ruby-dtracers: A D repository from eac

A collection of dtrace scripts and analysis tools for use with dtrace on Ruby 2. Contributions welcome.

Scripts

Allocate.d

Details object allocations:

   28702 Array(3) ruby-2.0.0-p195/gems/mongo-1.7.1/lib/mongo/util/core_ext.rb:46
   33000 Hash(0) ruby-2.0.0-p195/gems/treetop-1.4.14/lib/treetop/runtime/compiled_parser.rb:66
   33600 String(4) ruby-2.0.0-p195/gems/activerecord-2.3.18/lib/active_record/attribute_methods.rb:1
   35868 Array(3) ruby-2.0.0-p195/gems/active_record_shards-2.7.4/lib/active_record_shards/shard_selection.rb:37
   41150 Array(1) ruby-2.0.0-p195/gems/mail-2.5.4/lib/mail/header.rb:273
   85400 String(1) ruby-2.0.0-p195/gems/treetop-1.4.14/lib/treetop/runtime/compiled_parser.rb:100
  100042 Array(2) ruby-2.0.0-p195/gems/treetop-1.4.14/lib/treetop/runtime/compiled_parser.rb:110
  186065 Array(3) ruby-2.0.0-p195/gems/activesupport-2.3.18/lib/active_support/memoizable.rb:32
  186065 String(17) ruby-2.0.0-p195/gems/activesupport-2.3.18/lib/active_support/memoizable.rb:32

Flow.d

Outputs a simple human readable call flow graph.

  -> OpenSSL::Buffering`read_nonblock
   -> OpenSSL::SSL::SSLSocket`sysread_nonblock
    -> Class`new
     -> Exception`initialize
     <- Exception`initialize
    <- Class`new
    -> Exception`exception
    <- Exception`exception
    -> Exception`backtrace
    <- Exception`backtrace
   <- OpenSSL::SSL::SSLSocket`sysread_nonblock
  <- OpenSSL::Buffering`read_nonblock

Method Count.d

Counts all method calls.

     149 Net::BufferedIO readuntil /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/protocol.rb:131
     156 String index /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/protocol.rb:133
     165 Net::BufferedIO rbuf_consume /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/protocol.rb:171
     165 String slice! /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/protocol.rb:172
     236 String [] /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/http/response.rb:56
    1980 Symbol == /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/http.rb:885

Method Time.d

Measures CPU time per method.

      27 OpenSSL::SSL::SSLSocket connect /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/http.rb:918
     147 OpenSSL::X509::Store set_default_paths ruby-2.0.0-p195/gems/faraday-0.8.8/lib/faraday/adapter/net_http.rb:104
     147 Faraday::Adapter::NetHttp ssl_cert_store ruby-2.0.0-p195/gems/faraday-0.8.8/lib/faraday/adapter/net_http.rb:106
     147 Faraday::Adapter::NetHttp configure_ssl ruby-2.0.0-p195/gems/faraday-0.8.8/lib/faraday/adapter/net_http.rb:98
     408 TCPSocket initialize /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/http.rb:878
     408 IO open /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/net/http.rb:878
     436 Timeout timeout /rubies/ruby-2.0.0-p195/lib/ruby/2.0.0/timeout.rb:52

  <off cpu> (IO)                                                19071

Analysis

Browsing Allocations

Object allocations tend to point to the most expensive Ruby code. ./browse_allocations simplifies exploring allocation data.

 Usage:
./allocate.d -p <pid> -o cached.out

 APP_PATH=zendesk/zendesk ./browse_allocations cached.out
 Options:
 --package: Allocations by package
 --file: Allocations by file
 --line: Allocations by line (most detailed)
 --size: Allocated size per package

 Examples:

 APP_PATH=zendesk/zendesk ./browse_allocations cached.out --package
 Packages by allocation count
 ----------------------------
 93976 active_record_shards-2.7.4
 110082 actionpack-2.3.18
 460727 treetop-1.4.14
 547922 activerecord-2.3.18
 590897 activesupport-2.3.18
 880476 mail-2.5.4
 Total: 3419286

Flamegraph

Flamegraphs are a good starting point when trying to get a sense of where the application is spending its time. In the following example, we can see that most of Faraday::Connection#get time is spent making an HTTP call.

 Usage:
./flamegraph.d -p <pid> -o http.out
GRAPHER=~/FlameGraph/flamegraph.pl ./analyze/generate_flamegraph http.out

Due to inaccurate return probes in Ruby (or perhaps bugs in this script) the output isn't 100% accurate. It does give a fairly good sense of the most heavily used code paths, and is a good guide for where to focus more accurate profiling methods (e.g. object allocations).

Tips

Often the biggest wins will come from reducing object allocations.
For applications, try to mimic your production environment as much as possible. This includes things like setting the correct log level, sending a couple warmup requests, and making sure your app isn't in some error state (like throwing a bunch of memcached errors).

eac/ruby-dtracers