Benchmarking

Matthew Dillon dillon at apollo.backplane.com
Mon Mar 8 10:55:47 PST 2004


:Some time ago I had a plan to benchmark several OS'es for fun&learning. As 
:it turned out, due to the lack of time the process was very slow. I intend 
:to do DragonFlyBSD next, so I'm asking: what would be the best way to 
:install it? Download the bootable ISO (from 
:http://www.dragonflybsd.org/main/download.cgi) and then upgrade to -current 
:via cvs? Is there a somewhat step-by-step walkthrough of the process since, 
:as I understand it there is no install/setup program?
:
:Also, if anybody has some performance-related patches, now would be a good 
:time to commit them to CVS :)
:
:(Current results are in http://geri.cc.fer.hr/~ivoras/osbench/osbench.sxc; 
:Summary: FreeBSD 5.2 is slower in synthetic tests but appears surprisingly 
:better in a "real-world" test than 4.9).

    Pretty comprehensive.  The numbers are about what I expect.  5.x will
    do better with pure userland cpu tests simply because it has a more
    recent GCC installed, and will do worse on benchmarks requiring lots
    of kernel interaction (like small-block writes) due to higher kernel
    overheads.  The 5.x pipe code is very well optimized so the results
    there make sense, and the 5.x network code has a number of fixes in it
    which 4.x never got.

    I think DFly is in pretty good shape performance-wise but there is
    at least one mbuf related patch we ought to throw in.  Also, the
    recent token code work has not been well optimized yet but I don't know
    how much that will impact things relative to what we had before.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>





More information about the Kernel mailing list