Req #53866 [Asn]: Zend engine's hashtable performance tweaks
Edit report at https://bugs.php.net/bug.php?id=53866&edit=1
ID: 53866
Updated by: kalle@php.net
Reported by: marcin dot babij at nasza-klasa dot pl
Summary: Zend engine's hashtable performance tweaks
Status: Assigned
Type: Feature/Change Request
Package: Scripting Engine problem
PHP Version: trunk-SVN-2011-01-28 (SVN)
-Assigned To: dmitry
+Assigned To: nikic
Block user comment: N
Private report: N
New Comment:
@nikic, can you confirm/close the report if this is no longer relevant with ZE3 please
Previous Comments:
------------------------------------------------------------------------
[2012-08-10 10:14:03] reeze dot xia at gmail dot com
This seems dead. no response for a long time :(
------------------------------------------------------------------------
[2011-01-28 13:27:48] marcin dot babij at nasza-klasa dot pl
Description:
------------
What was done:
- Hash function in zend_hash.h was rebuilt and became much faster, without losing the most important
properties.
- Hashtable implementation was changed from Simple chaining to Open addressing with linear probing,
but with linked bucket, not included in hash array, which causes:
-- Bucket structure to lose 2 pointers.
-- Searching works similar, but don't have to jump with pointers stored in different memory
locations, inserting, deleting and rehashing don't need to update linked list, but must search
for first empty bucket, which is fast, because it scans continuous memory.
-- Load factor decreases from 1.0 to 0.5-0.75 to make less collisions and faster hashtable, which in
turn increases memory footprint a little.
- Open addressing doesn't change significantly performance, but next thing was to create new
array (arEmpty), which is of size nTableSize bytes, which keeps track of used/empty buckets and
makes inserting and rehashing much faster. In future it can be tested as bit-array with size of
nTableSize/8 bytes.
- More macros were added to replace repetitive constructs.
- New constants were added to allow:
-- Creating new hashtables of size at least X (where 4 and 8 are reasonable), which makes no
rehashing and reallocing memory while changing size to 2 and then to 4.
-- For small tables it's better to extend them by a factor of 4 times, not 2, to make rehashing
cost smaller for most hashtables, of cost of little higher memory consumption.
-- For large tables it's better to have other load factor, closer to 1, while for small tables
it's better to use load factor closer to 0.5.
What should be done:
- http://lxr.php.net/xref/PHP_TRUNK/ext/standard/html_tables/html_table_gen.php#722
should be changed and html_tables.h regenerated, but this will need to rewrite hashtable engine from
C to PHP
- APC should be fixed
What can be done:
- Make new constants configurable by php.ini.
- Test if changing arEmpty from byte-array to bit-array helps on performance.
- Tweak default constants' values using some real-live benchmarks.
- Prove (or modify and prove) hash function to have property, that it has no collisions if two keys
don't differ on no more than 6 bytes, which will lead to memcmp omit first (or last) 6 bytes of
key. Also simpler thing may be proven, that is it has no collisions if two keys are not longer than
6 bytes, which will make most string keys omit memcpy at all.
------------------------------------------------------------------------
--
Edit this bug report at https://bugs.php.net/bug.php?id=53866&edit=1
Thread (5 messages)