SlideShare a Scribd company logo

DrupalCampLA 2011: Drupal backend-performance

•
•
A
Ashok Modi

The DrupalCampLA 2011 presentation on backend performance. The slides go over optimizations that can be done through the LAMP (or now VAN LAMMP stack for even more performance) to get everything up and running.

1 of 49
Download to read offline
DRUPAL BACKEND PERFORMANCE AND SCALABILITY Ashok Modi – DrupalCampLA 2011
About me   Systems Analyst at California Institute of the Arts.   Working with Drupal since 2006.  View my profile at http://drupal.org/user/60422  View my thoughts at http://btmash.com   Strong interest in server optimization.
About presentation   Will reference some of the sites I’ve worked on   Main CalArts site with (too many) modules kept under control.   CalArts (Photo Site with about 150k pieces of content)   Zimmer Twins (1M content, ½ M users)   There is a lot of material to go through.   Avoid talking about software that has successor.   May have to speed through some areas (or are you ok with staying around for longer?)   Doesn’t really apply to shared hosting   Only a part of this presentation (code related optimizations?) might apply.   Have a question? Ask!   Have something to share? Come on up!   I am certain there are some very knowledgeable folk in the room…
Resources   Khalid Baheyeldin  http://2bits.com   Drupal.org infrastructure group  http://groups.drupal.org/high-performance   Peter Zaitsev  http://www.mysqlperformanceblog.com/   Lullabot  http://lullabot.com/ideas/blog   Community 
Goals   Define your objectives and goals first   Do you want a faster response to the end user per page?   Do you want to handle more page views?   Do you want to minimize downtime?   Each is related…but different   Gets harder and harder to achieve more performance   More infrastructure   Patching / Hacking Drupal   Revisions to architecture (Fields, Views)
Diagnosis   Proper diagnosis is essential before proposing and implementing a solution.   Based on proper data.   Analysis of collected data.  Few possible paths of optimization.
Validation   Avoid the ‘wild goose chase’.   Validate results on a test server.   Replicate the data on a development server.   Backup and migrate will help.   Migrate can also help.   Recreate the site.   Gather a time difference between test and production server.   Measure again and see if relative times remain the same.
Points of optimization   Introduction.   Tools to measure and diagnose issues.   Speed optimizations.
Introduction – Hardware   Physical server matters   Dedicated   VPS   Cloud-based   Anyone here to argue it doesn’t work?   Multiple cores are the norm   32 > 16 > 8 > 4 > 2 > 1   Lots of RAM (caching the file system and db as much as possible)   Multiple disks (split of the server to various disks and/or servers)   SSD is much faster than a reg. HDD.   Look into https://github.com/facebook/flashcache   FusionIO also looks very promising.   Tuning DB on SSD is different from tuning DB on HDD
LAMP Stack   Traditionally most commonly used stack for hosting Drupal and similar applications.   Linux   Apache   MySQL   PHP   A chunk of the presentation will focus on the above.   Though we need an acronym involving V(arnish), A(pachesolr), M(emcache), M(ongo), N(ginX)   VLAMNMP? Or VAN LAMMP?   Not discussing Windows.   Anyone host Drupal on Windows Server? Would like to know more.
Multiple Servers   Master DB Server, multiple web servers.   Use a load balancer (or something like HAProxy or some other DNS round robin)   Set up slave database servers for SELECT queries.   Do it only if you have the budget and resources.   Complexity is expensive (running cost, maintenance time)   Tuning a system can avoid or delay a split.   A site by 2bits runs on one server.   Handles3+ million hits per day.   Read more at http://goo.gl/XueVY
Testing tools   Apache Benchmark (DEMO)   ab –n 100 –c 10 http://www.example.com   ab –n 10 –c 10 –C PHPSESSID=<sessid> http://www.example.com   Do 10 concurrent requests for up to 100 requests.   Average response time per second.   How many requests handled per second.   Jmeter   Similar to Apache Benchmark.   Can natively use on Windows.   Can test POST functionality.   LoadStorm   Web service to test site.   Will give pretty graphs in real time.
Console Monitoring tools   Top  Real time monitoring.  Load average.  CPU utilization.  Memory usage.  List of processes.   htop  Similar to top but for multiple cores.  Faster.
Console Monitoring tools (cont’d)   atop  Shows network statistics.  Runs a collection daemon in the background.   vmstat  Report memory statistics   netstat  Shows active network connections  netstat –an  netstat –an | grep EST
Graphical Monitoring tools Cacti    http://www.cacti.net  Available as a package on Ubuntu, Debian (various other *nix/bsd flavors)  Easy to understand graphs.  Displays history over day, week, month, year.  Graphs available to display stats for CPU, memory, network, Apache, MySQL  Many others written by others available online. Munin    http://munin.projects.linpro.no  Very similar to Cacti (doesn’t require a db)  Can create own monitoring scripts. Nagios    Heard its very powerful (alerts by email, sms, etc).  Drupal module for integration. Panopta and New Relic offer hosted monitoring  
Linux / BSD   Use a proven stable distribution (Debian Stable, Ubuntu LTS, RHEL, CentOS)   Use recent versions   *Use whatever distro your staff has expertise with   Try to avoid bloat   Don’t install PostGreSQL if you are only using MySQL, no desktop server, java, etc.   Balance compiling own programs versus using packages.   Old versions of GD on older versions of Debian, Ubuntu.   Compiling provides full control.   Can be a pain to upgrade.
Apache   Most popular, supported, and feature rich.   Stable.   Can also be enabled with too many unnecessary modules.   mod_proxy, cgi_module, etc may be unnecessary.   Smaller process = more users can access site   apachectl -M – Display all enabled apache modules.   apachetop   Reads and analyses Apache access logs.   Good to detect crawlers.   ‘apachetop –f /path/to/site/access.log
Apache Optimizations   MaxClients (prevent swapping / thrashing)   Too low – cannot serve enough clients   Too high – you run out of memory and you start swapping. Server dies and you cannot serve any clients.   MaxRequestsPerChild   Tune it to terminate the process faster and free up memory.   KeepAlive   Keep it enabled.   New connects will not get opened constantly.   mod_gzip/deflate   Serve more content quickly.
NginX   http://nginx.net   Quite stable.   More lightweight than Apache.   Reasonably easy to set up.   http://wiki.nginx.org/Drupal for drupal settings    Seeing some promising results in our dev environment so far…   Run PHP as FastCGI process.   Canalso do this with Apache (current method for DrupalCampLA website).   Also uses less memory.
Varnish  HTTP accelerator  Set it up as a reverse proxy to send the call to Apache if it cannot server something itself  Serve anonymous page requests and static files  D6 core will not get served anonymous – use pressflow.  D7 and varnish play nicely.  Requires some tuning  Used on Drupal.org  RiotGames uses it (@dragonwize might be able to say some more if he’s in the room?)  Serve millions of pages of content with little impact on server.  Look at http://drupal.org/project/varnish  http://goo.gl/8l7gI has some configuration info.  http://goo.gl/9xQDz has a better explanation 
Varnish (cont’d)   Define IP/port in backend <name> for each web server.   Define multiple backends for different servers   backend b1 {.host=“127.0.0.1”; .port=“8082”}   Use director to group backends for round robin.   return (pass) // do not cache any checks you make.   If (req.url ~ "^.*/ajax/.*$") { return (pass); }   return (lookup) // lookup in cache or pass it on to backend to get cached.   unset beresp.http.Set-Cookie; // Unset cookies; What actually allows caching.   Lullabot’s article: http://goo.gl/7JFrP   Basic setup for D7: http://goo.gl/l7601   Tested on own blog…handled about 3k requests per second (couldn’t figure out way to break it). Previously handled 30 – 50 requests per second.
MySQL   Most popular database for Drupal.   Not necessarily the best database but still does a good job.   Easy to set up, lots to tune.   Less to tune in D6 but D7 requires tuning, even for small sites.   Various pluggable engines (InnoDB, Archive)   Forks   MariaDB   Percona   Drizzle   MySQL 5.5 is a big difference.   More to tune.   http://goo.gl/hU8tW
MySQL Monitoring   mtop / mytop   Like top but for MySQL   Real time monitoring (no history)   Shows slow queries and locks.   If you have neither – SHOW FULL PROCESSLIST   Mysqlreport   http://hackmysql.com/mysqlreport   Reports on server – no recommendations (documentation on site explains a lot)   Slow Query Log   Can be enabled in you’re my.cnf file   List queries more than N seconds   List queries without indexes.   Helped identify bottlenecks (one involved a bad count(*) query which I removed)   mysql_slow_log_parser script (http://goo.gl/4ZHCT)
MySQL Engines   MyISAM   Fast reads   Less overhead   Poor concurrency   InnoDB   Transactional   Slower in some cases (SELECT COUNT(…))   Lots of settings to analyze and change   Better concurrency.   Forks   Percona comes with XtraDB (replacement for InnoDB).   Maria comes with XtraDB.   Both currently looking to be better options than InnoDB (use Google Patches).   Same tuning settings.
MySQL Tuning  There are many (many) settings that could be tuned.  Talk about the ones most likely to give the largest benefits. (D7 Focused)  innodb_buffer_pool_size  Very important.  Set up to 80% of memory allocated for DB to this.  If db is small, memory could be used elsewhere. innodb_log_file_size    Important for sites with large write workloads  64 – 512M innodb_flush_log_at_trx_commit    By default, each update transaction flushes the log which is expensive.  Set to 0 (write to buffer but no flushes on transaction) or 2 (flush cache instead of disk; both still flush disk/sec) so flushes happen to OS cache.  Will lose 1-2 seconds of data if set to 0 if OS crashes. Will have data loss only with full OS crash if set to 2.
MySQL Tuning (cont’d)   table_cache   Opening tables can be expensive.   Keeps tables open in cache.   1024 is good place to start.   thread_cache   If you have a lot of quick connections, increase the value.   query_cache_size   Will cache query results.   Generally 32M – 512M   Use mysqlreport to get an idea of what settings to tune.   Use mysqltuner to help guide you in right direction.   http://mysqltuner.pl   Read http://mysqlperformanceblog.com
MySQL Replication   Used on Drupal.org   INSERT/DELETE/UPDATE queries go to master   SELECT queries go to slave(s)   Provide noticeable improvements.   D7 supports replication.   For D6, Pressflow is best bet.   Beware of complexity (connection bet. Master/slave goes down, bad things happen).   Did extensive tuning on Zimmer Twins   Noticeableimprovement despite lack of querying slave server.   Removed slave server for good.
MongoDB  Public release in 2009  Document-oriented  ‘No-SQL’  b.collection.insert|add|update({parameters});  Retrieve subsets (for certain fields / objects)  Manages collections of objects in JSON-like format.  { "username" : "bob", "address" : { "street" : "123 Main Street", "city" : "Springfield", "state" : "NY" } }  Currently supports up to 64 indexes.  1 for ascending order, -1 for descending order.  b.collection.ensureIndex({username: 1, address.state: 1});  Nested fields can also be indexed.  Supports master-slave replication. Has automated database sharding.  Easily create a cluster.
MongoDB (cont’d)   Module for Drupal exists!   http://drupal.org/project/mongodb   Cache, Field Storage, Blocks, Queues, Sessions, Watchdog currently supported.   Does a lot of the heavy lifting.   Very fast (Examiner.com uses it and has disabled page caching despite high load).   For anything exported into MongoDB, previous SQL queries will need to be modified so they become mongo queries.   For entities, use entityfieldquery   Anything in this format will actually allow you to switch between SQL and Mongo (and anything else) without changing code.   Look at http://drupal.org/project/efq_views for promising work into even more flexible views 
PHP   Use a recent, stable release.   D7 requires 5.2.x, as do a few 6.x contributed modules.   D8 will require 5.3.x (yes, it’s a ways away ).   Install an Op-code cacher / accelerator.   Useful in bringing down memory usage for a site.   eAccelerator   APC   Xcache   Zend optimizer (commercial)   Anyone try HipHop?
Running PHP   mod_php   Standard php module used by apache.   Few issues.   Well tested and supported.   Can be a resource hog.   FastCGI (PHP-FPM)   Can be used with NginX, Apache.   Runs as a separate process.   More stable with lower memory usage than mod_php.   Trickier to install (lots of good doc. online).
Debugging PHP   Xdebug  http://xdebug.org  Display traces on error conditions.  Trace functions.  Profile PHP Scripts  Lots of docs online for installing.   kCacheGrind  Provides a visualization on bottlenecks in code.
Op-code Caching   Benefits   Lowered memory usage.   Significant decrease in CPU utilization.   Usage on http://www.zimmertwins.com lowered memory usage per process from 20M down to less than 4M   Usage on http://calarts.edu lowered memory usage per process from 45M down to less than 10M.   Drawbacks   May crash   May require restarts after updating code (apc.stat = 0)
Op-code Caching (cont’d)   Op-code caching will not work in all circumstances  Network connections.  Sorting arrays.  DB queries   Bad modules are bad.
Drupal   Database intensive.   Can be a resource hog.   Memory intensive.   D7 > D6 > D5   Site may not be affected by bottleneck.   Quick Tips   Disable unnecessary modules.   Ifperformance is such a concern, sites *can* be made without Fields / CCK (ZimmerTwins was such a site)   Views UI, Field UI, Rules UI, <module> UI on production.   Make sure cron runs regularly.
Drupal Tools   Devel   http://drupal.org/project/devel   Totalpage execution.   Query execution time.   Query log.   Memory utilization.   Can be combined with Stress Testing.   Trace   http://drupal.org/project/trace   Use for debugging.   Traces output, invocations, warnings.   Filter by Query Type.
Module calls over network   Email users (og)   Call web widgets / APIs (youtube, twitter, facebook)   Cache as much data as possible.  Usehelper modules to aid with reducing bog.  Queues.  SMTP Mail module.
Drupal Caching   Helpful in not querying / processing same bits of content over and over.   Especially for anonymous users all of whom may be viewing the same content on your site.   Many caches in core.   Bootstrap   Block   Field   Filter   Form   Image   Menu   Page (only for anonymous)   Many from contrib modules (like views, rules, media)
Useful contrib caching modules EntityCache    http://drupal.org/project/entitycache  Caches all core entities (node, user, taxonomy) on entity_load()  Stays in cache until expiry or until content is updated/deleted. Boost    http://drupal.org/project/boost  Creates HTML for pages and stores it in files.  Requires changes to .htaccess file  Does not load up Drupal once content is cached to file for anon. users.  Can also use module to display site while in maintenance mode.  Varnish has mostly replaced this module (though they could play with each other) on sites not in shared hosting.  Views content cache  Block cache alter  Performance hacks
Pluggable caching   Use $conf variable in settings.php   $conf['cache_backends'][]= ‘./path/to/cache_first_mechanism.inc’;   $conf['cache_backends'][]= ‘./path/to/cache_second_mechanism.inc’;   $conf[‘cache_class_<bin>’] = ‘CACHECLASS’;   $conf[‘cache_default_class’] = ‘SECONDCACHECLASS’;   Allows you to use a custom caching module.   Can even use this to completely disable caching (DO NOT USE ON A PRODUCTION SITE!)   Contrib modules   Cache Router (D6)   APC (D7)   Very fast.   Limited to caching on one web server (cache cannot get reused over multiple servers – can be good or bad.)
Memcached   Distributed object caching in memory.   Written by danga for LiveJournal   Lives in memory   Can span multiple servers.   Seamless for D6, D7   D7 is still undergoing some changes.   Hash own requirements (apache needs to be restarted, have to clear old caches)   Slower than APC, but scalable.   Takes load off DB server (yay!)
Search Mechanism   Drupal core search   Need I say anything?   Search API looks to be a much more promising (not to mention flexible!) option.   Pluggable system to support various types of backends.   Search API MongoDB, Search API Xapian   Supports Views.   Proposal to include in D8 (core conversation at DC London).   LuceneAPI   Not as fast as ApacheSolr.   Easy to setup.   Google CSE might also be a good fit.
ApacheSolr   ApacheSolr   Very fast.   Easy to configure on *nix systems.   Requires a server on which Java can be installed.   More and more companies offering Solr as a Service (take load off your systems altogether).   Available as Drupal Module.   Apachesolr (very mature)   Search API apachesolr (very promising)   Views Plugin so even drive non-search related pages using Solr!
Other options   Using an optimized distribution   Pressflow (for D6)   http://fourkitchens.com/pressflow-makes-drupal-scale   Only supports MySQL   Cleanly supports reverse proxies such as Varnish   Optimized for PHP5   Pressflow for D7 is currently identical to D7   Talksabout abandoning MySQL in favor of a MongoDB/ Cassandra DB architecture.   Many of the improvements made for Pressflow are in D7 core.   Keep in mind that a faster Drupal Core won’t save you from contrib modules behaving badly.
Other options (cont’d)   Patching Drupal / Contrib  ‘Hack core’  Need to know what you are doing.  Sometimes necessary.  Create a patches directory where all the changes you make for a core/contrib file can be tracked and easily applied on updates.  Create own module and do any necessary schema updates / alters from there.
Past experiences from Drupal Core   User login on zimmertwins.com was painfully slow (5+ seconds per user)   Gist of problem: DB not using index on username due to lower()   Bug had been around since 2006.   Solution: Modified patch on D.O for site.   User login time down to 0.1 seconds.   Pressflow avoids this by not allowing case-insensitve login.   Comments did not have index on user ID   Created index on user id as an update from my module.   If added in future, can remove my version of index.   Loading for comment by user no longer an issue.
Advice for developers   Take advantage of caching.   Use memory wisely.   Unsetthe variable if you don’t have a need for it later.   Save variable to memory for future use so processing isn’t done multiple times (see drupal_static()).   Take advantage of AHAH functionality   Fewer queries.   Not reloading the page.   Saving bandwidth.   Learn to use jQuery (same as above).   Pay close attention to what
Possibly related sessions   Note: Some have passed (but check out their screencasts)   Drupal on the Cloud   Patrick Wall   Drupal Development Q&A   Drupal 7 Q&A   Building APIs   Adam Gregory   Professional Staging and Deployment   Christefano   Understanding Ctools   Helior Colorado
Questions   Have a question?   Want to talk more about performance?  Let’s talk after    Think you can help with Drupal performance?  http://goo.gl/I3PN2   Thank you 

More Related Content

DrupalCampLA 2011: Drupal backend-performance

  • 2. About me   Systems Analyst at California Institute of the Arts.   Working with Drupal since 2006.  View my profile at http://drupal.org/user/60422  View my thoughts at http://btmash.com   Strong interest in server optimization.
  • 3. About presentation   Will reference some of the sites I’ve worked on   Main CalArts site with (too many) modules kept under control.   CalArts (Photo Site with about 150k pieces of content)   Zimmer Twins (1M content, ½ M users)   There is a lot of material to go through.   Avoid talking about software that has successor.   May have to speed through some areas (or are you ok with staying around for longer?)   Doesn’t really apply to shared hosting   Only a part of this presentation (code related optimizations?) might apply.   Have a question? Ask!   Have something to share? Come on up!   I am certain there are some very knowledgeable folk in the room…
  • 4. Resources   Khalid Baheyeldin  http://2bits.com   Drupal.org infrastructure group  http://groups.drupal.org/high-performance   Peter Zaitsev  http://www.mysqlperformanceblog.com/   Lullabot  http://lullabot.com/ideas/blog   Community 
  • 5. Goals   Define your objectives and goals first   Do you want a faster response to the end user per page?   Do you want to handle more page views?   Do you want to minimize downtime?   Each is related…but different   Gets harder and harder to achieve more performance   More infrastructure   Patching / Hacking Drupal   Revisions to architecture (Fields, Views)
  • 6. Diagnosis   Proper diagnosis is essential before proposing and implementing a solution.   Based on proper data.   Analysis of collected data.  Few possible paths of optimization.
  • 7. Validation   Avoid the ‘wild goose chase’.   Validate results on a test server.   Replicate the data on a development server.   Backup and migrate will help.   Migrate can also help.   Recreate the site.   Gather a time difference between test and production server.   Measure again and see if relative times remain the same.
  • 8. Points of optimization   Introduction.   Tools to measure and diagnose issues.   Speed optimizations.
  • 9. Introduction – Hardware   Physical server matters   Dedicated   VPS   Cloud-based   Anyone here to argue it doesn’t work?   Multiple cores are the norm   32 > 16 > 8 > 4 > 2 > 1   Lots of RAM (caching the file system and db as much as possible)   Multiple disks (split of the server to various disks and/or servers)   SSD is much faster than a reg. HDD.   Look into https://github.com/facebook/flashcache   FusionIO also looks very promising.   Tuning DB on SSD is different from tuning DB on HDD
  • 10. LAMP Stack   Traditionally most commonly used stack for hosting Drupal and similar applications.   Linux   Apache   MySQL   PHP   A chunk of the presentation will focus on the above.   Though we need an acronym involving V(arnish), A(pachesolr), M(emcache), M(ongo), N(ginX)   VLAMNMP? Or VAN LAMMP?   Not discussing Windows.   Anyone host Drupal on Windows Server? Would like to know more.
  • 11. Multiple Servers   Master DB Server, multiple web servers.   Use a load balancer (or something like HAProxy or some other DNS round robin)   Set up slave database servers for SELECT queries.   Do it only if you have the budget and resources.   Complexity is expensive (running cost, maintenance time)   Tuning a system can avoid or delay a split.   A site by 2bits runs on one server.   Handles3+ million hits per day.   Read more at http://goo.gl/XueVY
  • 12. Testing tools   Apache Benchmark (DEMO)   ab –n 100 –c 10 http://www.example.com   ab –n 10 –c 10 –C PHPSESSID=<sessid> http://www.example.com   Do 10 concurrent requests for up to 100 requests.   Average response time per second.   How many requests handled per second.   Jmeter   Similar to Apache Benchmark.   Can natively use on Windows.   Can test POST functionality.   LoadStorm   Web service to test site.   Will give pretty graphs in real time.
  • 13. Console Monitoring tools   Top  Real time monitoring.  Load average.  CPU utilization.  Memory usage.  List of processes.   htop  Similar to top but for multiple cores.  Faster.
  • 14. Console Monitoring tools (cont’d)   atop  Shows network statistics.  Runs a collection daemon in the background.   vmstat  Report memory statistics   netstat  Shows active network connections  netstat –an  netstat –an | grep EST
  • 15. Graphical Monitoring tools Cacti    http://www.cacti.net  Available as a package on Ubuntu, Debian (various other *nix/bsd flavors)  Easy to understand graphs.  Displays history over day, week, month, year.  Graphs available to display stats for CPU, memory, network, Apache, MySQL  Many others written by others available online. Munin    http://munin.projects.linpro.no  Very similar to Cacti (doesn’t require a db)  Can create own monitoring scripts. Nagios    Heard its very powerful (alerts by email, sms, etc).  Drupal module for integration. Panopta and New Relic offer hosted monitoring  
  • 16. Linux / BSD   Use a proven stable distribution (Debian Stable, Ubuntu LTS, RHEL, CentOS)   Use recent versions   *Use whatever distro your staff has expertise with   Try to avoid bloat   Don’t install PostGreSQL if you are only using MySQL, no desktop server, java, etc.   Balance compiling own programs versus using packages.   Old versions of GD on older versions of Debian, Ubuntu.   Compiling provides full control.   Can be a pain to upgrade.
  • 17. Apache   Most popular, supported, and feature rich.   Stable.   Can also be enabled with too many unnecessary modules.   mod_proxy, cgi_module, etc may be unnecessary.   Smaller process = more users can access site   apachectl -M – Display all enabled apache modules.   apachetop   Reads and analyses Apache access logs.   Good to detect crawlers.   ‘apachetop –f /path/to/site/access.log
  • 18. Apache Optimizations   MaxClients (prevent swapping / thrashing)   Too low – cannot serve enough clients   Too high – you run out of memory and you start swapping. Server dies and you cannot serve any clients.   MaxRequestsPerChild   Tune it to terminate the process faster and free up memory.   KeepAlive   Keep it enabled.   New connects will not get opened constantly.   mod_gzip/deflate   Serve more content quickly.
  • 19. NginX   http://nginx.net   Quite stable.   More lightweight than Apache.   Reasonably easy to set up.   http://wiki.nginx.org/Drupal for drupal settings    Seeing some promising results in our dev environment so far…   Run PHP as FastCGI process.   Canalso do this with Apache (current method for DrupalCampLA website).   Also uses less memory.
  • 20. Varnish  HTTP accelerator  Set it up as a reverse proxy to send the call to Apache if it cannot server something itself  Serve anonymous page requests and static files  D6 core will not get served anonymous – use pressflow.  D7 and varnish play nicely.  Requires some tuning  Used on Drupal.org  RiotGames uses it (@dragonwize might be able to say some more if he’s in the room?)  Serve millions of pages of content with little impact on server.  Look at http://drupal.org/project/varnish  http://goo.gl/8l7gI has some configuration info.  http://goo.gl/9xQDz has a better explanation 
  • 21. Varnish (cont’d)   Define IP/port in backend <name> for each web server.   Define multiple backends for different servers   backend b1 {.host=“127.0.0.1”; .port=“8082”}   Use director to group backends for round robin.   return (pass) // do not cache any checks you make.   If (req.url ~ "^.*/ajax/.*$") { return (pass); }   return (lookup) // lookup in cache or pass it on to backend to get cached.   unset beresp.http.Set-Cookie; // Unset cookies; What actually allows caching.   Lullabot’s article: http://goo.gl/7JFrP   Basic setup for D7: http://goo.gl/l7601   Tested on own blog…handled about 3k requests per second (couldn’t figure out way to break it). Previously handled 30 – 50 requests per second.
  • 22. MySQL   Most popular database for Drupal.   Not necessarily the best database but still does a good job.   Easy to set up, lots to tune.   Less to tune in D6 but D7 requires tuning, even for small sites.   Various pluggable engines (InnoDB, Archive)   Forks   MariaDB   Percona   Drizzle   MySQL 5.5 is a big difference.   More to tune.   http://goo.gl/hU8tW
  • 23. MySQL Monitoring   mtop / mytop   Like top but for MySQL   Real time monitoring (no history)   Shows slow queries and locks.   If you have neither – SHOW FULL PROCESSLIST   Mysqlreport   http://hackmysql.com/mysqlreport   Reports on server – no recommendations (documentation on site explains a lot)   Slow Query Log   Can be enabled in you’re my.cnf file   List queries more than N seconds   List queries without indexes.   Helped identify bottlenecks (one involved a bad count(*) query which I removed)   mysql_slow_log_parser script (http://goo.gl/4ZHCT)
  • 24. MySQL Engines   MyISAM   Fast reads   Less overhead   Poor concurrency   InnoDB   Transactional   Slower in some cases (SELECT COUNT(…))   Lots of settings to analyze and change   Better concurrency.   Forks   Percona comes with XtraDB (replacement for InnoDB).   Maria comes with XtraDB.   Both currently looking to be better options than InnoDB (use Google Patches).   Same tuning settings.
  • 25. MySQL Tuning  There are many (many) settings that could be tuned.  Talk about the ones most likely to give the largest benefits. (D7 Focused)  innodb_buffer_pool_size  Very important.  Set up to 80% of memory allocated for DB to this.  If db is small, memory could be used elsewhere. innodb_log_file_size    Important for sites with large write workloads  64 – 512M innodb_flush_log_at_trx_commit    By default, each update transaction flushes the log which is expensive.  Set to 0 (write to buffer but no flushes on transaction) or 2 (flush cache instead of disk; both still flush disk/sec) so flushes happen to OS cache.  Will lose 1-2 seconds of data if set to 0 if OS crashes. Will have data loss only with full OS crash if set to 2.
  • 26. MySQL Tuning (cont’d)   table_cache   Opening tables can be expensive.   Keeps tables open in cache.   1024 is good place to start.   thread_cache   If you have a lot of quick connections, increase the value.   query_cache_size   Will cache query results.   Generally 32M – 512M   Use mysqlreport to get an idea of what settings to tune.   Use mysqltuner to help guide you in right direction.   http://mysqltuner.pl   Read http://mysqlperformanceblog.com
  • 27. MySQL Replication   Used on Drupal.org   INSERT/DELETE/UPDATE queries go to master   SELECT queries go to slave(s)   Provide noticeable improvements.   D7 supports replication.   For D6, Pressflow is best bet.   Beware of complexity (connection bet. Master/slave goes down, bad things happen).   Did extensive tuning on Zimmer Twins   Noticeableimprovement despite lack of querying slave server.   Removed slave server for good.
  • 28. MongoDB  Public release in 2009  Document-oriented  ‘No-SQL’  b.collection.insert|add|update({parameters});  Retrieve subsets (for certain fields / objects)  Manages collections of objects in JSON-like format.  { "username" : "bob", "address" : { "street" : "123 Main Street", "city" : "Springfield", "state" : "NY" } }  Currently supports up to 64 indexes.  1 for ascending order, -1 for descending order.  b.collection.ensureIndex({username: 1, address.state: 1});  Nested fields can also be indexed.  Supports master-slave replication. Has automated database sharding.  Easily create a cluster.
  • 29. MongoDB (cont’d)   Module for Drupal exists!   http://drupal.org/project/mongodb   Cache, Field Storage, Blocks, Queues, Sessions, Watchdog currently supported.   Does a lot of the heavy lifting.   Very fast (Examiner.com uses it and has disabled page caching despite high load).   For anything exported into MongoDB, previous SQL queries will need to be modified so they become mongo queries.   For entities, use entityfieldquery   Anything in this format will actually allow you to switch between SQL and Mongo (and anything else) without changing code.   Look at http://drupal.org/project/efq_views for promising work into even more flexible views 
  • 30. PHP   Use a recent, stable release.   D7 requires 5.2.x, as do a few 6.x contributed modules.   D8 will require 5.3.x (yes, it’s a ways away ).   Install an Op-code cacher / accelerator.   Useful in bringing down memory usage for a site.   eAccelerator   APC   Xcache   Zend optimizer (commercial)   Anyone try HipHop?
  • 31. Running PHP   mod_php   Standard php module used by apache.   Few issues.   Well tested and supported.   Can be a resource hog.   FastCGI (PHP-FPM)   Can be used with NginX, Apache.   Runs as a separate process.   More stable with lower memory usage than mod_php.   Trickier to install (lots of good doc. online).
  • 32. Debugging PHP   Xdebug  http://xdebug.org  Display traces on error conditions.  Trace functions.  Profile PHP Scripts  Lots of docs online for installing.   kCacheGrind  Provides a visualization on bottlenecks in code.
  • 33. Op-code Caching   Benefits   Lowered memory usage.   Significant decrease in CPU utilization.   Usage on http://www.zimmertwins.com lowered memory usage per process from 20M down to less than 4M   Usage on http://calarts.edu lowered memory usage per process from 45M down to less than 10M.   Drawbacks   May crash   May require restarts after updating code (apc.stat = 0)
  • 34. Op-code Caching (cont’d)   Op-code caching will not work in all circumstances  Network connections.  Sorting arrays.  DB queries   Bad modules are bad.
  • 35. Drupal   Database intensive.   Can be a resource hog.   Memory intensive.   D7 > D6 > D5   Site may not be affected by bottleneck.   Quick Tips   Disable unnecessary modules.   Ifperformance is such a concern, sites *can* be made without Fields / CCK (ZimmerTwins was such a site)   Views UI, Field UI, Rules UI, <module> UI on production.   Make sure cron runs regularly.
  • 36. Drupal Tools   Devel   http://drupal.org/project/devel   Totalpage execution.   Query execution time.   Query log.   Memory utilization.   Can be combined with Stress Testing.   Trace   http://drupal.org/project/trace   Use for debugging.   Traces output, invocations, warnings.   Filter by Query Type.
  • 37. Module calls over network   Email users (og)   Call web widgets / APIs (youtube, twitter, facebook)   Cache as much data as possible.  Usehelper modules to aid with reducing bog.  Queues.  SMTP Mail module.
  • 38. Drupal Caching   Helpful in not querying / processing same bits of content over and over.   Especially for anonymous users all of whom may be viewing the same content on your site.   Many caches in core.   Bootstrap   Block   Field   Filter   Form   Image   Menu   Page (only for anonymous)   Many from contrib modules (like views, rules, media)
  • 39. Useful contrib caching modules EntityCache    http://drupal.org/project/entitycache  Caches all core entities (node, user, taxonomy) on entity_load()  Stays in cache until expiry or until content is updated/deleted. Boost    http://drupal.org/project/boost  Creates HTML for pages and stores it in files.  Requires changes to .htaccess file  Does not load up Drupal once content is cached to file for anon. users.  Can also use module to display site while in maintenance mode.  Varnish has mostly replaced this module (though they could play with each other) on sites not in shared hosting.  Views content cache  Block cache alter  Performance hacks
  • 40. Pluggable caching   Use $conf variable in settings.php   $conf['cache_backends'][]= ‘./path/to/cache_first_mechanism.inc’;   $conf['cache_backends'][]= ‘./path/to/cache_second_mechanism.inc’;   $conf[‘cache_class_<bin>’] = ‘CACHECLASS’;   $conf[‘cache_default_class’] = ‘SECONDCACHECLASS’;   Allows you to use a custom caching module.   Can even use this to completely disable caching (DO NOT USE ON A PRODUCTION SITE!)   Contrib modules   Cache Router (D6)   APC (D7)   Very fast.   Limited to caching on one web server (cache cannot get reused over multiple servers – can be good or bad.)
  • 41. Memcached   Distributed object caching in memory.   Written by danga for LiveJournal   Lives in memory   Can span multiple servers.   Seamless for D6, D7   D7 is still undergoing some changes.   Hash own requirements (apache needs to be restarted, have to clear old caches)   Slower than APC, but scalable.   Takes load off DB server (yay!)
  • 42. Search Mechanism   Drupal core search   Need I say anything?   Search API looks to be a much more promising (not to mention flexible!) option.   Pluggable system to support various types of backends.   Search API MongoDB, Search API Xapian   Supports Views.   Proposal to include in D8 (core conversation at DC London).   LuceneAPI   Not as fast as ApacheSolr.   Easy to setup.   Google CSE might also be a good fit.
  • 43. ApacheSolr   ApacheSolr   Very fast.   Easy to configure on *nix systems.   Requires a server on which Java can be installed.   More and more companies offering Solr as a Service (take load off your systems altogether).   Available as Drupal Module.   Apachesolr (very mature)   Search API apachesolr (very promising)   Views Plugin so even drive non-search related pages using Solr!
  • 44. Other options   Using an optimized distribution   Pressflow (for D6)   http://fourkitchens.com/pressflow-makes-drupal-scale   Only supports MySQL   Cleanly supports reverse proxies such as Varnish   Optimized for PHP5   Pressflow for D7 is currently identical to D7   Talksabout abandoning MySQL in favor of a MongoDB/ Cassandra DB architecture.   Many of the improvements made for Pressflow are in D7 core.   Keep in mind that a faster Drupal Core won’t save you from contrib modules behaving badly.
  • 45. Other options (cont’d)   Patching Drupal / Contrib  ‘Hack core’  Need to know what you are doing.  Sometimes necessary.  Create a patches directory where all the changes you make for a core/contrib file can be tracked and easily applied on updates.  Create own module and do any necessary schema updates / alters from there.
  • 46. Past experiences from Drupal Core   User login on zimmertwins.com was painfully slow (5+ seconds per user)   Gist of problem: DB not using index on username due to lower()   Bug had been around since 2006.   Solution: Modified patch on D.O for site.   User login time down to 0.1 seconds.   Pressflow avoids this by not allowing case-insensitve login.   Comments did not have index on user ID   Created index on user id as an update from my module.   If added in future, can remove my version of index.   Loading for comment by user no longer an issue.
  • 47. Advice for developers   Take advantage of caching.   Use memory wisely.   Unsetthe variable if you don’t have a need for it later.   Save variable to memory for future use so processing isn’t done multiple times (see drupal_static()).   Take advantage of AHAH functionality   Fewer queries.   Not reloading the page.   Saving bandwidth.   Learn to use jQuery (same as above).   Pay close attention to what
  • 48. Possibly related sessions   Note: Some have passed (but check out their screencasts)   Drupal on the Cloud   Patrick Wall   Drupal Development Q&A   Drupal 7 Q&A   Building APIs   Adam Gregory   Professional Staging and Deployment   Christefano   Understanding Ctools   Helior Colorado
  • 49. Questions   Have a question?   Want to talk more about performance?  Let’s talk after    Think you can help with Drupal performance?  http://goo.gl/I3PN2   Thank you 