Efficient Memcached Usage

Lessons learned while implementing memcached on zvents.com

Tyler Kovacs

First and Foremost

• Typical memcache operation takes a fraction of a millisecond (for a local instance of memcached). Realistically, memcached optimization is one of the last things you'll need to do. Start with the low-hanging fruit which will have a much larger effect: optimize your database, enable page, action and fragment caching, etc.
• But that's no excuse not to plan ahead.

Get The Latest Software

• What versions are you running?
• memcached
• The memcached NEWS page implies that there haven.t been any releases since 2004.
• Make sure you're running latest memcached (1.2.0 released 09/2006)
• Lots of CPU and memory consumption fixes committed in 05/2006
• Ruby client libraries
• Run MemCacheClient instead of RubyMemCache API
• gem install memcache-client
• Make sure you're on latest version (1.2.0 as of 11/2006)

Watch Your Footprint

• run memcached -vv
• Analogous to watching SQL statements in development.log
• How often do you touch the cache per page request? You might be surprised.
• On a busy site, cache touches can be a bottleneck
• How many hits can memcached handle?
• Lots - benchmark it yourself. Here's a best-case benchmark:

CACHE["a"] = 1
n = 10000
get_total = Benchmark::measure{ for i in 1..n; CACHE["a"]; end }.real
set_total = Benchmark::measure{ for i in 1..n; CACHE["a"] = rand; end }.real
puts "#{n} GET requests in #{get_total} seconds: #{n/get_total} req/sec"
puts "#{n} SET requests in #{set_total} seconds: #{n/set_total} req/sec"

Watch Your Footprint (more)

• Things will be slower in production
• Caching more complex data types (ActiveRecord objects) that have higher marshaling costs
• memcached probably running on separate hosts (network overhead)

Memcached Statistics

• memcached collects useful performance and utlization statistics

pid	uptime
time	version
rusage_user	rusage_system
curr_items	total_items
curr_connections	total_connections
connection_structures	cmd_get
cmd_set	get_hits
get_misses	bytes_read
bytes_written	limit_maxbytes
bytes

• Refer to protocol docs for explanation: http://www.danga.com/memcached/apis.bml

Performance Monitoring

• Use Cacti (or something similar) to monitor statistics over time
• Cacti
• Memcached Cacti Template

• or roll your own using stats command from memcache-client_extensions

Plugins

• Zvents Tools
• A collection of plugins released while developing zvents.com. Many plugins provide features related to memcached.
• http://blog.zvents.com/2006/10/31/zvents_tools
• memcache-client_extensions
• Adds 3 commands to memcache-client API
• get_multi: retrieve multiple keys in parallel
• stats: return memcached performance and utilization statistics
• flush_all: empty all information stored in the cache

extended_fragment_cache

• In-process memory cache
• cuts cache touches in half for fragment caching. Fragment caching hits the fragment_cache_store twice for each cached fragment
1. The read_fragment method invoked in a controller to determine if a fragment has already been cached. e.g.,

          unless read_fragment("/x/y/z")
            ...
          end
        

2. The cache helper method invoked in a view that renders the fragment. e.g.,

          <% cache("/x/y/z") do %>
            ...
          <% end %>
        

• With this plugin, the second cache op is served out of in-process cache

extended_fragment_cache (more)

• Content Interpolation
• Easily substitute dynamic values into cached data
• Reduces complexity in views
• Reduces cache touches
• We can replace this:

<% cache("/first_part") do %>
  This content is very expensive to generate, so let's
  fragment cache it.
<% end %>
<%= Time.now %>
<% cache("/second_part") do %>
  This content is also very expensive to generate.
<% end %>
        

extended_fragment_cache (more)

• with this:

<% cache("/only_part", {}, {"__TIME_GOES_HERE__" => Time.now}) do %>
  This content is very expensive to generate, so let's
  fragment cache it.
  __TIME_GOES_HERE__
  This content is also very expensive to generate.
<% end %>
        

• also integrates time-based fragment caching from memcache_fragments plugin

  <% cache("/only_part", {:expire => 15.minutes}) do %>
    This content is very expensive to generate, so let's
  fragment cache it.
  <% end %>

custom_benchmarks

• Allows you to log arbitrary information to the summary line in the Rails default log

Completed in 0.07425 (13 reqs/sec) | Rendering: 0.02631 (35%)
| DB: 0.02668 (35%) | Search: 0.00000,0 (0%) 
| memcache: 0.00073,2 (0%)| PID: 27605 | 200 OK
[http://zvents.dyndns.org/events/show/355761?nowsr=1]

• Next release includes this default integration with memcache-client