gain (217 vs 166 req/sec)
On some apps we show better results than other PHP implementations. It will
be great if others here could test this on their apps and compare to their
existing PHP version to get additional results.
The re-factoring is not finished yet as the focus was to first test whether
this effort would deliver results. Not all extensions are supported, some
tests are failing, and we also have more ideas for additional improvement.
But we feel, we’ve proven enough out to open it up for review, feedback
and assistance, and wanted to involve the community as soon as we managed
to get on a promising direction. There’s more work to do in finishing
support of all extensions and continue to make some additional engine
Please try the refactored PHP engine and provide feedback re: performance,
memory usage and any issues that come up. You may find it in *phpng* branch
at *php.net <http://php.net>*. Some instructions may be found at
<http://wiki.php.net/phpng>*. As mentioned, there are some missing
extensions so not everything will run.
I would like to say many thanks to Xinchen and Nikita who made significant
part of presented work.
I think that this engine can make the new major version of PHP we’re
talking about a lot more interesting.