On Fri, 2006-02-03 at 10:39 +0100, Sebastian Hagedorn wrote: > --On 2. Februar 2006 22:36:25 +0100 Kjetil Torgrim Homme > <[EMAIL PROTECTED]> wrote: > > > we have an old installation, so we're using LVS (2 directors in > > active/active) in front of Perdition (2 frontends) in front of 16 Cyrus > > instances running on three nodes sharing a SAN. the cluster software > > handling the three nodes is HP ServiceGuard. > > Wow, that's a lot of nodes. I wonder if and why so many are necessary in > your case?
obviously the LVS directors are used for other services as well, and they aren't exactly sweating... but it keeps complexity down to run it on separate servers. during peak hours (8000 simultaneous IMAP users), a Perdition server would only barely be able to handle the load on its own, so we're actually bringing in an extra server to make sure we have true redundancy. the Cyrus nodes have lots of CPU to spare, it's really only during routine tasks at night that they run without idle time. still, it's very comfortable to be able to run on only two nodes without having to worry about performance. it is true that I/O is the biggest problem, and the hardest piece to scale up. we're using a HP EVA as the storage solution, but we share it with other uses (Oracle databases etc.) I guess the idea is that hardware is cheap. if it enables us to do maintenance during office hours, it saves labour costs. we also routinely do upgrades of kernel, firmware, etc. on just one of the nodes and run a single Cyrus instance on it (the one with e.g. my personal e-mail ;-) to do realistic testing before putting the new version into production. > We have a HA cluster using Red Hat Enterprise Linux 3 and its > Cluster Suite. There are just two nodes and only one is active at a time > (the other handles other jobs, though). We also use LVS but only for SMTP > and virus scanning (running on four nodes). We have roughly 40,000 accounts > and this one server handles them all quite well. Usually I/O is the > limiting factor for Cyrus, and that's where our SAN infrastructure is > invaluable. we have seven SMTP boxes handling incoming mail for our 77k users. this could quite clearly be optimised, but SpamAssassin is quite resource hungry if you're just passing it everything. so far, just bringing in an extra server has been considered cheaper than allocating time to do such optimisation. -- Kjetil T. (University of Oslo, Norway) ---- Cyrus Home Page: http://asg.web.cmu.edu/cyrus Cyrus Wiki/FAQ: http://cyruswiki.andrew.cmu.edu List Archives/Info: http://asg.web.cmu.edu/cyrus/mailing-list.html
