Issue #9262 has been updated by tmm1 (Aman Gupta).


I ran some benchmarks with funny-falcon's patch. Memory usage is increased by 5-10mb RSS in our app, and response time is improved by 4-8%.

before:

  500 requests to https://github.com/dashboard as tmm1
      peak memory:  349.4 MB RSS
         cpu time:  7,171ms total                  (14ms avg/req,  13ms - 21ms)
  500 requests to https://github.com/github/github/pull/17695 as tmm1
      peak memory:  404.6 MB RSS
         cpu time:  98,713ms total                 (197ms avg/req,  184ms - 285ms)

after:

  500 requests to https://github.com/dashboard as tmm1
      peak memory:  359.2 MB RSS
         cpu time:  6,806ms total                  (13ms avg/req,  12ms - 20ms)
  500 requests to https://github.com/github/github/pull/17695 as tmm1
      peak memory:  410.3 MB RSS
         cpu time:  95,559ms total                 (191ms avg/req,  177ms - 280ms)


----------------------------------------
Bug #9262: global_method_cache should be configurable or grow automatically
https://bugs.ruby-lang.org/issues/9262#change-43840

Author: tmm1 (Aman Gupta)
Status: Open
Priority: Normal
Assignee: ko1 (Koichi Sasada)
Category: 
Target version: next minor
ruby -v: trunk
Backport: 1.9.3: UNKNOWN, 2.0.0: UNKNOWN


The global_method_cache is currently a fixed 2048 entries. This is not configurable, and can be inadequate for large applications with thousands of classes and methods.

In our app, increasing the size of the cache to 32768 entries reduced time spent in search_method and overall pressure on st_lookup:

before
      420  14.0% st_lookup
      182   6.1% vm_search_method (inline)

after
      265   9.5% st_lookup
      125   4.5% vm_search_method (inline)

It's possible the VM could grow global_method_cache automatically, using some heuristic based on the number of long-lived classes or method_entries that are defined.
However this would break the hashing in the current implementation.


-- 
http://bugs.ruby-lang.org/