From: John M. <jm...@iv...> - 2002-05-24 14:53:34
|
> In our current situation, we have approximately 500 users, with up to > 200 users at any one time hitting the web server. Each apache process > is using between 10 and 20MB of RAM each, and the CPU is usually 95% > idle or higher. Yeah, this is definitely an interesting situation. I've got a rather lengthy story about my college's run-in with this, but I'll try to condense it down here... We've got about 2600 users in 24 campuses across Indiana. In short, we never know exactly what's going to happen until it does. :) SM is providing us with excellent web mail, but who knows how well the load is going to be handled when everyone gets migrated over? This past Monday, we found out - the hard way. We were running Courier-IMAP, Postfix, PHP 4.1.2, SM 1.2.6 (+in-house mods), Apache 1.2.3, and Linux (kernel 2.4.14), all on a dual PIII/1gig with 1GB ram, 100GB scsi raid-5. All went well on Monday until about 11am, when several hundred users decided to hit web mail at once. I saw the load avg go up to about 20 before the box became unusable. Sometime later on, it did a mostly-crash, and I ended up rebooting it. Ew, I felt so dirty... Fortunately, we had planned for a rush of users moving from LookOut to web mail, and had an LVS cluster (linuxvirtualserver.org) contingency plan. For anyone considering having lots of users, I highly recommend the solution. A single PIII/700 acts as the load balancer and failover controller, 3 PIII/1gig+1GB Dell 1550's are the web "farm." They NFS export all of SM and the PHP sessions directory from the mail server itself, then do the IMAP work to the mail server itself as well. LVS works by transparently passing incoming tcp requests off to a cluster of machines based on queuing decisions (round robin, weighted least connections, etc.). The whole system works quite well, and handles several hundred users (over SSL) without so much as a hiccup. We should be able to easily scale to move all of our users to web mail, even if we have to add more nodes to the cluster. Points to remember: - Lots of SSL is bad on your cpu's. - Pick a good IMAP daemon. Anything that uses mbox storage is going to roll over and die very quickly.- Again, Maildir is good. You can even serve it up over NFS if you need to. - If you can, split the IMAP and web server load onto separate machines, whether you're technically clustering or not.- If you're doing lots of virus scanning, mailing lists, or other apps on your SMTP box, you may want to split that off from the IMAP server as well. Hope that helps... John -- John Madden UNIX Systems Engineer Ivy Tech State College jm...@iv... |