Facebook's HipHop Also a PHP Webserver

darthcamaro writes "As expected, Facebook today announced a new runtime for PHP, called HipHop. What wasn't expected were a few key revelations disclosed today by Facebook developer David Recordan. As it turns out, Facebook has been running HipHop for months and it now powers 90 percent of their servers — it's not a skunkworks project; it's a Live production technology. It's also not just a runtime, it's also a new webserver. 'In general, Apache is a great Web server, but when we were looking at how we get the next half percent or percent of performance, we didn't need all the features that Apache offers," Recordon said. He added, however, that he hopes an open source project will one day emerge around making HipHop work with Apache Web servers.'"

Comments

A stupid question...

[__aaclcg7560]

For all the trouble you're going through to convert PHP into C++ (300,000 lines and 5,000 unit tests), wouldn't programming in C++ in the first place be easier?

Re:A stupid question...

[shutdown -p now]

As a programming language, PHP is simple. Simple to learn, simple to write, simple to read, and simple to debug.

PHP is not a simple language. A keymark of a simple language is consistency, and PHP is anything but - I won't even touch on the mess that is the standard function library, but just the language itself. For example, this gem, taken directly from the language spec, regarding array indices/keys:

A key may be either an integer or a string. If a key is the standard representation of an integer, it will be interpreted as such (i.e. "8" will be interpreted as 8, while "08" will be interpreted as "08"). Floats in key are truncated to integer. The indexed and associative array types are the same type in PHP, which can both contain integer and string indices.

This is awesome on many levels. The obvious fubar is the treatment of "8" vs "08" (and note that, while it is clearly obvious when a string literal is used in the source code, how about a string variable, or other expression computed at runtime?). But the bit about silent float->int truncation is also interesting, especially the "silent" part. Combined with rounding errors and the overall non-obviousness of binary floating-point arithmetic (especially to a typical PHP coder), this design decision is just hilarious.

I've long held the opinion that C/C++ rules on mixed signed/unsigned arithmetic and comparisons are a good example of awful language design, but PHP beats that by a margin so large it's not even funny.

Oh, I also don't know of any other language that has what effectively amounts to synactic sugar for try/catch with an empty catch block. Good programming practices FTW!

I find it curious, by the way, that PHP coders like to compare the language to C++ or Java - where it actually has some subjective advantages, such as dynamic typing - but very rarely to Perl, Python or Ruby, where all such advantages disappear, but design flaws immediately stand out.

Re:A stupid question...

[MightyMartian]

PHP is a lot better environment to develop new features quickly and doesn't get you into so many security pitfalls. And they're already using C++ for some parts of the site:

Except for the problem that, historically, PHP is been one big vast security pitfall.

Re:A stupid question...

[shutdown -p now]

Because it's really easy to create memory leaks and similair bugs in C++.

It's very easy to get rid of memory leaks in C++, as well. A very simple rule: never, ever write a type declarator with a * in it. In other words, no raw pointers - use Boost/TR1 shared_ptr, or roll out your own, it doesn't matter - just use it consistently. At that point, you can still get reference cycles (which are also leaks), but you can do that in PHP 5.3 - which also uses reference counting with no GC for cyclical references - just as well. And the usage of 5.3 so far is minuscule.

Alternatively, just use any of third-party tracing GCs, such as Boehm.

By the way, from personal experience, I find that languages with built-in reference counting and no cycle detection (those I know of are VB6 and PHP) are actually more prone to memory leaks when coding that languages with explicit memory management. The reason is that, in, say, C++, coders are actually aware of issues such as memory allocation, and view smart pointers as convenient helpers, not as some kind of magic fairy dust. Because of that, the question of "what happens if two smart pointers reference each other" is a rather obvious one, and the issue is noticed and rectified early on. In contrast, in VB6 and PHP, you don't have to deallocate explicitly, so refcounting is magic - many people don't even understand how the algorithm works! - that is, until you run into a cyclic reference that leaks...

Re:A stupid question...

[abulafia]

No, I think I'll stick with attacking it for being a truly crappy language. I don't care that it is slow or wastes memory. If you're paying$20/month for your dance-school-business-calendar installed and customized by a local teenager, the idea of writing a web app in C is silly for efficiency is silly. Likewise, whatever the intent of @ is for, I most certainly expect people have and will abuse it in exactly the way described to "fix" problems. People endlessly bash, for instance, Perl as being write only, and there's truth to that. But there is truth to that because the language tends to encourage hard to read code. You can say that's not the intent, and you'd be right, but that doesn't matter. (Though I do still love Perl.)

I do agree that PHP is fine for toy web sites, and that people get themselves in trouble using the executable web page model because the don't know what they're doing. These things are true for the same reason: PHP is full of poorly thought out magic that allows people to get in over their head, and doesn't provide the tools to easily dig back out. I'm all for making programming more accessible, but encouraging people to foot-bullet themselves in predictable ways doesn't strike me as a good approach.

I dislike it for other reasons, but for instance Ruby on Rails is a much more solid approach, in my opinion - the path of least resistance is generally the right thing to do, once a newbie internalizes the MVC idea and a couple conceptual points the learning curve is pretty gentle, and Ruby is a pretty well constructed language that lets people grow into using more conceptually useful techniques over time without the up-front demands of learning, say, Lisp.

(While I'm chasing people off my lawn, the whole RoR mindset seems to lead people down a rabbit hole of writing dumb little DSLs -- who on earth thinks a toy language for generating CSS is a good idea? You just push yourself one more indirection layer away from what's going on and end up dinking around with yet another silly new syntax for your effort. Muppet coding at its worst.)

Re:Ambitious

[sopssa]

He said they were struggling just to get half a percent more performance with Apache. That had nothing to do with "HipHop".

In reality their CPU usage dropped average 50%

With HipHop we've reduced the CPU usage on our Web servers on average by about fifty percent, depending on the page.

Re:i can hear it now

[bill_mcgonigle]

the php haters: "look how awful php is, you need to convert everything into c++ before you can use it in really large scale deployments!"

"Look how awful C++ is, you have write bits in assembly to get it to really run."

"Look how awful assembly is, you really optimize when you can write machine opcodes."

And the microcode guys just glare out from their caves with their glowy little eyes in incredulity.

Elsewhere is heard, "You guys still use CPU's? It's the GPU decade, dude."

And somebody down the hall builds an ASIC to solve a specific problem and thinks he's so smart.

But, the analog EE understands his elegant circuit doesn't enable a team of 200 developers to build the top social networking site.

Re:Well, I Guess "HipHop" Is the New Champ!

[NevDull]

My guess is that it was probably a progression from "Haiping's PHP" to HPHP to HipHop.
Two syllables vs four or more... looks like they're not just computing more efficiently, but also speaking more efficiently!

Re:GUI applications

[dgatwood]

What makes PHP nice is that it is so close to C. For people who are comfortable working in C, PHP is just a few dollar signs away.

Re:GUI applications

[ls671]

Application servers based on Java are heavy only on start up, the allocated memory is then reused which makes it light on system load once started.

Java uses some of its memory to cache machine code in order to re-execute it the next time it is needed and this also makes it light on system load.

Simply by using top, you could understand what I am talking about. Java uses more memory but it is otherwise very light on system load and guess what ?

Machines typically have 4 GB of ram nowadays.

Most people bitching about Java being heavy do not understand what I am trying to explain to you here ;-)

Re:GUI applications

[eihab]

Hopefully he has upgraded to the "once in a while" switch replacement technique.

That struck me as weird, because as a programmer you usually start with conditionals and then move on to loops. I had a hard time believing that someone would know of "while(true)" and not "else if".

So I decided to run some tests over dinner. I'm no C++ programmer but here's how I went with this.

First I wrote a tests.cpp that looks like this:

#include
int main () {
    int subType, mainType = 11;

    Slashdot_Filter_Sucks // Editable section
    while (true) {
        if (mainType == 7) {
            subType = 4;
            break;
        }
        if(mainType == 9) {
            subType = 6;
            break;
        }
        if(mainType == 11) {
            subType = 9;
            break;
        }
        break;
    }
    Slashdot_Filter_Sucks // End of editable

    std :: cout

I compiled that and it resulted in a 8120 bytes binary that ran in 0.005ms.

I thought about other obvious and simple ways to write this code and I created four more versions that are identical except for the code between the dividers (I had pretty asterisk lines but Slashdot's junk filter made me take it off). They are:

testif.cpp (test using an if/else statement):

if (mainType == 7) subType = 4;
else if (mainType == 9) subType = 6;
else if (mainType == 11) subType = 9;

testifonly.cpp (no else, only ifs):

if (mainType == 7) subType = 4;
if (mainType == 9) subType = 6;
if (mainType == 11) subType = 9;

testswitch.cpp (using a switch statement):

switch(mainType) {
    case 7: subType = 4;
    case 9: subType = 6;
    case 11: subType = 9;
}

testp.cpp (subtract 3 from mainType since that seemed like a pattern):

subType = mainType - 3;

I compiled everything using g++ then I ran time ./output. All the versions ran on average in 0.005ms, however, the binary sizes were different:

#ls -l (ordered by size)
8072 testp
8109 testifonly
8120 tests
8121 testif
8125 testswitch

Ok, no case here in terms of size. So I tried compiling again with -O3, and the results were:

#ls -l (ordered by size)
8024 testp_o3
8024 tests_o3
8025 testif_o3
8029 testifonly_o3
8029 testswitch_o3

Here it seems that the subtraction and the weird while/break method have the smallest file size. Without code context, one can imagine that subType was to be left alone if mainType was not 7,9 or 11. Which would mean the subtraction code wouldn't work in that scenario.

Now, I don't know the intricacies of C++ or Assembly, but I have to wonder if this was the work of a moron or someone who knew exactly what they were doing and did so for a reason.

Again, without context, none of this matters.