very slow syncing, any ideas?

Michael Loftis mloftis at wgops.com
Thu Oct 19 19:11:32 EDT 2006



--On October 19, 2006 11:25:07 PM +0200 Marten Lehmann <lehmann at cnm.de> 
wrote:

> Hello,
>
>> Uhm... LA of 30 is very high.  What OS?  I assume Linux, vmstat 5 will
>> tell you where you're hitting the wall, but unless you've got an 8 CPU
>> machine LA 30 is rather quite high.  Linux LA is a measurement of
>> processes blocked on I/O, processes running and processes waiting to run
>> on a CPU.
>
> yes, Linux (2.6.9, RHEL4):

Looking at taht i'd say you're VERY badly CPU bound.  a simple dd/cp 
doesn't do anything to the mail but IMAP ops will require some CPU 
work....Cyrus also will probably be forcing syncs but your I/O load doesn't 
look that high (my mfe's run more I/O and they're not storing any mail, 
just logs and temporary files for virus/spam scanning heh, and they only 
have a little IDE HDD each)

the numebrs in the procs->r column indicate you've got a lot of processes 
vying for CPU time.  the b column indicates you've got a little bit of 
blocking going on.  Processes in the 'D' state may show up in ps or top 
output (doubtfully top unless you change the default sort).

I'd guess you're being CPU bound and cyrus is probably running (usually 
does by default) at a bit lower priority than your interactive cp 
operations.  2.6 kernel scheduler does a lot of weird stuff if it 
determines you're interactive, and then tends to give you priority over 
daemons.  it's very non-deterministic, but I don't think you're running 
into this.  quite likely some inefficiencies in the cyrus code (or 
intentional backoffs?) but mostly just being CPU bound.  I know our MBEs 
can generally scp or rsync quite a bit faster than they can perform a 
XFER/RENAME command that moves between mail store servers (IE MURDER host 
to MURDER host).




More information about the Info-cyrus mailing list