Web performance mercadolibre - ECI 2013

Techniques, tips and tools for improve
and measure web performance.
Web Performance
MercadoLibre
Santiago Aimetta
Nicolas Brizuela

Why performance?
Reducing time to response, impact
directly in your revenues.
Impact directly in the bounce rate,
conversions rate and is very important for
user experience.

Some numbers of Meli Search
● 75MM searches/day (870 searches/second)
● Peak traffic 102k rpm (1.700 searches/second)
● Avg response time: 320ms

Amazon test 2008
● + 100ms >> -1% sales
Bing test 2009
● + 2000ms >> -4.3% revenues/user
MercadoLibre 2013
● + 3000ms >> + 3% in Bounce rate
-1% in Revenues

Performance golden rules
● 80-90% of the end-user response time is spent on the
frontend.
We start there
● Greater potential
● Simple
● Proven to work

What is this?
Is the amount of time between the client
makes an HTTP request and the browser
starts receiving the first byte.
How much time is spent making the
request until receive the first byte of the
response.

1. DNS LOOKUP
2. INITIAL CONNECTION
3. WAITING
4. RECEIVING DATA
5. CLOSING CONNECTION

1. DNS LOOKUP
2. INITIAL CONNECTION
3. WAITING
4. RECEIVING DATA
5. CLOSING CONNECTION
<< Time to first Byte = TTFB

● Static content
○ Such as Html, Js, Css and images
○ Should be under 100 miliseconds
● Dinamic content
○ Includes all the server side processing plus the
network infrastructure work
○ Should be beetween 200 and 500 miliseconds

Possible problems
● To many connections to the server
● Disk IO

Hardware check..
● Disk IO
● RAM usage
● Swap usage
● Network bottlenecks

Configuration check..
● Webserver config (Apache,Jboss,..) / Php config
● Database settings
● Network settings
● Api / webservices latency

Hardware
● CDN- Content delivery network (Akamai,
CloudFront, BitGravity)
● Multiple servers with load balancing ( f5 , nginx )
● NAS - Filers ( T-com, IBM, HP)
● Web caches ( Varnish, Polipo, Squid, TrafficServer )

Software
● Parallel processing
● Database tuning
● Sql tuning
● API / Webservices response caching
● NoSql (MongoDB, Bigtable, Redis)
● Chunking - Early buffer flush

Analyzers
● Google Page Speed
● WebPageTest
● Firebug

Custom measuring
● Navigation timing api
var timing = window.performance.timing;
var ttfb = timing.responseStart -timing.connectEnd;

CDN (Content Delivery Network)

● Group of servers distributed in multiple datacenters
across the internet
● The CDN serves the content using the servers that
are closer to the client
● The network latency is reduced by the proximity
between client and server.

● The resources can be cached
● Multiple servers prevent bottlenecks
● Useful for static resources like Html, Css, fonts , Js,
videos , images, documents, etc

● Type of http compression like deflate
● This saves bandwidth and increases speed.
● Web client (i.e Browser) sends an Accept-Encoding :
gzip, deflate header
● Web server responds Content-Encoding : gzip if the
data is compressed

● Reduce the 70%-90% of the response size
● Use in Html, Css, Js, Xml, Json
● Dont use in Pdf and images, they are already
compressed
● Better compression tips:
○ Sorted key values : Css, html attributes
○ use one type of quotes, " or '
○ Css and Js minification

Benefits
● Saves requests to resources that changes
infrequently.
● HTTP caching saves the resources in the
browser or the proxy.
● Should be cached: CSS, JS, Static HTML, Images,
Flash, Pdf, media files.

Response Headers
● Strong ones:
○ These headers express the resource lifetime.
○ The value is a date or a timestamp.
○ A resource is downloaded again when the
expiration date is reached.
○ Expires and Cache Control.

Response Headers
● Weak ones:
○ Specifies characteristics to identify if the
resource change
○ The browser sends conditional GETs to check
○ Last-Modified, Etag

Cache Control
● Strong header
● Cache-Control:public
● Cache-Control:private
● i.e: Cache-Control:public, max-age=3600

Expires
● Sets an expiration date in the future.
● if Cache-control and expires are set for the same.
resource Cache-control takes precedence.
● i.e: Expires: Mon, 8 Jul 2013 21:31:12 GMT.

Last modified
● Is a time based header.
● The application specifies the last modified header
i.e: Last-Modified: Tue, 09 jul 2013 17:45:57 GMT.
● The next time the browser sends a conditional GET
asking if the resource has changed
i.e If-Modified-Since: Tue, 09 jul 2013 17:45:57 GMT.
● If the resource hasn't changed the server return an
empty response with the 304 code (Not Modified)

Etag
● Use an md5 hash to identify if the resource change.
ETag: "15f0fff99ed5aae4edffdd6496d7131f".
● In the next request the header If-None-Match is sent
with the ETag value
i.e: If-None-Match: "15f0fff99ed5aae4edffdd6496d7131f"
● If the ETag match, the server responds 304

Tools
● Most of browser tools has a network analyzer
● The example below were made with Chrome dev tool

Tips
● For static content: use Cache-Control.
● Cache-Control is easy to check.
● Avoid conditional Gets.
● Use the app version or a fingerprint in the url.

Tips
● For private content: use Cache-Control :
private to avoid proxy caching.
● Prevent caching: use Cache-Control:no-cache,
no-store.
● Urls with query string.

Keep alive (reuse connections)

● Client and server keep the connection open,
unless the client indicates otherwise (via
Connection: close header).
● Http connections are expensive.
● Saves TCP handshake ( 150 ms average ).

● Persistent connections send multiple request and
response interactions over single connection.
● If the connection is not persistent you can specify a
time out.

Advantages
● CPU & memory savings, less tcp connections and
fewer TCP control blocks.
● Allows request and response pipelining.
● Reduce network load, less packets sent.
● Supported by modern browsers.

Parallel downloads
The biggest impact on end-user response
times is the number of components in the
page

How it works?
● Images can be downloaded in parallel
● JS and CSS..other story

● Loading steps
○ downloading (can be parallel )
○ parsing
○ executing
● Rules
○ Scripts prevents other scripts to be downloaded
and parsed
○ Stylessheets prevent scripts to be downloaded
and parsed
○ Modern browsers start looking ahead in the
document and pre-loading stylesheets and scripts

The HTTP/1.1 RFC
A single-user client SHOULD NOT
maintain more than 2 connections with
any server or proxy.

IE 6 and 7: 2
IE 8: 6
IE 9: 6
IE 10: 8
Firefox 2: 2
Firefox 3: 6
Firefox 4 to 17: 6
Opera 9.63: 4
Opera 10: 8
Opera 11 and 12: 6
Chrome 1 and 2: 6
Chrome 3: 4
Chrome 4 to 28: 6
Safari 3 and 4: 4
How browsers handle it?
● Browsers don't have to follow
this guideline.
● Parallel connections.

Nice trick!
● The number of parallel connections applies to a
server.
● Use multiple domain names
○ i.e resources1.domain.com, resources2.
domain.com
○ Expands per server connection limit.
○ If the domains are CNAMEs of the same ip,
works too!

Trade off
● DNS lookup ~ 150 ms
● Browser cpu per parallel download
● Bandwidth

Reduce client request time by reducing the
request size
Small request

Objective
● Minimize the request overhead
● Cut down on client request time by reducing the
number of bytes uploaded as request header data
● Average request size is 1500 bytes.

How
● Keeping cookies and request headers as small as
possible ensures that an HTTP request can fit into a
single packet.
● Small urls.
● Small cookies.
● Remove unused header.

For static content
Cookieless domain

Static content
● Objective:
○ If you set a cookie in particular domain, all
subsequent HTTP requests for that domain must
include the cookie.
○ Static content, such as images, JS and CSS files,
don't need to be accompanied by cookies.
○ Avoid caching user info.

Static content
● How:
○ Create a domain for static content
○ Use caching headers
○ CDNs avoid cookies

Use browser idle time
Prefetching

Objective
● Use the browser idle time to download or
prefetch documents that the user might visit in
the near future.

Which content prefetch?
● Images commonly used.
● The next page of the search results.
● Prefetch common DNS.

● Image:
● Full page
● DNS
● Be aware of
○ Bandwidth, website statistics

Where and why
Javascript load & execution

Where and why?
● In the head: RUM, analytics
● before </body>: scripts needed by page load
● After page load: scripts needed soon after
page load
● On demand: In reaction to users

With and without blocking
Javascript loading

Motivation
● Scripts blocks downloads and render.

Several ways avoid it
● XHR Eval
● XHR Injection
● Script in Iframe
● Script DOM Element
● Script Defer

Avoid blocking - Simple approach

Avoid blocking - adding a callback

Avoid blocking - cross browser

Avoid blocking - onload event
● Blocks onload event until the script have been
downloaded and executed
○ script defer
○ script async
○ script dom element
● Fix
○ If you want to ensure that the JavaScript doesn't
start to download or execute until after the load
event, you can insert it using the window.onload
event handler:

Sometimes the image weight its 40-50% of
the complete page weight
Image Optimization

Lossless optimizations
● Are those that take an image and produce another
image, which renders exactly the same and it's
smaller in file size than the original
● The lossless file size savings come from:
○ Using better compression algorithms to store the
pixel information.
○ Removing unneeded metadata that goes with the
image file.

GIF
● The best way to optimize a GIF image is to convert it
to PNG8.
● It can store up to 256 colors, just like GIF.
● PNG8 supports alpha transparency.
● Software:
○ Photoshop
○ OptiPNG
● Animated GIF
○ Don't convert to PNG
○ Software:
■ GIFSicle

JPEG
● Edit image metadata
○ Software: JPEGTran, EXIFTool
● Optimizing compression
○ Software: JPEGTran
● Cropping
○ Rotation to 90, 180, 270 degrees

PNG
● Icons, illustrations and photos with high contrast.
● Support transparency (alpha channel).
● Optimizations
○ Strip PNG chunks
○ Better pixel compression
● Software
○ TinyPNG
○ OptiPNG
○ PNGOptimizer

● Stoyan test over 1000 sites
○ Convert GIFs to PNG ( -23% )
○ PNG optimization tools ( -17% )
○ Run JPEGTran on all JPEGs ( -13% )
○ Optimize animations with GIFSicle ( -4% )

JPEG Progressive
Perceived speed is more important that actual speed

JPEG Progressive
● Two types of images, baseline and progressive
● Baseline jpeg: is a full-resolution top-to-bottom scan
of the image

JPEG Progressive
● Progressive jpeg: is a series of scans of increasing
quality, loads from low quality to high in several
"passes"

JPEG Progressive
● The progressive jpeg’s first pass is low-resolution, but
it contains as much information, or more, as the small
image
● Software:
○ jpegtran
○ jpegcrop
● Images of file size 10K and over have a better chance
of being smaller when using the progressive JPEG
format

WebP
A new image format for the Web

WebP
● What is?
○ is a new image format that provides lossless and
lossy compression for images on the web
● 26% smaller than PNG
● 25-34% smaller than JPEG
● Supports transparency ( alpha channel )

WebP
● Software
○ CwebP
○ DwebP
○ libwebp
● Support
○ Chrome 9+
○ Opera 12+
○ Android 4+
○ Opera mobile 11+
○ Chrome for android 27+

Links
● http://www.rackaid.com/resources/time-to-
first-byte/
● http://jackwhitey.hubpages.com/hub/cdn
● https://developers.google.
com/speed/articles/gzip
● https://devcenter.heroku.
com/articles/increasing-application-
performance-with-http-cache-headers#http-
cache-headers
com/speed/docs/best-practices/caching

Links
● http://www.nczonline.
net/blog/2009/06/23/loading-javascript-
without-blocking/
● http://www.stevesouders.
com/blog/2009/04/27/loading-scripts-without-
blocking/
com/blog/2008/12/27/coupling-async-scripts/
● http://calendar.perfplanet.com/2010/the-
truth-about-non-blocking-javascript/

Links
● http://www.igvita.com/2011/10/04/optimizing-
http-keep-alive-and-pipelining/
● http://tools.ietf.org/id/draft-thomson-hybi-
http-timeout-01.html#rfc.section.2
● http://www.dcs.bbk.ac.
uk/~ptw/teaching/http/slide15.html
● http://www.dcs.bbk.ac.
uk/~ptw/teaching/http/slide16.html
● http://abdussamad.com/archives/169-
Apache-optimization:-KeepAlive-On-or-Off.
html

Links
● http://davidwalsh.name/javascript-domready
● http://api.jquery.com/ready/
● http://alexsexton.com/blog/2010/01/dont-let-
document-ready-slow-you-down/
com/speed/docs/best-practices/request
● http://www.nczonline.
net/blog/2009/05/05/http-cookies-explained/

Links
● http://www.catswhocode.
com/blog/mastering-html5-prefetching
● http://davidwalsh.name/dns-prefetching
● http://davidwalsh.name/html5-prefetch
● http://statichtml.com/2011/link-prefetching-
broken-in-chrome.html
● http://www.bookofspeed.com/chapter5.html
● http://calendar.perfplanet.
com/2012/progressive-jpegs-a-new-best-
practice/

Links
● http://caniuse.com/webp
● https://developers.google.com/speed/webp/?
hl=es-ES
● http://news.cnet.com/8301-1023_3-
57585114-93/google-cuts-network-usage-
by-terabytes-by-switching-to-webp/
● https://developers.google.com/speed/spdy/
● http://www.chromium.org/spdy
● http://googlecode.blogspot.com.
ar/2012/01/making-web-speedier-and-safer-
with-spdy.html

Links
● http://dev.chromium.org/spdy/spdy-best-
practices
● http://www.slideshare.net/nzakas/enough-
withthejavascriptalready
● http://ejohn.org/blog/browser-page-load-
performance/
com/blog/2008/03/20/roundup-on-parallel-
connections/
● http://www.yuiblog.
com/blog/2007/04/11/performance-research-

Web performance mercadolibre - ECI 2013

More Related Content

What's hot (17)

Similar to Web performance mercadolibre - ECI 2013 (20)

Recently uploaded (20)

Web performance mercadolibre - ECI 2013