Date: Wed, 12 Mar 2003 07:08:02 -0500 From: "Troy Settle" <troy@psknet.com> To: <freebsd-isp@freebsd.org> Subject: RE: Mail Storage Reccomendations (3Ware vs Adaptec vs ....) Message-ID: <000601c2e890$08c973f0$aa8ffea9@abyss> In-Reply-To: <20030311235714.GA27853@dragon.stack.nl>
next in thread | previous in thread | raw e-mail | index | archive | help
> -----Original Message----- > From: Dean Strik [mailto:dean@stack.nl] > Sent: Tuesday, March 11, 2003 6:57 PM > To: Troy Settle > Cc: freebsd-isp@freebsd.org > Subject: Re: Mail Storage Reccomendations (3Ware vs Adaptec vs ....) > > > Troy Settle wrote: > > > > Besides NFS, this machine will also be running a MySQL > > server with 1 or 2 very small databases (for storing the > > vpopmail database). > > Whoa. "Besides NFS". So you export the mailboxes to other > hosts? That's an important difference. Yeah, I suppose it is for those folks who are still stuck on mbox-style mail storage. For those who have seen the light of Maildir/, the difference isn't that signifigant. I've been doing mail storage over NFS with woefully under-powered machines (celerons) for 2 years now without serious issue (except reading .cdb files over NFS). > Anyway, if disk performance is a problem, I'd > go with SCSI RAID10: high performance read, good performance write. > RAID5 means a performance penalty, even in hardware. With RAID10, > even software RAID would give acceptable performance: software RAID is > mostly expensive with RAID5. U320 SCSI probably isn't worth it. I never said that disk performance was a problem. Any admin would be stupid not to be concerned, but at my current volumes, I'd have decent storage performance with a Vinum RAID5 running on a P2 w/ATA66 drives. Anyways, I'm on UW SCSI now, and will be going up to U160. I've thought about U320, but agree that it's probably not worth it at the current volume of mail that I'm pushing. The cost of disks isn't a huge issue, in fact I'll probably end up with U320 disks. The actual controller I get will simply depend on the final solution I end up with. It's likely that I'll just go with the box w/onboard U160 and a ZCR solution. > If possible, make the sytem disks (and swap) independent from the mail > storage disks. The only processes on the server will sshd and mysqld, save for when I log in. The only major I/O will be with the NFS mounts. The box will be configured without softupdates, and with at least 2GB of RAM, MySQL should have plenty of room to play. Everything related to email will be FreeBSD. SMTP/POP/IMAP will be spread across a couple boxes. Webmail will be on a box by itself (because SquirrelMail/php/apache combo seems to have some resource issues). > > Of course make sure that your NFS link is good. If you export the > mailboxes to NFS, and the clients are not running FreeBSD, but e.g. > Solaris, you may have a low performance NFS link because of > differences > in window sizes. You may need to tweak a lot then. Even when running > FreeBSD, setting readahead and blocksizes in NFS mount options is > advised. But note: NFS performance means tweaking. > > How does the mail get delivered to the mailboxes? I strongly advise > against delivering over NFS. Put your sendmail/postfix/* logfiles on > another disk than the storage files. Just say NO to sendmail, Postfix is a possibility, but Qmail is my MTA of choice at the moment. I want the mailstore to be once-removed from the 'net. By creating a SAN, I feel that I'm reducing the potential security risks while decreasing the complexity of administering everything. > > How do clients access the mail storage? Do they use IMAP or POP3 to a > machine which mounts the files over NFS? Consider cutting out > the middle > man: run IMAP/POP3 service on the storage machine if > possible. Make sure > you know where the bottleneck is. Most use POP3, but there are a growing number of people who are using Webmail (IMAP). The only bottleneck I have right now, is the resource issues with Squirrelmail/PHP/Apache combo. I fully expect that a pair of XP1800 boxes to run all SMTP/POP3/IMAP/Anti-Virus functions will last me until I outgrow the ~100-150GB that my storage server will end up having (I've actually already decided to go with the SCSI solution). > > The MySQL server should not have a big impact. Unless the machine also > works as a relay (avoid the storage/relay combination when possible), > the databases are not big and not that often queried. If this > would pose > a problem, consider making regular dumps of the databases to standard > Berkeley/*DBM files and using those in your daemons. As for avoiding the "storage/relay" combination, this is already the plan. I'm building a STORAGE server, not an SMTP server. The queue will be handled on a WD JB disk (8MB cache), which is fast enough for my purposes (SMTP and Anti-Virus), and cheap enough that I can let it get torn up. When the time comes, my SMTP/AV processes will be running on Dual XEON boxes w/scanning taking place on an md(4) device for sure. Currently, authentication and verification takes place from from cdb files and/or flat-files, which works fine in terms of performance, but has certain limitations in terms of administrivia. For many reasons, I want to move all authentication to a central database. My billing software (Platypus) will add/remove/suspend/alter accounts via ODBC connection to the MySQL server. This will serve for the authentication of RADIUS/POP3/IMAP/SMTP, which will end up being 1000's of requests per minute, but shouldn't present a signifigant load on the box. If necessary, I can always move the MySQL database to a separate server. Radius Accounting, while important, is a non-critical process for me right now, and will be handled by another MySQL server (with a single ata100 drive). Before you ask, I have considered LDAP, but I'm not comfortable enough with it that I feel I can use it in a production environment. Quite frankly, the few times I've tried to set it up, I got nowhere, even though I've followed at least 2 different How-To docs to the letter multiple times. > > On ATA controllers: the 3ware vs. Mylex is almost a holy war. In > archives and on google you can find many experiences, but often > contradicting. I have no personal experience with these. CPU should > however not really be the problem. Actually, I've already decided to go with the SCSI solution. Both will cost me $3-4k, and while ATA disks are half the cost, they seem to need replacement 3x as often. In the long run, I feel that I'll be better off with a SCSI solution for this project. Also, it looks like the final configuration will be a 108GB of RAID0+1, though I may break down and get 73GB drives instead, giving me 221GB. -- Troy Settle Pulaski Networks http://www.psknet.com 540.994.4254 ~ 866.477.5638 Pulaski Chamber 2002 Small Business Of The Year To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-isp" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?000601c2e890$08c973f0$aa8ffea9>