On 01/14/12 1:56 AM, Rafał Radecki wrote:

We are thinking about replacing these storage hosts with one solution, maybe a storage array with appropriate disk space and I/O capacity. What are pros and cons of that solution? Do storage arrays have appropriate I/O capacity (X*software RAID)? Does it scale good? Does storage capacity also scale good in those solutions?

those questions are all extremely vague. 'storage arrays' vs 'storage hosts' could mean almost anything. without a thorough analysis of your workloads, its impossible to guess what your performance or capacity requirements are.

Rafa³ Radecki radecki.rafal@gmail.com wrote:

We are thinking about replacing these storage hosts with one solution, maybe a storage array with appropriate disk space and I/O capacity.

As John had mentioned the solution really depends on a complete analysis of your requirements (average and maximum message size, number of messages, number of submailboxes, number of users, number of concurrent users, interconnect, etc). I'm not sure if you're talking about replacing just the storage subsystem or the entire mail system, but I'll assume the latter.

Although I mentioned Cyrus Murder http://cyrusimap.web.cmu.edu/mediawiki/index.php/Cyrus_Murder on the other thread, that is intended for truly large installations and may (or not) be overkill in your situation.

One nice thing is that a basic Cyrus installation (without the Murder part of the architecture) installs onto a single host (and is available in the CentOS distribution), and can scale up when circumstances dictate. Cyrus is extremely easy on resources; probably a decade ago I had (going by memory) over 100,000 users on one machine.

One configuration I like for "single instance" (non-Murder) installations is to run Cyrus in a two-node pacemaker/corosync active-passive cluster using DRBD and local disk as the storage. In this configuration both nodes have mailscanner + sendmail running (equal MX priorities in DNS), and both are configured to hand off the messages via lmtp to the active node's cyrus processes. These are sealed servers (no general login access) with authentication tied into a master-replica LDAP configuration. And of course the local storage on the nodes is Linux software RAID (so it's replicated at not only the host level but also the DRBD level).

Even ignoring the usual HA aspects, it's nice to be able to do most maintenance without bringing down the mail system.

And of course if your site is large enough that the "single instance" configuration runs out of capacity, you can still expand it into the Cyrus Murder architecture quite easily.

With respect to NFS, trying to use it as the back end storage has the usual problem with NFS locking and perhaps performance. Using has in the past been contraindicated, but I've not looked into that issue recently to see if there are any changes. If you're leaning toward a storage appliance (netapp or whatever), I would expect cyrus to work just fine with iSCSI.

Cyrus also has nice searching and mail filtering capabilities.

Devin