Thomas, thank you for taking the time to write such a detailed response.
> Firstly, it'd be interesting to know what you consider high volume > (in terms of concurrent users, total new emails per day and maybe > average number of emails per folder) It's a little unknown for now, which is why I left it open. For starters, at least hundreds of concurrent IMAP users (growing quickly), tens of thousands of new messages per day. Folder sizes will start small but we expect them to grow quite quickly without much expunging. > Also performance probably depends quite a lot on the usage patterns and > other requirements. If your users largely just pull there emails > using POP3, the extra indexing features of Dovecot probably don't help > you very much. But if your users do a lot of listing/searching in a > scenario where many people keep most of their email in IMAP folders some > people have claimed quite considerable performance improvements up to an > order of magnitude (a dedicated webmail software development and webmail > service provider company with a total user base of several million, > don't know how large the particular installation was, don't know what > storage infrastructure they were using). Most users on webmail, some using IMAP desktop clients. I've heard at least anecdotal evidence suggesting the same thing. I hate to be saying stuff like that on this list. :( > An interesting point is probably also if you plan to use share nothing > nodes with dedicated storage block devices per node and lots of RAM per > User (i.e. a situation where most of the indexes/directory tables are > located in the local main memory), or if you are going for a setup with > a NAS (e.g. NFS) to provide for high availability with failover time > measured in seconds, with very little caching at the local IMAP node. In > the latter case, the NAS part quickly becomes a bottleneck and thus > Dovecots ability to reduce the number of IO requests required is > probably a helpful thing to improve scalability. Haven't decided yet, but leaning toward some sort of NFS-ish solution. Would love to hear more advice on this. My understanding is if we used Dovecot's dbox where filenames don't change as they do in maildir++ then the biggest problem with IMAP over NFS which is lots of small writes becomes not such a big problem. Is this correct? Problem there is that we have lots of maildrop scripts that would be hard/impossible to migrate to sieve - we'd have to consider continuing to use maildrop but make it call Dovecot LDA for ultimate delivery or figure out how to convert maildrop scripts to shell scripts that sieve can pipe to. > In the latter case, spending money on a good NAS and client hardware/OS > that deal well with your NAS protocol is probably a good idea. > > I.e. go for guys like NetApp rather than ancient Linux kernels > put in a fancy x86 chassis and labelled NAS. Or best avoid this > problem altogether by scaling out with many NASes rather > than scaling up with a big one (especially if you expect lots > of growth in terms of users). Budget may dictate the latter. Does something like GFS make sense or add to the layers that sap latency and performance? > Linux' NFS client implementation used to suck quite badly > several years ago at least, FreeBSD was better and Solaris Sparc > (on Niagara) was better still. I ran into big problems trying to use Courier over NFS a few years so I agree! But it's improved recently enough to give it a try? Maybe if I can use Dovecot dbox it would help more too. > In the NAS scenario, a colleague recently ran a recent Courier Imap > on fairly loaded Solaris/Niagara 1 boxen serving thousands of > parallel sessions per box against a Dovecot 2.X running on a > CentOS6/x86-VM serving just the test mailbox containing about > 4000 emails. Accessing that test mail box with a webmail > software that needs to read in email listings every time > you log in, he was quite impressed with how much faster > it seemed, but he didn't do any accurate measurements yet. > > But if you use a shared storage, Dovecot's use of index files could > also become a drawback because it's much harder keep these in sync > and in working order if you allow concurrent access > to these structures on several nodes. Courier imap is pretty rock > solid with many nodes accessing the same NAS. Hmm interesting. > Using an MDA that can create the same index structures that > your IMAP server of choice can read in order to reduce the need > to scan/rebuild folder indices is probably a good idea > too. Thus the idea of calling Dovecot LDA from inside maildrop. Yuck, but it's CPU cycles at delivery time versus better IMAP performance, so maybe it's alright. > Also, think about backup and disaster recovery right from the start. > Traditional backup software is quite bad at dealing with many millions > of small files that change a lot every day. A storage > backend that allows you to do filesystem level snapshots and mirror > these to a place far away is better. Gotcha. Was considering XFS too, but I don't think it allows snapshots. With GFS...? You're right, this will require difficult thought > If you are aiming for 100% uptime, invest in a storage-technology that > can survive blowing up at least one component device while > continuing to serve data. Dealing with thousands of irate users > that cannot access their email, who are fearing for their data or whose > new emails get rejected because you have to do some unexpected offline > maintenance on your storage infrastructure during the day can be > quite challenging. And the damage to the reputation of your service > can linger for years. Agree 100% > BTW, if you read German, or at least can gather what you need from > some of the stats pictures, here's a 5 year old article > on the subject. > > http://www.linux-magazin.de/Heft-Abo/Ausgaben/2007/06/Auf-der-Teststrecke/ Don't read German but thanks. What I got out of that is maybe I should be considering Cyrus! :) > On Thu, 2012-06-21 at 10:26 -0700, email builder wrote: >> Hi, >> >> I'm a long-time Courier user and very happy with it (thank you, Sam). > I'm currently tasked with spec'ing out what will be a high volume email > environment and I'm a little concerned that Dovecot might be the better > choice (due to indexing). I have only started to do research, so forgive me > if > I jump to conclusions. >> >> I'd like to know if anyone has practical or anecdotal advice about how > Courier stands up to Dovecot under high use and what tweaks can be made to > Courier to keep up. >> >> TIA >> ------------------------------------------------------------------------------ Live Security Virtual Conference Exclusive live event will cover all the ways today's security and threat landscape has changed and how IT managers can respond. Discussions will include endpoint security, mobile security and the latest in malware threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ _______________________________________________ Courier-imap mailing list Courier-imap@lists.sourceforge.net Unsubscribe: https://lists.sourceforge.net/lists/listinfo/courier-imap
