PHP Performance: Principles and tools

PHP Performance
Principles and Tools
By Kevin Schroeder
Technology Evangelist
Zend Technologies
I have stickers!

About Kevin
Past: Programming/Sys Admin
Current: Technology Evangelist/Author/Composer
@kpschrade

Follow us!
Zend Technologies
http://twitter.com/zend
http://twitter.com/kpschrade (me!)

Join us at ZendConThe premier PHP conference!October 17-19, 2011 – Santa Clara, CA
Conference Themes
Cloud Computing
Learn about the latest developments in PHP Cloud infrastructure, management and application services
Mobile and User Experience
Learn how to build engaging mobile apps with the latest PHP technologies and tools
Enterprise and Professional PHPExplore PHP best practices, new technologies and practical tips with industry experts
Conference Highlights
Sessions focused on how to best develop and deploy PHP
Sessions designed for all knowledge levels
Intensive tutorials for accelerated learning
PHP Certification crash courses and testing
Exhibit hall showcasing the latest products
Special networking opportunities during meals and events
www.zendcon.com

Quiz
Who here is a sysadmin?
Who here is a developer?
Who here is both?

WARNING!
Performance is a HIGHLY context sensitive topic
There are no answers here; only possible directions
Do not micro-optimize
20ms cf. 100ms is not a 5 fold improvement in performance, it is a 5 fold waste of your time
Unless you’re Yahoo, Google, Facebook, etc.
The perception of performance is as important as actual performance
Use the scientific method to find performance problems
Observe
Hypothesize
Experiment
Solve one problem at a time and MEASURE RESULTS!

General categories of performance problems

Performance – Database
Symptoms
Pages take a long time to load with Apache/IIS/PHP using proportionally less of the CPU time compared to DB
Causes
The list could go on for 4.23x105slides
There are so many ways you can screw up your queries
Tools
Intra-application query monitoring, usually via a DB adapter abstraction layer
Zend Server – A distinct group of monitoring options for DB calls in PHP

Performance - Database
Possible remedies
Simplify your queries
Make your queries more complex
Tune your database
Don’t just automatically allocate more memory
Don’t just create indexes
Keeping this light because there are other, better, MySQL performance experts (many of them here)

Performance – Network Latency
Symptoms
Depends on where the latency is occurring
Often slow page responses with low CPU usage
Causes
Network Saturation (probably not too likely)
DNS Reverse Lookups
TCP Handshakes
High number of “hops”
TCP Backlog
… and much more
Tools
Wireshark
Tcpdump
Zend Server – monitor slow socket function execution

The TCP handshake
T/TCP SYN + Response Data + FIN
T/TCP SYN + Response Data + ACK +FIN
ACK

Performance – IO Contention
Symptoms
Unpredictable and difficult to reproduce page load times, but often corresponds to request frequency
Causes
Too few spindles serving too much data
High write concurrency
Insufficient free system RAM
Insufficient depth of data directory structure
Tools
Zend Server (in extreme cases, long fopen() calls)
vmstat & top (Linux)
System Internals Process Explorer (Windows)

Performance – IO Contention
Remedies
Increase the number of disks
Favor directory depth over width for storing raw data files
Increase RAM so read operations can be pulled from memory

Performance – CPU Utilization
Symptoms
Page load times either correspond directly to CPU utilization or are consistently slow
Causes
Badcode or over-architected solutions
Excessive recursion
Excessive IO
High server utilization
too few servers serving too much traffic
Tools
top & vmstat (Linux)
System Internals Process Explorer (Windows)

Performance – CPU Utilization
There are often 2 different paths this will usually follow
User time
Time spent executing the code that you wrote
System time
Time spent running functionality that your code calls, such as networking, file system, etc.

Performance – Network Connectivity
Symptoms
Database or web service calls take long to execute
Causes
If read functions die unexpectedly, bad network hardware might be a cause
Slow DNS queries
Full TCP backlog
Tools
Zend Server – monitor failed calls to socket_connect(), fsockopen() or fread()
Wireshark
Tcpdump

OO
Beware of magic!
__get, __set, __call
Situations where they are often used
When representing another object that it has no knowledge of - I.e. Soap requests, COM objects, Java objects, database tables
When there are property-level permissions on an object’s properties. Keeping the properties private and using __get/set to determine access
If used sparingly they won’t affect you, but many modern applications have lots of recursion or large underlying architectures that can have unpredicted consequences

!Preg
Use preg_match for data checking only when you have to
stripos(‘http://’, $website) is over twice as fast as preg_match(‘/http:///i’, $website)
ctype_alnum() is almost 5 times as fast as preg_match(‘/^s*$/’);
Casting “if ($test == (int)$test)” is well over 5 times faster than preg_match(‘/^d*$/)
Preg will probably not bite you unless you have a lot of recursive logic that depends on it

OS tools to use
Devs are really bad about thinking about the OS
Use OS tooling!!!
vmstat
top
tcpdump & wireshark
strace
Try stracehttpd –X at least once
lsof
You can map file descriptors found in strace to descriptors in lsof

Zend Server
Monitors for several events like slow requests, slow DB calls, high memory usage, etc.
Provides the context for those requests
Can be configured to provide profile-like data from production installations
Has a built in queuing mechanism that allows you to offload non-immediate processing (which can actually be quite a lot)

Do you queue?
Offloading long running processes to a queuing system can help make your front end servers scale more vertically
Examples include
Job Queues
Zend Server 5 (http://www.eschrade.com/ tag: job-queue)
Gearman
Beanstalk
Message Queues
ActiveMQ
RabbitMQ

TODO || ! TODO

What not to do
This is not complete!!!
“Oh we’ll just cache it”
“Our NFS server is just fine for serving PHP files”
Use a frigging deployment mechanism
“That LIKE query is necessary”
You may need to structure your data a little different
Don’t make your DB gues
You might need to use a search engine instead
Zend_Search_Lucene
SOLR
“Using arrays is faster than objects, we’ll use them instead”
“Monitoring/diagnostic tools are too expensive”

List of things to do
This is not complete!!!
THINK before you code!!
Minimize require_once calls/use an autoloader
Try to make your architecture more horizontal rather than vertical
People from Java backgrounds tend to go more vertical
There is often a correlation between long stack traces and poor performance
Use Lazy Loading/Load on Demand
SPL arrays and such are great for this

Use diagnostic tools!
Zend Server Monitoring and Code Tracing
Some kind of profiler
Utilize compiled PHP functionality for iterative complex functionality.
Use Ajax
Allows you to make a more responsive page by having fast content available immediately and slow content available as it’s ready

With multiple service calls, consider batching multiple NB-IO requests
Facebook does this, so it must be cool
Synchronous requests – 4.2 seconds
Asynchronous requests using curl – 1.5 seconds

Questions?

To contact me after my presentation, text XIK to INTRO (46876)
Follow us!
Zend Technologies
http://twitter.com/zend
http://twitter.com/kpschrade (me!)

Join us at ZendConThe premier PHP conference!October 17-19, 2011 – Santa Clara, CA
Conference Themes
Cloud Computing
Learn about the latest developments in PHP Cloud infrastructure, management and application services
Mobile and User Experience
Learn how to build engaging mobile apps with the latest PHP technologies and tools
Enterprise and Professional PHPExplore PHP best practices, new technologies and practical tips with industry experts
Conference Highlights
Sessions focused on how to best develop and deploy PHP
Sessions designed for all knowledge levels
Intensive tutorials for accelerated learning
PHP Certification crash courses and testing
Exhibit hall showcasing the latest products
Special networking opportunities during meals and events
www.zendcon.com

Similarly, be careful when dealing with lots of Ajax data
Developers running test data often under-estimate how much data their code will send back
The browser needs to process that data
With large data sets the browser may take a while to update the DOM
2 options
Use staggered loading
Pass rendered content to the browser and attach with innerHTML

Be aware of your page size
Oracle – 104kb – 861ms
Zend – 183kb – 2.4 seconds
Microsoft – 238kb – 4.4 seconds
IBM – 640 kb – 4.6 seconds
There is a relatively direct relationship between the amount of data on your web page and the render time

http://lanyrd.com	32
http://localhost	8
http://coderwall.com	1

PHP Performance: Principles and tools

by Kevin Schroeder

Accessibility

Categories

Upload Details

Usage Rights

3 Embeds 41

Statistics

PHP Performance: Principles and tools Presentation Transcript