Web Performance tuning presentation given at http://www.chippewavalleycodecamp.com/
Covers basic http flow, measuring performance, common changes to improve performance now, and several tools and techniques you can use now.
2. John McCaffrey: Its all about me
• Doing Java since 2000, Rails since 2007
• Presented at WindyCityRails
– 2008: Advanced Firebug and JS unit testing
– 2009: PDF Generation in Rails
– 2010: Rails Performance Tuning
• Addicted to Performance Tuning
• railsperformance.blogspot.com
• Feedback: http://spkr8.com/t/4961
@J_McCaffrey
railsperformance@gmail.com
(the slides will be available, with additional references)
3. What are we gonna cover?
• Performance touches a lot of layers
• We have a very diverse audience
6. def organization_branches
get_organizations.flatten.uniq.select{|org| org.children.size > 0}
end
def organization_leaves
get_organizations.flatten.uniq.select{|org| org.children.size == 0}
end
def load_organizations
branches = current_user.organization_branches
leaves = current_user.organization_leaves
@branches = branches.to_json(:only => [:id, :name])
@leaves = leaves.to_json(:only => [:id, :name])
@organizations = @criteria.branch ?
leaves.collect {|o| [ o.name.titleize, o.id ] }.sort :
branches.collect {|o| [ o.name.titleize, o.id ] }.sort
end
It all started...
2811 queries!
Down to 17!!
7. And then...
Invoicing issue:
If a customer resubmits an order with 2 years,
they should be charged $9 but were only being
charged $4.50
How common are ROI calculations on software projects?
53626 * $4.50 = $241,317 (just in the last year)
8. Our theme of the day: Leverage
to achieve the greatest value
10. Performance and Business metrics: Speed matters
Amazon: 100 ms delay caused a 1% drop in revenue.
Google: 400 ms delay caused a 0.59% decrease in search requests per user.
Yahoo!: 400 ms delay caused a 5-9% decrease in traffic.
Bing: 2 sec delay caused a 4.3% drop in revenue per user.
Mozilla made their download page 2.2 seconds faster and saw an
increase of 15.4% in downloads.
Google Maps reduced the file volume by 30% and observed a
30% increase in map requests.
Netflix enabled gzip on the server; pages became 13-25% faster and
saved 50% of traffic volume!
11. Performance and Business metrics: Shopzilla.com
Went from 6 sec
down to 1.2sec
http://www.phpied.com/the-performance-business-pitch/
12. Fred Wilson: Speed is the most important feature
10 Golden Principles of Successful Web Apps
Fred Wilson, Venture Capitalist www.avc.com
•Speed• Instant Utility
• Software is media
• Less is More
• Make it programmable
• Make it personal
• Make it RESTful
• Discoverability
• Clean
• Playful
14. Give yourself a fighting chance, make a plan
Alois Reitbauer, dynatrace
Anti-patterns
• Thinking Scalability something you just sprinkle on later
• Guessing, not testing
• Ad-hoc, unstructured
• Waiting until it hurts, not proactively planning
• Thinking that any early planning is Premature optimization
Video on Parleys.com
18. Premature optimization
"Premature optimization is
the root of all evil'
-Donald Knuth
"Quoting Knuth, as a way to
avoid planning for performance
issues, is a cop-out"
@J_McCaffrey
Don't waste time on making
optimization changes until you
KNOW they are necessary
19. If you can not measure it,
you can not improve it.
— Lord Kelvin
20. Let's do this the right way
Measure:
• Understand the tools to use, at each layer, to measure
and monitor performance
Create repeatable tests:
• Be able to invoke the system and easily test
• A/B, before/after tests, long-term monitoring
Isolate your changes:
• One thing at a time
That's it: Measure, Test, and Isolate your changes
21. Terminology
Load:
How much work is
there?
Utilization:
How much of the
resources are in use?
Scalability:
How well can we
handle the load?
Throughput:
How many tasks can
be done per unit of
time?
Concurrency:
How many tasks can
we do at once?
Capacity:
How big can our
throughput go before
things fall apart?
Response time (avg) == Latency
Requests per second == Throughput
22. Terminology
Performance:
• The resources and Time for a single request
• Measured in Response Time
Throughput:
• The number requests that can be handled per/time
• Measured in Requests per second
Scalability:
• Ability to grow and handle more requests, without
significantly degrading the experience
Performance != Scalability
23. Terminology
Let's say:
1 worker that can compute a job in .5 sec
Throughput = 2 jobs per sec
Latency = .5 sec per job
Performance != Scalability
Adding more workers will not improve the latency
But not having enough workers can increase the
response time, due to Queuing and delay
25. Waterfall: Twitter.com
Waterfall
Total time: 3815 ms
Dynamic generation: 250 ms
(~7%)
Static: 3369 ms (~89%)
* Will not add to 100%
When the html is downloaded
When the rest
of it is loaded
28. Firebug Netpanel
Simple and repeatable way to measure page load
Usually, less than 20% of the time is spent in fetching the HTML
So how do we optimize the other 80%?
29. YSlow
• Created by Steve Souders
• Observes the netpanel data, applies rules and gives you a
score.
• Deep understanding of how the browser works
• Its the best way to have a repeatable, measurement of your
site's load performance
http://developer.yahoo.com/yslow
30. YSlow
• Minimize HTTP Requests
• Use a Content Delivery Network
• Add an Expires or a Cache-Control Header
• Gzip Components
• Put StyleSheets at the Top
• Put Scripts at the Bottom
• Avoid CSS Expressions
• Make JavaScript and CSS External
• Reduce DNS Lookups
• Minify JavaScript and CSS
• Avoid Redirects
• Remove Duplicate Scripts
• Configure ETags
• Make AJAX Cacheable
• Use GET for AJAX Requests
• Reduce the Number of DOM Elements
• No 404s
• Reduce Cookie Size
• Use Cookie-Free Domains for Components
• Avoid Filters
• Do Not Scale Images in HTML
• Make favicon.ico Small and Cacheable
32. Google Page speed
• Similar to YSlow
• includes paint events
• separates out ads, trackers from your content
• better recommendations and optimizations
• 1 click to see optimized version of your files
http://code.google.com/speed/page-speed
36. Summary
• Page load is the place to start
• Read everything in the Yslow and PageSpeed rules
• Encourage other devs to play with YSlow, PageSpeed
• Show Webpagetest.org, and zoompf.com to the boss
Now that we have a way to measure page
performance, we can think about making changes
38. Isn't everyone already doing this?
Analysis of the Alexa top 1000 sites found:
• 42% don't gzip
• 44% have more than 2 css files
• 56% serve css from a cookied domain
• 62% don't minify
• 31% have more than 100k size css
Top 1000 retail sites
50% aren't doing both keep-alive and compression
(the 2 easiest things!!)
39. Compress with Gzip
• Can save you 50% to 80% of your bandwidth
• Simplest performance improvement you can make
• Beware differences in browsers
• File types: html, js, css, xml, rss, txt, font, json
Probably the simplest and smartest
thing you can do for your app!
40. Combine and minify javascript and css
Rails
• javascript_include_tag :all, :cache => "all"
• Asset_packager, Jammit
Load common js libraries from google
http://github.com/rpheath/google_ajax_libraries_api
Jsmin or YUICompressor
Closure compiler
Let's check it out!
41. Combine and minify javascript and css
Sprites and image optimization
• spriteme.org
• JqueryUI already gives you sprited images
Image optimization
• Smush.it
42. Expires and Browser Caching
Setting a far future expires tells the browser not to ask for the file
Expires:Mon, 02 Sep 2030 21:23:08 GMT
/images/logo.png?1234567890
Using the query string lets us break the caching by having the
browser ask for a new file
There is a problem with using a URL query string this way
http://blog.eliotsykes.com/2010/05/06/why-rails-asset-caching-is-broken/
/images/logo-fp-839b180ff39a24f8d6e0ee70e4c40fed.png
43. Rails Caching
Take your dynamic content, and make it static
• Page
• Action
• Fragment
Full length tutorials
• http://railslab.newrelic.com/
• http://railscasts.com/episodes?search=caching
• Greg Pollack Aloha on Rails http://vimeo.com/10860860
49. Commercial tools
New Relic
• Developer plugin is a good start
• Free version doesn’t get you much (Bronze version with heroku)
• Works for .NET
• Scoutapp.com
• Similar to new Relic
• Large selection of community plugins
50. Testing tools
Apache bench
ab -n 10 -c 2 http://www.somewhere.com/
Httperf
httperf --server localhost --port 3000 --uri / --num-conns 10000
Jmeter
yes, its ugly, but its powerful
52. Database issues
Common database issues
• Bad queries
• Not utilizing explain
• Inadequate indexes
• N+1 queries
• Selecting more data than is needed
• Inconsistent queries for the same data
53. Query_reviewer
Common database issues
• Bad queries
• Not utilizing explain and slow query log
• http://github.com/dsboulder/query_reviewer
• Runs explain on all of your queries, outputs to div in
page
55. Bullet plugin
Bullet plugin
• http://github.com/flyerhzm/bullet
Help you reduce the number of queries
with alerts (and growl).
56. Slim_scrooge
Be more specific in your select
• http://github.com/sdsykes/slim_scrooge
• Instruments your code
• Observes the usage pattern
• Suggests/changes your select statement
• When invoked within the exact same context
57. Ruby Issues
Most are due to memory problems
• Slow GC
• Not release references
• Capturing scope
• Profiling will reveal what’s going on
• http://guides.rubyonrails.org/performance_testing.html
If you want to learn everything there is to know about profiling
• Aman Gupta & Joe Damato
• http://timetobleed.com/
• http://memprof.com/
• They’ve already identified and fixed several memory leaks
58. Ruby issues: Faster Ruby Libraries
C Extension++
• XML parser http://nokogiri.org/
• JSON parser http://github.com/brianmario/yajl-ruby/
• CSV parser http://www.toastyapps.com/excelsior/
• HTTP client http://github.com/pauldix/typhoeus
Date
http://github.com/jeremyevans/home_run
59. Take home points
Performance
• Performance affects the bottom line
• The biggest performance win is usually in improving the load time
• Continue to monitor and test your apps performance over time
• Gzip, combine and minify to get a big boost
• Lack of indexes is likely to be one of your biggest backend issues
• Try to stay up to date with Libraries and patches
• A well tested codebase is easier to tune!
• Upgrading to ruby 1.9 will give you a huge performance boost
60. Links: great info on performance
Performance
• Scaling Rails series http://railslab.newrelic.com
• RailsCasts.com
• Velocity Conf http://en.oreilly.com/velocity2010
• Google io conf http://code.google.com/events/io/2010/
• dynatrace site http://ajax.dynatrace.com/pages/
• http://www.igvita.com
• http://www.mysqlperformanceblog.com
61. Wait, I made it through that?
Designing for performance
• Watch how they really use it
• Multiple tabs?
• Back and forth between list and detail view
• Great presentation at WindyCityRails2009 by David Eisinger
• Optimizing perceived performance
• Make it feel fast
62. Wait, I made it through that?
Architectural musings
•Background long running tasks
• Email, image processing, reports, feeds
•Leverage Rack, Sinatra or Padrino
•Parallel processing, Event Machine
•Learn from heroku, scalable from the start
•Revaluate your cloudy-ness
• Great RailsConf presentation by James Golick
64. Beyond ySlow:
improving page
performance
outside the
initial load
MySql and
Postgres
Database tuning
Heroku: Building
Scalable apps
from the start
Tuning
performance
for Mobile
devices
Performance
Improvements
coming in Rails
3, and 3.1
Speeding up
your tests
65. Questions for you
Architectural musings
• Moving to or already on Ruby 1.9?
• Serving your content from a separate domain
• Cookieless?
Hinweis der Redaktion
I really enjoy helping people get the most out of their apps
add some color to these slides!
There is so much to cover, and I want to make sure that everyone gets something valuable out of this talk, whether you are just starting out, or you work for Groupon, serving mils of hits a day.
My main objective is to help you get started. To know where to look, which tools to use, which techniques to investigate, and where to find the resources you need.
I won't be able to teach you everything about performance, but I can at least get you excited, and point you in the right direction.
If i don't teach you something new, or you don't get what you were looking for, please come up and talk to me. We can chat afterwards, if you have a specific issue, I'm sure I can help you out.
how I got started
customer complained report page was slow in prod
first 2 methods are from the model, the 3rd is from the controller
This code basically sets up 2 drop downs to pick which organization you want to run a report for.
I knew it was slow, and I knew it was wrong, but I didn't know all the ways that it was wrong
got my first taste of poorly performing Rails code
I learned:
test with real data
review your queries
let the database do what it do
watch out for N+1 queries
avoid unnecessary columns
add indexes
optimize the page
lack of testing, leads to an odd pattern of reuse without refactoring or improving
It was a rewarding experience knowing that a few basic changes were now going to improve the performance in several places in the app, and make a lot of people very happy.
The next day they asked me if I could look into a discrepancy they had in the billing code.
This only occurs if they have to resubmit due to their own delay, and then only if they have more than one year ordered.
I fixed the issue, but then became ridiculously curious about how much this bug had cost.
Its not just having 2yrs, its having more than one, and there are several orders with 3 or 4yrs.
It was one of the first times I knew how much a bug cost, and could measure the return on the time it took to fix it.
i then became obsessed with knowing how much a unit of development was going to be worth.
It was a powerful feeling, to have real numbers.
This is info that the rest of the business world is used to having, and how McDonald's knows that it can sell 59cent cheesburgers.
What's the ROI?
When will we break even?
These 2 experiences, where I was able to observe massive improvements with a small but focused effort, encouraged me to seek out more opportunities to maximize return
Leverage is an important theme here. When it comes to making the right performance decisions, we want to use the greatest leverage, and get big gains for small effort.
Gzip is a great example of leverage
Open source software
Jake’s talk on metric_fu is a great example of leverage
Nick’s talk on improving test performance is extremely valuable
Improving performance is all about leverage, and its all about money
The major sites know this, and you can see it in the work they do.
Google just released instant search, because they know that faster search means more money
google includes it in their adwords calculation
velocity conference 2009
google did a similar experiment, adding a delay of only 400ms, and saw a .6% drop in searches, but most interestingly, even after they were done with the experiment,and removed the delay, those users still did significantly less searches,, showing that bad performance leaves a lasting impression
Yahoo also added 400ms delay, and saw a 5-9% drop in full page traffic, meaning that as soon as the users started to feel a slow experience, they killed the page and went somewhere else (something to think about if you are the #3 player)
http://www.phpied.com/the-performance-business-pitch/
Improving performance is all about leverage, and its all about money
The major sites know this, and you can see it in the work they do.
Google just released instant search, because they know that faster search means more money
google includes it in their adwords calculation
velocity conference 2009
google did a similar experiment, adding a delay of only 400ms, and saw a .6% drop in searches, but most interestingly, even after they were done with the experiment,and removed the delay, those users still did significantly less searches,, showing that bad performance leaves a lasting impression
Yahoo also added 400ms delay, and saw a 5-9% drop in full page traffic, meaning that as soon as the users started to feel a slow experience, they killed the page and went somewhere else (something to think about if you are the #3 player)
http://www.phpied.com/the-performance-business-pitch/
Shopzilla wasn't doing an experiment, they had a slow site and they needed to fix it
big comparison site
100mil impressions aday
8000 requests per sec
Improved response time from 6
down to 1.2
These slides from a deck that was put together for the purpose of explaining performance tuning to business users,
http://thinkvitamin.com/business/fred-wilsons-10-golden-principles-of-successful-web-apps/
Said what they look for in the web apps they invest in
"not just a feature, but the most critical feature"
especially true of non-power users, users that have a choice
As part of their due dillegence they investigate the speed of the app, and they continue to monitor it, like watching a stock ticker.
They know that when the app gets slow, less users use it.
Its a great video, you should check it out
Any discussion of performance monitoring or measurement is met with sneers and the oft quoted 'Premature optimization is the root of all evil'
http://wiki.parleys.com/display/PARLEYS/Home#talk=35061764;title=Common%20Performance%20Antipatterns
I like Donald Knuth
ka-nooth
http://www-cs-faculty.stanford.edu/~uno/faq.html
author of the are of software programming
its true, and its worth reminding ourselves.
We should forget about small efficiencies, say about 97% of the time: Premature optimization is the root of all evil
We should forget about small efficiencies, say about 97% of the time: Premature optimization is the root of all evil
some people quote it, and to them it means even thinking about performance is premature
Let's make a distinction between premature optimization, and baseline measuring.
we're gonna make sure we measure, and tackle the most critical problems
Measuring your performance is not premature optimization
Makes sense for capacity planning
SLA
Should line up with performance expectations of the project
You need to establish the expectations, state your assumptions, and keep an eye on things
why do people say that?
scalability is a term that is thrown around a lot, its meaning becomes unclear, and people make bad decisions because of it
Its possible to have an app with great performance, that can not scale
And its possible to have an app that scales, but its performance is horrible
We need to know what we are measuring
Today we're going to focus on Response time and Requests per second
why do people say that?
scalability is a term that is thrown around a lot, its meaning becomes unclear, and people make bad decisions because of it
Its possible to have an app with great performance, that can not scale
And its possible to have an app that scales, but its performance is horrible
why do people say that?
scalability is a term that is thrown around a lot, its meaning becomes unclear, and people make bad decisions because of it
10 workers = 20req sec
If latency goes up to a full sec...
Its possible to have an app with great performance, that can not scale
And its possible to have an app that scales, but its performance is horrible
This kind of queuing and delay is what we want to eliminate, it represents time not working, just waiting
Here's your typical stack
at each one of these layers there are performance considerations
Our goal will be to get each layer to do less work, and distribute the load from busy workers to not so busy workers
Traditionally, performance tuning focuses on the backend, as it is logical that dynamic data takes more effort to serve up
How do we know if our site is slow?
What is it doing?
waterfalls from
zoompf
How do we know if our site is slow?
What is it doing?
How do we know if our site is slow?
What is it doing?
steve souders
steve souders
http://stevesouders.com/cuzillion/
steve souders
now we have 2 tools that allow us to measure the speed of our app
anyone on the team can do it.
I encourage you to test out your apps, if you haven't already
look up your competition, industry, etc
You can ask showslow to check on a site for you everyday
There is so much to cover, and I want to make sure that everyone gets something valuable out of this talk, whether you are just starting out, or you work for Groupon, serving mils of hits a day.
My main objective is to help you get started. To know where to look, which tools to use, which techniques to investigate, and where to find the resources you need.
I won't be able to teach you everything about performance, but I can at least get you excited, and point you in the right direction.
If i don't teach you something new, or you don't get what you were looking for, please come up and talk to me. We can chat afterwards, if you have a specific issue, I'm sure I can help you out.
www.stubornella.org
netflix saw 43% decrease in bandwidth after they turned on compression
google ajax library allows you to load local files, great when you are on the train
to minimize delays due to network performance, and the browsers limitations on parallel downloads, we want to reduce http calls
google ajax library allows you to load local files, great when you are on the train
to minimize delays due to network performance, and the browsers limitations on parallel downloads, we want to reduce http calls
There is so much to cover, and I want to make sure that everyone gets something valuable out of this talk, whether you are just starting out, or you work for Groupon, serving mils of hits a day.
My main objective is to help you get started. To know where to look, which tools to use, which techniques to investigate, and where to find the resources you need.
I won't be able to teach you everything about performance, but I can at least get you excited, and point you in the right direction.
If i don't teach you something new, or you don't get what you were looking for, please come up and talk to me. We can chat afterwards, if you have a specific issue, I'm sure I can help you out.
no instrumentation
Rawk
Pl_analyze
Splunk
There is so much to cover, and I want to make sure that everyone gets something valuable out of this talk, whether you are just starting out, or you work for Groupon, serving mils of hits a day.
My main objective is to help you get started. To know where to look, which tools to use, which techniques to investigate, and where to find the resources you need.
I won't be able to teach you everything about performance, but I can at least get you excited, and point you in the right direction.
If i don't teach you something new, or you don't get what you were looking for, please come up and talk to me. We can chat afterwards, if you have a specific issue, I'm sure I can help you out.
Images from ihower presentation
Make sure you repeat their question
what we couldn't cover this time, but might make for the latter half of a full day's tutorial...