[Kolab-devel] Cyrus Murder and kolab

[Matt Douhan]

On Monday 16 May 2005 14.28, Martin Konold wrote:
> Am Montag, 16. Mai 2005 14:13 schrieb christophe robert:
>
> Hi,
>
> > Do you use an external card like "Umem (cache)" to improve the writing .
> >
> > Is it possible to use disks "10000 RPM"  without  having a bottleneck  
> > (in Writing ) for this configuration
>
> Use a decent RAID level like RAID 10 and don't worry about the write
> performance too much.

 and when it goes big use different channels for different filesystems.

>
> Most access is read not write and most important the write IO latency is
> easily hidden by the OS while reading typically means that the user process
> has to really wait until the data arrives.
>
> According to my measurements having a HW write cache is a waste while
> having some hw cache in the raid controller for reading helps a little bit.

I am interested in these tests, the write capabilities for a busy kolab server 
we have found is not to be underestimated, yes IMAP synchs/reads are the 
majority but delivering 8-12000 emails an hour to disk takes its toll.

>
> In generell giving the host machine enough memory helps most as access to
> main memory is about 1-2 orders faster than access to the cache memory on a
> PCI card.
>
> Last but not least SCSI instead of ATA pays for a kolab server!
>

it sure does, but it is fair to say that large SATA arrays with controllers 
from 3ware is a good budget solution, we currently have 3 servers in place 
that uses SATA with great success, they are not at all busy but the 
installations are growing slowly.

rgds

Matt

-- 
Matt Douhan
www.fruitsalad.org
kolab + toltec + horde + kontact == success