HPHP VM Tech Talk

Report
HipHop:
High-Performance
PHP
Ali-Reza Adl-Tabatabai
HipHop Team
Facebook
Facebook: Move fast & build things
3
PHP
• General-purpose scripting language tailored for
web development
• Interactive
• Weakly typed & dynamic
<?php
var_dump(6 + 7);
var_dump('6' + '7');
var_dump('six' + 'seven');
var_dump((1<<63) + 0);
var_dump((1<<63) + (1<<63));
int(13)
int(13)
int(0)
int(-9223372036854775808)
float(-1.844674407371E+19)
4
PHP
• General-purpose scripting language tailored for
web development
• Interactive
• Weakly typed & dynamic
function foo($x) {
echo “foo: “. $x. ”\n”;
}
foo(“hello”); // prints “foo: hello”
foo(10);
// prints “foo: 10”
5
PHP
• General-purpose scripting language tailored for
web development
• Interactive
• Weakly typed & dynamic
if (...) {
class B { ... }
} else {
class B { ... }
}
class C extends B { ... }
6
PHP
• General-purpose scripting language tailored for
web development
• Interactive
• Weakly typed & dynamic
$a = ‘f’;
$b = ‘c’;
$c = ‘oo’;
$func = $a . $ $b;
$func();
7
PHP
• General-purpose scripting language tailored for
web development
• Interactive
• Weakly typed & dynamic
class C {
public $declProp = 1;
}
$obj = new C;
$obj->dynProp = 2;
// This is OK!
echo $obj->declProp . “\n”; // prints “1”
echo $obj->dynProp . “\n” // prints “2”
8
PHP
• General-purpose scripting language tailored for
web development
• Interactive
• Weakly typed & dynamic
if (function_exists(‘foo’)) {
...
}
if (class_exists($c)) {
...
}
9
Memory management
• PHP is reference counted
• Precise destruction semantics
class C {
function __destruct() { echo “bye!”; }
}
$x = new C;
$x = 1;
// prints “bye!”
10
Concurrency
• Single-threaded programming model
• Multiple requests run in parallel
• No shared memory, synchronization, or direct
communication
11
Performance... 
45
40
35
30
25
CPU Time
20
15
10
5
0
C++
Java
C#
Ocaml
PHP
Ruby
Python
Source: http://shootout.alioth.debian.org/u64q/benchmark.php?test=all&lang=all
12
Implications for Facebook
1. Bad performance can limit user features
2. Poor efficiency requires lots of re$ource$
INTERNET
...
webservers
storage
...
13
What have we done?
Facebook
(PHP)
PHP/Zend
Facebook
(binary)
Apache
14
HipHop compilation flow
Facebook
(PHP)
hphpc
Facebook
(C++)
PHP
Runtime
Gcc
Webserver
Facebook
(binary)
15
HipHop compiler (hphpc)
PHP
AST
C++
Parser
C++ Code
Generator
Optimize, Infer Types
Variant
Double
String
Array
Object
Integer
Boolean
16
Representing PHP data
type
data
KindOfStri
ng
& “Lorem
ipsum.”
KindOfInt
13
Type inference: fewer tags!
type
data
& “Lorem
ipsum.”
13
Basic operations
$a + $b
Basic operations: type dispatch
$ $a + $b
a
+
$
b
switch (a->m_type) {
case KindOfInt:
switch (b->m_type) {
…
}
case KindOfArray:
switch (b->m_type) {
…
}
…
}
…
}
Type inference: avoiding dispatch
add %rcx, %rdx
$a + $b
given
$a ::
Int,$b ::
Int
HipHop compiler: performance
45
40
35
30
25
CPU Time
20
15
10
5
0
C++
Java
C#
Ocaml
PHP
Ruby
Python
Disclaimer: estimated based on running Facebook
22
HipHop compiler: pros & cons
• Good for production servers
• Inadequate for development
– Solution: the HipHop interpreter (hphpi)
• Leverages HipHop runtime & webserver
• Open problem:
≠
Can we get the
best of both
worlds?
23
HipHop Virtual Machine (hhvm)
• Ambitious goal: replace both the HipHop
Interpreter and Compiler
HHVM
Interpreter
PHP
AST
Parser
HHBC
Bytecode
Generator
HHVM
JIT
Optimize, Infer Types
24
HipHop bytecode (hhbc)
• In-house design
• Stack-base VM
• Closer to PHP than
machine code
function lookup($cache, $key) {
if (isset($cache[$key])) {
echo “Hit! “ . $cache[$key];
return true;
} else {
echo “Miss!”;
return false;
}
}
96:
101:
106:
113:
Loc 0
Loc 1
IssetM <H E>
JmpZ 32
118:
123:
128:
133:
140:
141:
142:
String “Hit! “
Loc 0
Loc 1
CGetM <H E>
Concat
Print
PopC
143:
144:
True
RetC
145:
150:
151:
String “Miss!”
Print
PopC
152:
153:
False
RetC
25
Hhvm JIT
• Beyond static type inference: dynamic type
specialization
1. Observe types
2. Generate specialized code
3. Guards to check types
$n = 3 * $n + 1;
224:
229:
238:
243:
244:
245:
254:
255:
256:
Loc 0
Int 3
Loc 0
CGetH
Mul
Int 1
Add
SetH
PopC
;; Typecheck: int($n)?
cmpl $4, -4(%rbp)
jne
__retranslate
;; Type-spec xlation
mov $3, %r12d
mov -16(%rbp), %r13
mov %r13, %r14
imul %r14, %r12
add $1, %r12
mov %r12, %r13
mov $0x40000000, %r8
mov %r8, -8(%rbp)
mov %r13, -16(%rbp)
26
Translation cache:
Reuse & specialization
$n = 1.5;
...
$n = 3 * $n + 1;
Translator
...
__retranslate:
...
Translation Cache
T1:
;; Typecheck: INT($n)?
cmpl $4, -4(%rbp)
jne
__retranslate
T2
;; Type-spec INT
;; translation
. . .
T2: ;;
Typecheck: DOUBLE($n)?
cmpl $8, -4(%rbp)
jne
__retranslate
;; Type-spec DOUBLE
;; translation
. . .
27
Current state
• hhpc
– Runs www.facebook.com
– Paper to appear in SPLASH ‘12
• hhvm
– www.facebook.com works
– Developers using it
– ~27% slower than hphpc
• Download from github:
https://github.com/facebook/hiphop-php/
28
Perf progress
6/11-7/14
Ongoing & future work
• Performance
– Profile-guided, SSA-based 2nd gear JIT
– Type prediction
– Tuning for the HW
– Array shapes: turn hash tables into structs
• Tracing garbage collection
– Copy-on-write arrays
– Precise destruction semantics
• Language extensions
30
Summary
• PHP enables us to move fast
• Performance suffers because of interpreter
• Hiphop compiler
– Compiles PHP to C++ offline
– Significantly improves user experience & data center
efficiency
• HipHop virtual machine
– A new language VM tailored to PHP
– Brings dynamic JIT compilation & optimization to PHP
• Both open sourced on github
31
Thanks!
Questions?

similar documents