Benchmarks
KPHP has strong limitations in its area of applicability, but when code compiles, it runs much faster than PHP.
Zend benchmarks
Comparing performance on Zend benchmarks is absolutely synthetic and done just formally.
We'll take micro_bench.php and bench.php from the PHP repository.
Compiling them without modification fails, but after removing a couple of tests and correcting errors, they can be easily launched.
It comes about such results for 5 runs (PHP 7.4 with OpCache and without XDebug):
| micro_bench — PHP 7.4 | micro_bench — KPHP | bench — PHP 7.4 | bench — KPHP |
|---|---|---|---|
| 3.383 | 0.459 | 0.620 | 0.120 |
| 3.183 | 0.471 | 0.632 | 0.121 |
| 3.261 | 0.454 | 0.773 | 0.121 |
| 3.336 | 0.460 | 0.660 | 0.117 |
| 3.320 | 0.456 | 0.610 | 0.116 |
| avg PHP 7.4 | avg KPHP | avg PHP 7.4 | avg KPHP |
| 3.297 | 0.460 | 0.659 | 0.119 |
According to these benchmarks, KPHP is found to be 5–7 times faster.
Although they are quite synthetic, this is pretty much similar to other measurements.
Compare PHP vs KPHP vs C++
Consider the dedicated page. It also shows a 4–10 times breakthrough.
The main idea to pick from that text: as far as your project grows, using KPHP becomes more reliable.
This can be called “zero-cost abstractions” — not zero of course, but considerably smaller than in PHP:
- all defines and class constants are inlined at compile-time
- all simple methods, like getters and setters, are inlined and have no invocation overhead
- access modifiers are checked at compile-time, they don't exist at runtime at all
- class fields access is a compile-time known memory offset, without any hashtable lookups
- … and many others
But still: the only adequate way to compare performance is to take a real working high-loaded site.
VK.com pages generation time
If we measure PHP 7.2 version and KPHP-compiled on the same server, for a random account, we get
vk.com/feed — PHP: 10.1 sec, KPHP: 0.95 sec
vk.com/im — PHP: 1.25 sec, KPHP: 0.46 sec
vk.com/ads — PHP: 0.73 sec, KPHP: 0.20 sec
In general, for VK.com we have from 3 to 10 times average profit. Pretty much to suffer a bit, right?
Having a goal to optimize particular places, you can achieve even better results.
Machine-learning models with lots of maths
VK.com has some heavy pieces of business logic, which perform ML calculations in real-time. Such areas were specially optimized, fulfilling all KPHP potential.
Measuring that isolated pieces of code shows tremendous superiority of strict typing:
neural networks with huge matrices — PHP: ~8 sec, KPHP: 0.24 sec
deep decision trees with tons of branches — PHP: ~5 sec, KPHP: 0.14 sec
FFI has almost zero overhead
KPHP supports FFI (foreign functions interface) — pretty the same as PHP does.
But FFI in KPHP works much, much faster, because KPHP internals is also C++ actually, like any FFI library. KPHP is able to remove excess CData casting, send internal pointers to FFI functions, etc. If you are sure that particular FFI calls are signal-safe, you can even disable critical sections for them, and achieve almost native interop speed, tens of times faster than in PHP.
But. It’s a bad idea to state “KPHP is always faster than PHP”.
It may be wrong. If PHP code just calls strpos() or parse_url(), KPHP would have almost no benefits, as they are just PHP standard functions implemented in C. But the more business logic is PHP-implemented, the faster KPHP becomes in comparison — especially if you write PHP code keeping types in mind.
You can speed up your code — if you really ought to.