Things to Know About Web Performance

An overview of performance related topics to make systems fast and scalable.

Transcript

1. PERFORMANCE! / Michael Hamrah @mhamrah Software Engineer @ Getty Images
2. PERFORMANCE VS. SCALABILITY VS. RELIABILITY
3. TALKING POINTS Networking and TCP Web Servers App Servers Rails Caching Databases Front End Code Iaas (Amazon) vs PaaS (Heroku) Example Architectures
4. WHY IS THIS IMPORTANT? Converstion Rates Attention Span, Frustration Levels Trust and Web Performance Lessons Superbowl Website Failures
5. IT'S NOT ALL ABOUT YOUR CODE! Time spent in your code? Overall percentage?
6. NETWORKING! IT'S MOSTLY ALL ABOUT NETWORKING What happens when you type an address in your browser? The TCP/IP Stack: Data Link, Network, Transport, Application (also Physical, Session, Presentation) This is called the OSI model.
7. TCP: HOW DATA FLOWS FROM A TO B A node sends "packets" Packets must be of a certain size (~1500 bytes, 64k max) Only a certain amount of packets are allowed "in flight" Packets are guaranteed delivery, in-order (different than UDP) Starts with a Three-Way handshake. Syn -> Syn Ack -> Ack
8. HOW LONG DOES THIS TAKE? NYC to SFC: ~2530mi, RTT in vacuum: 27ms, RTT in fiber: 41.04ms IP and Routes add delay. (Check out Ping and Traceroute) Remember the three way handshake? Latency. Bandwidth may not matter Connections are scarce, precious resources! (Why we have connection pooling)
9. HOW DOES HTTP FIT IN? Application level protocol Defines how requests and responses sit on top of TCP Http 1.1: Keep-alive, pipelining Browsers 6-8 connections per server Think of 1mb of data split 1, 10, 100 ways
10. Most client-server optimization is about tuning for network latency and managing TCP connections. Networking isn't just about a browser and a server. Intra- and inter- data center connections are important! Optimize for throughput: keep data small, reduce latency, reduce connections.
11. WEB SERVERS! Web servers like nginx and apache handle http requests. Application servers run programs which produce dynamic content. Think Thin, Unicorn, Puma, Node, etc. Sometimes it's a blurry line. Depends on architecture.
12. THREADED VS. EVENTED How connections are mapped to requests How threads are mapped to requests Memory usage and context switching How does your OS help?
13. ALL ABOUT ASYNC Non-blocking io is essential in all cases EventMachine, Node, Twisted: Event loops Heroku's big problem with Rails
14. A TYPICAL SCENARIO LOAD BALANCER -> WEBSERVER -> APP SERVER Load Balancer Distributes Requests Web Servers (nginx/haproxy/varnish) handles connections, static content Application Servers or Process Pools respond to requests
15. SO WHAT DO WEB SERVERS ACTUALLY DO? Very good at handling http connections Parse http requests Add filters like gzip, authentication Buffering, Caching Reverse-Proxy support
16. DIFFERENCES WITH APPLICATION SERVERS (AND RUNTIMES!) Allow you to program dynamic content Thin, Unicorn, Puma, Passenger MRI, Rubinus, JRuby Node vs. Rails vs. Django vs. Whatever?
17. RACK A common interface between Ruby web servers and web frameworks. Provides simple http primitives Supports middleware Can mix Rails and Sinatra endpoints
18. DATABASES! They're beasts! Sql - ACID NoSql - BASE Both have their own set of problems.
19. TUNING DATABASES In both cases, it's about indexes. Learn to find slow queries. Avoid n+1 queries (go for eager loading). Optimize for read/write patterns.
20. SCALING DATABASES How do you add more nodes? Again, understand read/write patterns. Understand latency.
21. POLYGOT DATA LANDSCAPE Postgres, Mysql Cassandra, Riak, HBase MongoDb Redis Memcache
22. CACHING! Html Caching Fragment Caching Object Caching How easy is it to expire your cache?
23. HTML CACHING Uses http headers ETags, Cache-Control, Expires Anonymous vs. Authenticated Users How are you expiring cache?
24. FRAGMENT CACHING Still Html! More surgical ESI make tools like Varnish quite powerful
25. OBJECT CACHING Most common Most flexible Easy db integration Memcache for reads Redis for structured data Compress (CPU cycles are there) You're just denormalizing data
26. QUEUES! THEY'LL BE YOUR BEST FRIEND. Very Async Great for handling spikes Offload low-priority requests Great for elastic workers Resque, IronWorker, Amazon SQS
27. CLIENT-SIDE TECHNIQUES Sprites Minification Domain Sharding cache-control Async javascript Image management (WebP) CDNs!
28. PAAS VS. IAAS How do you want to run your apps? What control do you want or need? Other AWS Services: ElasticCache, R53/ELB, ASG, Search, CloudFormation, OpWorks, Multi-AZ, Multi-Region Cloud Services: Most tech is offered as SaaS NewRelic, Loggly
29. GOOGLE'S SPDY One persistent SSL connection True Duplex, no waiting Better window scaling Compressed headers (cookies, etc) Content Prioritization Server Push
30. EXAMPLES How would you design a news site? How would you design an e-commerce site? How would you design a social site?
31. PROBLEM The server request/response time is very long.
32. PROBLEM The server request/response time is very fast, but when more load increases, times increase dramatically.
33. PROBLEM A page takes a long time to load, but the server time is very fast.
34. PROBLEM We are sending images to a bunch of customers. Sending images is very slow, but nothing is maxed out. The problem increases when we try and send more images.