Miserable performance of cyrus-imapd 2.3.9 -- seems to be lockingissues

[Jeff Fookson]

Vincent Fox wrote:

>David Lang wrote:
>  
>
>>raid 6 allows you to loose any two disks and keep going.
>>
>>  
>>    
>>
>
>This is turning into a RAID discussion.
>
>The orginal poster was doing a RAID-5 across 3 disks, and has stopped
>commenting but it's probably because that's all the hardware he could 
>scrounge.
>
>
>
>  
>
The "original poster" (me) has not commented further on the RAID 
discussion that has evolved
from my initial questions because we are in fact limited in what we can 
afford. The points on the
relative crappiness of RAID5 have been very illuminating and useful, 
however.

Yesterday, we moved the cyrus metadata off the md->drbd->lvm->ext3 
partition onto a vanilla ext3 one
(in fact, on the system PATA disk) and have seen modest improvement as 
measured by average system
load and user contentment. Our intent is next to junk the RAID5 and go 
with a simple mirrored
pair of 500Gb SATA disks for the mailspool and a second mirrored pair 
for the metadata. We had wanted
the potential for automatic fallover with drbd but will scrap that for 
now. We are planning to run the mirrors
off a 4-port 3ware RAID card even though we're not overly fond of 3ware 
(we have a fair amount of experience
with RAID5 arrays on 3ware cards on our research  machines where they 
perform adequately but
not more). We are hoping the 3ware RAID1 will be a bit better than 
software RAID1.

I would also like to comment at this point that the lvm2 layer in our 
original design was so that we could use
snapshots to insure a static filesystem for backups. So a secondary 
question concerns how much potential
troubles we might have just backing up (with rsync) the active system?

Thanks again for everyone's insight and suggestions. Once the production 
system is working better and users
have been pacified, we'll continue to tinker with the lvm and drbd 
layers on a pair of test machines, but without
RAID5. I don't really think those other layers are the show-stoppers, 
although it is certainly true that the
combined effects of the incremental inefficiencies of each layer might 
have had cumulative impact.

I note in passing that we had tested the original design by pounding the 
pair of machines with incoming mail
and that did not reveal the gross difficiencies that we saw when we 
started accepting user imap connections as well.
Can someone in fact suggest a good battery of testing software so that 
we don't prematurely certify that a new system
is ready for production use (or a useful statistic to monitor)?

Jeff Fookson


-- 
Jeffrey E. Fookson, PhD			Phone: (520) 621 3091
Support Systems Analyst, Principal	jfookson at as.arizona.edu
Steward Observatory
University of Arizona