Ez performance measurement
Upcoming SlideShare
Loading in...5
×
 

Ez performance measurement

on

  • 4,036 views

How to make sure a website can survive go-live and cope with ever increasing ...

How to make sure a website can survive go-live and cope with ever increasing
traffic and amounts of data: knowing what to measure and log, during both
development and production phases; load testing ; identifying bottlenecks;
preventing disasters

Statistics

Views

Total Views
4,036
Views on SlideShare
4,028
Embed Views
8

Actions

Likes
2
Downloads
17
Comments
1

2 Embeds 8

https://twitter.com 7
https://si0.twimg.com 1

Accessibility

Categories

Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
  • A note for all participants to the eZ Summer Camp: the slides are much longer and more detailed than what we had the time to go through during the workshop
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

Ez performance measurement Ez performance measurement Presentation Transcript

  • Make it scale! Tools and techniques for analyzing performances of eZ Publish websitesGaetano Giunta | eZ Summer Camp | Sept. 6 2012
  • Introduction
  • SynopsysHow to make sure a website can survive go-live and cope with ever increasingtraffic and amounts of data: knowing what to measure and log, during bothdevelopment and production phases; load testing; identifying bottlenecks;preventing disastersPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 3 View slide
  • Table of contentsThe workshop is comprised of 2 parts:• Theory • As you might guess, it’s all about slides • Can we skip this or do you want it really detailed? Raise hands!• Practice • Part 1: load testing • Part 2: performance logging Requirements: • a working eZ Publish 4 installation on Linux (Debian/Ubuntu preferred) [a VirtualBox image is available if you don’t have this] • Shell access, root access • Internet access • LibreOffice (or any other spreadsheet software)PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 4 View slide
  • Theory
  • GOALSFinishing the workshop early to go bathing does not count ;-)1. A scalable web site • It is impossible to go for infinite scalability • Expected traffic figures should ideally be known beforehand • If not, a round of load testing before go-live is highly recommended2. “Fast enough” pages • Definition of enough hash to be agreed upon: for a webshop it is smaller than for an institutional site • Page load times experienced by the user depends on user bandwidth as well as html/js optimization (but that takes a dedicated workshop of its own)scaling > faster pages • Typical developer mistake: test pages on his own laptop (concurrency = 1) • The fast page becomes extremely slow when concurrency increases • If traffic never increases, your career as web developer is on a wrong pathPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 6
  • What is “scalable” anyway?Lies, damn lies and statistics (W. Churchill)• Developers measure traffic in PVS (page views per second/minute/hour/day) • It is easy to relate to server load • PV != Hits • But serving static content should never pose a problem anyway• Customers measure traffic in concurrent users• It is a good idea to agree on metrics when defining goals• Analytics packages generally measure user session length and average page impressions per session => average page views per second per userPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 7
  • Math!Back-of-the-napkin typeA few useful formulas: • Apache MaxClients x max memory for a web page = server memory - OS memory (assuming you are not running other stuff on the webserver, which you shouldn’t) • Max PVs = Ap. MaxClients x 60 / ( page generation + delivery time ) • PVs = Max concurrent users x Avg user sess. impressions / Avg user sess. length• Tips • User session != webserver process • Using a reverse proxy is almost always a good idea • Apache processes never release memory until they are recycled • Clients with low bandwidth keep a webserver process occupied for a long time; R.P. acts as “buffer” • Avoiding server swapping gives better results under peak traffic • Limiting traffic at the webserver preserves the rest of the server farm from meltdownPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 8
  • Nosce te ipsum WE NEED TO MEASURE RESOURCE CONSUMPTION TO FIND AND REMOVE SCALABILITY BOTTLENECKS• Many resources are involved in serving web pages • Network • Server hardware • OS • Apache / PHP / APC / eZ Publish • Database • Solr / external services / more…• Scalability is determined by the most scarce resource (bottleneck)• …which is generally not known beforehand• Improving response time for a resource which is not loaded can have the perverse effect of overloading the bottleneck resource and actually decrease performances!• eZ Publish does many things “behind the back” of the developer• developers suck anyway*PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 9
  • The art of measurement, I What to measureThe more you measure, the slower the system (Heisemberg principle)The more you measure, the harder it is to grasp the overall system stateFor eZ Publish applications, start with: • Ram, CPU, IO (disk), DB requests • Can be measured either globally on the server (BLACK BOX) or “per page” (WHITE BOX) • “per page” numbers will usually not vary between environments • Time taken to generate web pages • will vary depending on many factors (dev != prod) • Split between the time eZ does “computation” and access to external resources • Other? • Number of active user sessions PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 10
  • Information overloadCan you spot the problem? Oracle AWR report (this is the «summary», the report actually goes on for 10 pages)Up to eZ 4.6 eZ 4.7 and laterPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 11
  • The art of measurement, IIWhen to measure 1. During development - to avoid nasty surprises when it’s too late • measured data should be easily understandable by developer • it should in fact always be straight in his face • it should be easy to drill down on specific problems • all the way down to profiling every php function call 2. Before go-live - to validate production HW and architecture • Never assume that production hw will magically solve all problems • Sysadmins are morons anyway* • This is the good time for some load-testing 3. In real-life usage - for post-mortem analysis, troubleshooting and more • A small percentage of users could be getting slow pages without overall stats being impacted • Things always change over timePRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 12
  • The art of measurement, IIIHow to measure • Black Box: measure load of the (web)server • CLI tools: vmstat, free, iostat, top, ps, atop, dstat, etc… • PHP: APC control panel • Apache: mod_status • Mysql: mtop, innotop, percona toolkit, mysqli_get_client_stats, monyog, mysql ent. monitor • Monitoring systems: munin, cacti, zenoss, etc… (nb: availability monitoring != performance monitoring) • Need to correlate data with web traffic • Need to execute load testing to simulate real-life traffic • White Box: measure load per page • eZ debug output is good - but it is too detailed • It does slow down the site a bit • Reporting needed to compare evolution over time: have to log data somewhere • Enter ezperformancelogger (but also ezsnpd, ezmunin, etc…)PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 13
  • The art of measurement, IVTips• Always check error logs if there is something unexpected in measured data• Testing should be done using a realistic data set (eg. 10.000 users, not 10)• The clock of all servers should be in sync to allow correlation (no, really)• Measuring VMs: time is generally a liePRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 14
  • Load Testing
  • Load testing strategies I1. Baseline test a) Test downloading a small static file, a big static file b) Test executing the most simple php page which executes a db query c) Increment concurrent users until you get no increase in hits / second • Useful to uncover configuration errors in network / db / AMP stack • This is the “idealistic” goal for your dynamic pages2. Bruteforce test a) Hit the homepage N times in a row b) Increment concurrent users until you get no increase in hits / second • Can be run on other most-visited pages on the site as well • Make sure you’re not testing 404 pages (or redirects) • Reset user sessions between runs if they are auto-generated • Keep open vmstat and iostat while test runs to quickly identify bottleneck • Does not really correlate to concurrent users • Results difficult to communicate to customerPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 16
  • Load testing strategies II3. Full site navigation test a) Enable logging of interesting data b) Use wget or httrack to navigate the whole site c) Get log files data into a spreadhseet • Useful to uncover pages with bad resource usage • Can be run with both cold and warm caches to gauge cache efficiency4. Scenario testing • Need support from end user to determine most likely/useful scenarios • Takes time to configure in load-testing tool • Do not believe tools that promise to automagically generate a scenario by “sniffing” browser sessions: manual intervention will be needed • Always validate first each single response before running the whole test • The one test which is closer to real life… • …but also one which is easy to manipulate (many knobs to tweak)PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 17
  • Load testing tips• Do not use the webserver to execute the client (to avoid impact on cpu)• Do not measure routers, firewalls or network card performance either (by testing from remote network), unless what you want is real-world measures• Always write down complete hw and sw specs – some of it will have changed next time you want to run the test for comparison (a good idea: zip and save complete apache and php config files, write down command line used for client in the report)• Automate tasks to avoid human error / getting boredPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 18
  • Load testing tools(the ones I know about)• Apache bench • Good: always available • Bad: not very flexible; limited support for advanced http features• Siege • Good: better than Apache Bench; some support for scenario testing • Bad: not available by default in many linux distros (or windows)• Jmeter • Good: allows complex scenario testing; can run tests from a farm of machines • Bad: has a learning curve; needs Java• Httperf, web polygraph, …• Web-based tools • Good: can test from many locations across the world; easy to use • Bad: usually do not offer too much flexibility; $$$• Roll-your-own (php) script • Good: flexible; can be used on servers where you can not install any other software • Bad: can not compare with other measures; guarantee of correctness • Excellent: it is there to use! https://github.com/gggeek/ezabPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 19
  • Load testing tools: ezabhttps://github.com/gggeek/ezabComes with two scripts:• ezab.php replacement for ApacheBench when it is not available• abrunner.php runs ab many times in a rows and produces reportsUseful for strategies I and IIExample: testing the VirtualBox VM from the host OS – baseline datac favicon.ico phpinfo.php 500 1000 2000 200 450 900 1800 180 400 800 1600 160 Requests per second Requests per second 350 700 1400 140 Time per request ms (mean) Time per request ms (mean) 300 600 1200 120 Time per request (90%) Time per request (90%) 250 500 1000 100 ms rps ms rps Time per request (min) Time per request (min) 200 400 800 80 Time per request (max) Time per request (max) 150 300 600 60 Time per request (median) Time per request (median) 100 200 400 40 50 100 200 20 0 0 0 0 1 2 4 8 16 32 64 128 1 2 4 8 16 32 64 128 concurrent clients concurrent clientsPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 20
  • Load testing tools: ezabThis is an exercice you are expected to carry outTesting performance of the eZ Publish 4 installation (homepage) on VirtualBox1. Install abrunner wget https://raw.github.com/gggeek/ezab/master/abrunner.php2. Execute: test homepage of the installed site php abrunner.php -s ezpublish4.ezsc -u / -c "1 2 4 8 16" -a3. Import into LibreOffice the resulting csv file: test_logs/_.csv4. Create a graph by selecting the first 7 columns • Chart Type: line • Data range: «1° column as label»5. Icing on the cake: set a separate Y axis for the number of requests/second6. Stop VM, add cpus, reboot and repeat steps 1-5, compare graphs • Use the -l option to ezab to get different file names for reportsPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 21
  • Load testing: eZ Publish homepage 6000 10See anything strange? 9NB: this VM has 4 VCPUs 5000 8 7 Requests per second 4000 6 Time per request ms (mean) 3000 5 Time per request (90%) Time per request (min) 4 Time per request (max) 2000 3 Time per request (median) 2 1000 1 0 0 1 2 4 8 16PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 22
  • Load testing: eZ Publish homepage 6000 10See anything strange? 9NB: this VM has 4 VCPUs 5000 8 7 Requests per second 4000 6 Time per request ms (mean) 3000 5 Time per request (90%)• Performance is waay too little: 4 Time per request (min) Time per request (max) 2000 6 page views per second 3 Time per request (median) 2• RPS does not increase going 1000 1 from 1 to 16 concurrent clients 0 0 1 2 4 8 16PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 23
  • Load testing: eZ Publish homepage 350 200See anything strange? 180NB: this VM has 4 VCPUs 300 160 250 140 Requests per second 120 Time per request ms (mean) 200 100 Time per request (90%)• Performance is waay too little: 150 80 Time per request (min) Time per request (max) 6 page views per second 100 60 Time per request (median) 40• RPS does not increase going 50 20 from 1 to 16 concurrent clients 0 0 1 2 4 8 16• Xdebug is ON, APC is OFF!sudo apt-get install php-apcsudo mv /etc/php5/apache2/conf.d/xdebug.ini /etc/php5/apache2/conf.d/xdbg.ini.baksudo service apache2 restart• Test again: we get 150 rps at concurrency 4 and 8 :-)PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 24
  • Load testing: eZ Publish homepageQ: Is the site CPU or memory bound?The answer is:Cpu boundAt 16 concurrent clients,cpu usage is near 100%RAM is more than enoughAnd RPS decrease No swap Cpu idle timePRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 25
  • Load testing tools: HTTRACK• Website copier• GUI app on windows, web-based (or command-line) on linux• Used for further exercices later on• Alternatives: wget -R• Install and launch: sudo apt-get install webhttrack sudo /usr/lib/httrack/htsserver /usr/share/httrack/• Connect to http://192.168.56.101:8080/Tips• Make sure the server can send requests to itself: add to /etc/hosts the ezpublish4.ezsc hostname• If a robots.txt file is present, it will be respected by defaultPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 26
  • eZ Performance Logger
  • eZ Performance Loggerhttp://projects.ez.no/ezperformancelogger• Allows the developer to define a set of variables (KPI) to be measured• Each variable is measured for every page view (rest / ajax pages as well)• Measured KPIs can be sent to multiple logging systems • Csv formatted log probably the easiest to later parse • Apache’s own access log probably the best suited to avoid any performance hit • Supports logging directly to Google Analytics or Piwik via rewriting of html pages• Common KPIs are available (eg. db queries, db time), custom ones can be added• Integrates with Munin to visualize the measured data• Throws in full integration with XHProf profiler as bonus • According to facebook “good enough” to keep enabled in production• Does NOT come with a nice GUI of its ownPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 28
  • InstallationRequirements• eZ Publish 4.x• Apache webserver recommended• Optional: Xhprof• Optional: a Google Analytics account or Piwik• Optional: MuninThe extension comes preinstalled in the Virtual Machine for the WorkshopTo install by hand, follow the standard procedure – no need to touch the databaseFor advanced users: in classes/tracers, alternative connectors are provided for mysqlidatabase and ezdfs cluster, which allow to measure performance data even inproduction environments (where ezdebug is turned off).PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 29
  • First steps1. Unzip and activate extension2. Set logging format to csv-formatted file: create file settings/override/ezperformancelogger.ini.append.php [GeneralSettings] LogMethods[] LogMethods[]=csv [csvSettings] FileName=var/log/ezperflog.csv3. Set performance indicators to be logged, eg: memory used, execution time [GeneralSettings] TrackVariables[] TrackVariables[]=mem_usage TrackVariables[]=execution_time4. Test that it is working: 1. Browse to the homepage 2. Check for presence of var/log/ezperflog.csvPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 30
  • Measuring database queries per pageThe eZPublish database connector measures all the commands sent to the database:• Number of queries executed• Time takenThis is only done when the debug output is enabled.eZPerformanceLogger allows to log any existing «timing point»1. Enable debug output (in settings/override/site.ini.append.php) [DebugSettings] DebugOutput=enabled2. Add number of queries and time taken to the performance indicators logged TrackVariables[]=accumulators/mysqli_query/count TrackVariables[]=accumulators/mysqli_query3. Rotate the csv file, since it will now have a different number of columns php extension/ezperformancelogger/bin/php/rotateperflogs.php4. Reload homepage, check log filePRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 31
  • Making sure all requests are measuredBy default ajax calls and requests which end up in a redirect are not logged.How to fix:• Edit index.php, on line 198 add eZExecution::addCleanupHandler( array( eZPerfLogger, cleanup ) );• Browse content in the Admin interface (which uses ezjscore)• Look for calls to ezjscore/call in var/log/ezperflog.csv• Other frontend controllers have to be patched as well • index_ajax.php (removed in recent versions) • index_treemenu.php, index_treemenu_tags.php, index_soap.php • index_cluster.php currently not supportedPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 32
  • Visualization of data: spreadsheetsQ: is the site database-bound ?1. Use httrack to navigate the whole site (cache warmup) Tip: exclude from files to be downloaded all images, css, js, m4v2. Rotate log file php extension/ezperformancelogger/bin/php/rotateperflogs.php3. Use httrack to navigate the whole site again4. Rotate log file5. Import log file into LibreOffice6. Graph db queries per page, db time per page as % of page timeQ: how effective is the view cache?i. Disable the view cacheii. Clear all cachesiii. Execute steps 1 to 6 againiv. Compare the number of queries per pagePRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 33
  • Visualization of data: spreadsheetsPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 34
  • Visualization of data: Munin Ihttp://www.munin-monitoring.org/• Munin is an open source monitoring tool• It generates daily and weekly graphs for collected data• It collects a lot of data from the operating system• It comes with a wide set of plugins for existing software, such as Apache and MySql• Creating plugins for new software is relatively easy• Agent-based architecture: a munin “master” server can collect and display data from multiple “node” servers• For our scenario, the webserver acts as both master and node • Master: runs a cronjob that generates reports by querying nodes and stores them in /var/cache/munin; the reports are made available via Apache • Slave: runs a daemon: munin-node, listening on port 4949PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 35
  • Visualization of data: Munin IPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 36
  • Visualization of data: Munin IIIntegration of eZ Performance Logger• The default interval for collecting data is 5 minutes (it should be more flexible in version 2)• All ezperformancelogger KPIs can be shown in a Munin graph• By default, the “per page” value of the PKI is shown• In every graph, the Average, Maximum and Minimum value are shown• Via eZ Publish settings, appearance of those graphs can be tuned• Note: the timestamp of the last time the munin plugin has collected data from ezperformancelogger for any specific KPI is stored in var/<vardir>/logPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 37
  • Integrating with Munin I1. make sure you have a valid munin-node installation on your webserver Connect to http://192.168.56.101/munin If you get an access denied method, edit /etc/apache2/conf.d/munin <Directory /var/cache/munin/www> Allow from all2. Symlink the file bin/scripts/ezmuninperflogger_ into /usr/share/munin/plugins/ and make it executablecd extension/ezperformancelogger/bin/scriptschmod 755 ezmuninperflogger_sudo ln –s/var/www/ezpublish4/extension/ezperformancelogger/bin/scripts/ezmuninperflogger_/usr/share/munin/pluginsFix an error in the script ezmuninperflogger_: on line 1 put #!/bin/bashinstead of #!/bin/sh PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 38
  • Integrating with Munin II3. Create a configuration file for the munin plugin: sudo vi /etc/munin/plugin-conf.d/ezmuninperflogger [ezmuninperflogger_*] env.php /usr/bin/php env.ezpublishroot /var/www/ezpublish44. Restart the munin node service sudo service munin-node restart5. check if the configuration works: sudo munin-node-configure --suggest If it does, you should see in the output a line similar to: ezmuninperflogger_ | no | yes (+execution_time +mem_usage) the "yes" in the 2nd column is important. Between parenthesis you get the list of variables which can be graphedPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 39
  • Integrating with Munin III6. activate the plugin: sudo munin-node-configure --suggest –shell You should get 3 lines with "ln -s ..." commands. Execute them (nb: as root)7. test that it works: run: sudo munin-run ezmuninperflogger_<$varname>8. restart munin-node again: sudo service munin-node restart9. navigate the site, wait 5 minutes, connect to Munin again. Troubleshooting tip: munin logs are available in /var/log/munin10. Integrate Munin in the eZ administration interface: edit ezperformancelogger.ini [MuninSettings] MuninURL=http://192.168.56.101/munin/11. Optionally, you can customize how the variables recorded will show up in Munin graphs by editing more ini settings in section [MuninSettings]PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 40
  • Integrating with Munin IIIPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 41
  • Drilling down on hot code pathsXHProf• http://pecl.php.net/package/xhprof• Profiler from Facebook• Designed to be fast enough to be used in production (at least faster than Xdebug ;-)• Comes with its own web-based GUI• Installation sudo apt-get install graphviz sudo pecl config-set preferred_state beta sudo pecl install xhprof sudo vi /etc/php5/apache2/conf.d/xhprof.ini => add extension=xhprof.so sudo service apache2 restart• While at it, disable apc (???) PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 42
  • Activating XHprof• edit your config.php file and add the following lines at the top: (if you miss the config.php file, copy config.php-RECOMMENDED into config.php) include( extension/ezperformancelogger/classes/ezxhproflogger.php ); eZXHProfLogger::start();• Log in to admin interface, go to Setup tab, bottom-left menu item: XHProf Profiling• You can see the data recorded for the pages you have just browsed to• Click on the name of a run to get profiling information in all its gory detailTips• To avoid logging profiling of all pages, you can start it in any place in the code• Links to profiling runs will be displayed in the debug output as well...• ...but enabling debug output does have an impact on profiling • A cronjob is available to periodically remove old profiling dataPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 43
  • Activating XHprofPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 44
  • Activating XHprofPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 45
  • Activating XHprofPRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 46
  • Questions?Thanks for participating!These slides https://dl.dropbox.com/u/520168/eZ%20Performance%20Measurement.pdfSource code, command snippets https://gist.github.com/gggeek (look for gists numbered 1 to 9)About me Consultant for eZ Systems since 2007 gg@ez.no @gggeek http://share.ez.no/blogs/gaetano-giunta http://projects.ez.no/users/community/gaetano_giuntaWith helpful support from Yannick Modah Gouez! ( ymg@ez.no )* = I hope you were not offended by jokes about developers and sysadmins.I consider myself a devop: someone incarnating the worst aspects of both ;-)PRESENTER: GAETANO GIUNTA 6/9/2012 SLIDE 47