fc5 kernels after 1788 have excessive, kernel overhead
davej at redhat.com
Sat Dec 31 20:30:46 UTC 2005
On Sat, Dec 31, 2005 at 08:30:13PM +0000, Andy Burns wrote:
> Anne & Lynn Wheeler wrote:
> >long running application doing enet to disk transfer sustains
> >2.6mbytes/sec transfer with 20% cpu busy (mostly non-kernel time).
> Even that seems pretty low, I'd have thought you should be able to
> sustain 10-12MB/s
> >post-1788 kernel, it is 100% cpu (mostly kernel time) and 600kbytes/sec
> >sustained (an increase of over 20times in total cpu executed per byte
> >transferred and possibly an increase of 100times in kernel cpu executed
> >per byte transferred).
> Just did a quick test here, a http wget of a 75MB file across a 100Mb/s
> full duplex link, where the file is fully cached on the source machine,
> I scrape 1.8MB/s and eat 75% CPU on a 3.0GHz HT P4, if I do some
> "intensive" monitoring with top, I can get it down to about 900KB/s
> Of course we're not really proving anything here, we know that
> kernel1805 is built for debugging not for speed ...
my people page will have a new kernel with it turned off in about
an hour (buildsys permitting).
More information about the test