Scaling an early stage startup

by Mark Maunder <mark@feedjit.com>

Why does performance and scaling quickly matter?
• Slow performance could cost you 20% of your revenue according to Google. • Any reduction in hosting costs goes directly to your bottom line as profit or can accelerate growth. • In a viral business, slow performance can damage your viral growth.

My first missteps
• Misconfiguration. Web server and DB configured to grab too much RAM. • As traffic builds, the server swaps and slows down drastically. • Easy to fix – just a quick config change on web server and/or DB.

Traffic at this stage
• 2 Widgets per second • 10 HTTP requests per second. • 1 Widget = 1 Pageview • We serve as many pages as our users do, combined.

Keepalive – Good for clients, bad for servers.
• As http requests increased to 10 per second, I ran out of server threads to handle connections. • Keepalive was on and Keepalive Timeout was set to 300. • Turned Keepalive off.

Traffic at this stage
• 4 Widgets per second • 20 HTTP requests per second

Cache as much DB data as possible
• I used Perl’s Cache::FileCache to cache either DB data or rendered HTML on disk. • MemCacheD, developed for LiveJournal, caches across servers. • YMMV – How dynamic is your data?

MySQL not fast enough
• High number of writes & deletes on a large single table caused severe slowness. • Writes blow away the query cache. • MySQL doesn’t support a large number of small tables (over 10,000). • MySQL is memory hungry if you want to cache large indexes. • I maxed out at about 200 concurrent read/write queries per second with over 1 million records (and that’s not large enough).

Perl’s Tie::File to the early rescue
• Tie::File is a very simple flat-file API. • Lots of files/tables. • Faster – 500 to 1000 concurrent read/writes per second. • Prepending requires reading and rewriting the whole file.

BerkeleyDB is very very fast!

• I’m also experimenting with BerkeleyDB for some small intensive tasks. • Data From Oracle who owns BDB: Just over 90,000 transactional writes per second. • Over 1 Million non-transactional writes per second in memory. • Oracle’s machine: Linux on an AMD Athlon™ 64 processor 3200+ at 1GHz system with 1GB of RAM. 7200RPM Drive with 8MB cache RAM.
Source: http://www.oracle.com/technology/products/berkeley-db/pdf/berkeley-db-perf.pdf

Traffic at this stage
• 7 Widgets per second • 35 HTTP requests per second

Created a separate image and CSS server
• Enabled Keepalive on the Image server to be nice to clients. • Static content requires very little memory per thread/process. • Kept Keepalive off on the App server to reduce memory. • Added benefit of higher browser concurrency with 2 hostnames.
Source: http://www.die.net/musings/page_load_time/

Now using Home Grown Fixed Length Records
• A lot like ISAM or MyISAM • Fixed length records mean we seek directly to the data. No more file slurping. • Sequential records mean sequential reads which are fast. • Still using file level locking. • Benchmarked at 20,000+ concurrent reads/writes/deletes.

Traffic at this stage
• 12 Widgets per second • 50 to 60 HTTP requests per second • Load average spiking to 12 or more about 3 times per day for unknown reason.

Blocking Content Thieves
• Content thieves were aggressively crawling our site on pages that are CPU intensive. • Robots.txt is irrelevant. • Reverse DNS lookup with ‘dig –x’ • Firewall the &^%$@’s with ‘iptables’

Moved to httpd.prefork
• Httpd.worker consumes more memory than prefork because worker doesn’t share memory. • Tuning the number of Perl interpreters vs number of threads didn’t improve things. • Prefork with no keepalive on the app server uses less RAM and works well – for Mod_Perl.

The amazing Linux Filesystem Cache
• Linux uses spare memory to cache files on disk. • Lots of spare memory == Much faster I/O. • Prefork freed lots of memory. 1.3 Gigs out of 2 Gigs is used as cache. • I’ve noticed a roughly 20% performance increase since using it.

Tools
• httperf for benchmarking your server • Websitepulse.com for perf monitoring.

Summary
• Make content as static as possible. • Cache as much of your dynamic content as possible. • Separate serving app requests and serving static content. • Don’t underestimate the speed of lightweight file access API’s . • Only serve real users and search engines you care about.