Playing with the PHP JIT

The PHP JIT RFC is a hot topic on the internals list right now and the voting has started for it to be included in PHP 8.0 and as experimental feature in 7.4.

I wanted to test it out myself, here are the steps necessary to get started on a Linux (Ubuntu) server (or desktop):

git clone https://github.com/php/php-src.git
cd php-src
git remote add zendtech https://github.com/zendtech/php-src.git
git checkout zendtech/jit-dynasm-7.4
./buildconf
./configure  --prefix=/opt/php/php-7.4 --enable-opcache --enable-opcache-jit --with-zlib --enable-zip --enable-json --enable-sockets --without-pear
make -j4
sudo make install

For testing I needed a more realistic problem that was bit more complex than PHPs internal benchmark (which the JIT doubles in speed).

Luckily I came across a good one at Symfony User Group in Cologne this week: The react-php-redis server by @another_clue re-implements redis server in PHP with almost zero PHP extension dependencies. That means the vanilla build from above with no dependencies is enough to get it running.

In addition it fully works with the redis-benchmark command that the original redis-server package includes, so it takes no effort to make some tests. The benchmark pegs the PHP redis server to 100% making it a good candidate for testing JIT.

The code is doing async I/O so the Redis protocol parsing and internal handling should play a significant role in this code that might be optimizable by the JIT.

git clone https://github.com/clue/php-redis-server.git
cd php-redis-server/
composer install

I ran it without the JIT:

/opt/php/php-7.4/bin/php bin/redis-server.php --port 6380

And with the JIT use these flags:

/opt/php/php-7.4/bin/php -dopcache.enable_cli=1 -dopcache.jit_buffer_size=50000000 -dopcache.jit=1235 bin/redis-server.php --port 6380

The -dopcache.jit=1235 only jits the HOT functions that are called often.

Then to check their performance in relation to each other, I ran redis-benchmark -p 6380 -q against the servers (from the redis-tools package).

Don’t take my numbers for gold (I ran them on my busy desktop machine), but you can see a 4-23% improvement depending on the benchmarked command.

Benchmark Nojit Jit % change
PING_INLINE 30674.85 31877.59 3.92%
PING_BULK 87873.46 95969.28 9.21%
SET 81766.15 87336.24 6.81%
GET 81433.22 91575.09 12.45%
INCR 77881.62 83682.01 7.45%
LPUSH 71275.84 79617.83 11.70%
RPUSH 67294.75 79239.3 17.75%
LPOP 73529.41 84530.86 14.96%
RPOP 76103.5 80450.52 5.71%
SADD 84745.77 89686.1 5.83%
HSET 82712.98 91074.68 10.11%
SPOP 87260.03 99700.9 14.26%
LPUSH (needed to benchmark LRANGE) 68493.15 83822.3 22.38%
LRANGE_100 (first 100 elements) 21743.86 26759.43 23.07%
LRANGE_300 (first 300 elements) 9825.11 11923.21 21.35%
LRANGE_500 (first 450 elements) 6819.42 8272.67 21.31%
LRANGE_600 (first 600 elements) 5120.33 5707.11 11.46%
MSET (10 keys) 45998.16 52631.58 14.42%

Looking at the results in a system profiler (perf) I can see that the PHP process is spending a lot of time in I/O functions, so these numbers are not showing the full potential of the JIT with CPU bound code.

Sign up for my newsletter to get notified when I post new content on this blog and with the occasional exclusive content only for subscribers.

By clicking on the Subscribe button I am giving my consent for Benjamin Eberlei to hold my name and email address for the purposes of contacting me with a newsletter on the topics of this blog. You can unsubscribe with one click at any time and withdraw your consent. No spam. I will never share your e-mail address. Privacy Policy