From: Masaki Matsushita <glass.saga@...>
Date: 2012-02-21T13:47:15+09:00
Subject: [ruby-core:42772] [ruby-trunk - Feature #5995] calling io_advise_internal() in read_all()


Issue #5995 has been updated by Masaki Matsushita.


Some file systems(e.g. ext3, ext4) use do_sync_read() for general read.
http://lxr.linux.no/#linux+v3.2.6/fs/ext3/file.c#L55
http://lxr.linux.no/#linux+v3.2.6/fs/ext4/file.c#L231

In read process, do_generic_file_read() is called finally.
http://lxr.linux.no/#linux+v3.2.6/fs/read_write.c#L338 ( do_sync_read() )
In ext3 and ext4, f_op->aio_read is generic_file_aio_read().
http://lxr.linux.no/#linux+v3.2.6/mm/filemap.c#L1395 ( It calls do_generic_file_read(). )

Then, do_generic_file_read() calls page_cache_sync_readahead() or page_cache_async_readahead().
http://lxr.linux.no/#linux+v3.2.6/mm/filemap.c#L1118

Both page_cache_sync_readahead() and page_cache_async_readahead() call ondemand_readahead() and its readahead size is limited by ra->ra_pages.
http://lxr.linux.no/#linux+v3.2.6/mm/readahead.c#L401

posix_fadvise() expands ra->ra_pages( http://lxr.linux.no/#linux+v3.2.6/mm/fadvise.c#L90 ) and it reduces the number of times of actual read.
Therefore, I think the patch makes sense on some file systems as stated above.
----------------------------------------
Feature #5995: calling io_advise_internal() in read_all()
https://bugs.ruby-lang.org/issues/5995

Author: Masaki Matsushita
Status: Rejected
Priority: Normal
Assignee: Motohiro KOSAKI
Category: core
Target version: 


=begin
I propose to call io_advise_internal() in read_all().
It will increase performance.

I created a dummy file:
 dd if=/dev/zero of=dummy bs=1M count=100

Then, I ran the following:

 require 'benchmark'
 
 Benchmark.bm do |x|
   x.report do
     f = File.open("dummy") # dummy file(about 100MB )
     f.read
   end
 end

I freed page cache before each test:
 sudo sysctl -w vm.drop_caches=1

results on Ubuntu 11.10(3.0.0-15-server):

r34462:

 user     system      total        real
   0.050000   0.220000   0.270000 (  0.356033)
 
       user     system      total        real
   0.050000   0.190000   0.240000 (  0.332243)

       user     system      total        real
   0.060000   0.210000   0.270000 (  0.347758)

patched ruby:

 user     system      total        real
   0.030000   0.130000   0.160000 (  0.225866)

       user     system      total        real
   0.040000   0.170000   0.210000 (  0.250172)

       user     system      total        real
   0.040000   0.150000   0.190000 (  0.254654)

It shows the patch increases performance.
=end


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