Bruce Guenter wrote: > > On Tue, Jan 15, 2002 at 05:26:55PM +0100, Martin Eriksson wrote: > > I was just wondering how Ext3 and Reiserfs compare. When I reinstalled my > > server (because of a stupid hacker) I took the opportunity to change to > > ReiserFS. And I have to say it's really much faster than Ext3. > > > > So what's some highlights on Ext3 vs. ReiserFS? I guess the Ext2 compability > > is one large factor for using Ext3, but otherwise? > > I ran some benchmarks recently to test performance of several Linux > filesystems under heavey synchronous load (ie a mail server). For this, > ext3 data=journal was nearly twice as fast as ReiserFS. > > See http://bruce-guenter.dyndns.org/benchmarking/ That's interesting - you've obviously done a lot of work there! The kernel version is a bit old. I think that 2.4.10 may have had broken fsync() behaviour as well, which may inflate the ext3 throughput. The way to get good synchronous throughput with ext3 is to ensure that the files are spread across lots of directories and lots of worker processes. This way, you can have lots of threads all running lots of synchronous operations in lots of dirfferent directories, and they all get coalesced into one operation. So a good mail spool setup is to hash the mailspool across 256 directories, and run 100 instances of the MTA. This is how postfix is normally used; not sure about qmail. If all the work is happening in one directory, or by one process then this parallelism is not possible.
