[Dovecot] reason to switch to 1.1
We've been running 1.1 on about half of our servers for about a week now. Ive mailed before that I was pleasantly surprised by its better use of resources. Here's a graph showing that fact. Server load in the last 10 days.
http://uwimages.smugmug.com/photos/286355874_9FNp2-L.png
Cor
Cor Bosman wrote:
We've been running 1.1 on about half of our servers for about a week now. Ive mailed before that I was pleasantly surprised by its better use of resources. Here's a graph showing that fact. Server load in the last 10 days.
It may be good to list your hardware, user count, mailbox backend, and file system information, as I suppose that this kind of improvement is not universal.
Anders.
We've been running 1.1 on about half of our servers for about a week now. Ive mailed before that I was pleasantly surprised by its better use of resources. Here's a graph showing that fact. Server load in the last 10 days.
It may be good to list your hardware, user count, mailbox backend, and file system information, as I suppose that this kind of improvement is not universal.
This specific server is a dual core 2.8ghz xeon with hyperthreading running on FreeBSD 6.2-STABLE. We have over 1 million mailboxes, with about 75,000 daily active users. At peak maybe 20,000 concurrent, in a mix of webmail and direct imap. (no POP, thats handled by different software).
The backend is a NetAPP 6070 with about 170 harddisks. All mail and control files are on this netapp, but all indexes are on local disks, and we try and make sure customers end up on the same imap server every time. As soon as all servers are 1.1 I'll start experimenting with NFS indexes.
Cor
On Mon, Apr 28, 2008 at 12:42 PM, Cor Bosman cor@xs4all.nl wrote:
It may be good to list your hardware, user count, mailbox backend, and file system information, as I suppose that this kind of improvement is not universal.
This specific server is a dual core 2.8ghz xeon with hyperthreading running on FreeBSD 6.2-STABLE. We have over 1 million mailboxes, with about 75,000 daily active users. At peak maybe 20,000 concurrent, in a mix of webmail and direct imap. (no POP, thats handled by different software).
The backend is a NetAPP 6070 with about 170 harddisks. All mail and control files are on this netapp, but all indexes are on local disks, and we try and make sure customers end up on the same imap server every time. As soon as all servers are 1.1 I'll start experimenting with NFS indexes.
Just in case I understand you wrong: You're serving 20k concurrent users with 1 (one) server?
What mailbox backend are you using? maildir?
Chris
This specific server is a dual core 2.8ghz xeon with hyperthreading running on FreeBSD 6.2-STABLE. We have over 1 million mailboxes, with about 75,000 daily active users. At peak maybe 20,000 concurrent, in a mix of webmail and direct imap. (no POP, thats handled by different software).
The backend is a NetAPP 6070 with about 170 harddisks. All mail and control files are on this netapp, but all indexes are on local disks, and we try and make sure customers end up on the same imap server every time. As soon as all servers are 1.1 I'll start experimenting with NFS indexes.
Just in case I understand you wrong: You're serving 20k concurrent users with 1 (one) server?
What mailbox backend are you using? maildir?
Actually..it's 2 NetApp 6070s. But those are not just simple servers. They are very expensive, dedicated NFS boxes each taking up a full rack doing multiple terrabytes each, connected with multiple gbit links.
It's 99.99% maildir, with a very small hidden mbox legacy namespace.
Cor
On Tue, Apr 29, 2008 at 09:42:10AM +0200, Cor Bosman wrote:
Actually..it's 2 NetApp 6070s. But those are not just simple servers. They are very expensive, dedicated NFS boxes each taking up a full rack doing multiple terrabytes each, connected with multiple gbit links.
It's 99.99% maildir, with a very small hidden mbox legacy namespace.
How large are the (individual) mailboxes you're hosting there?
GeertHow large are the (individual) mailboxes you're hosting there?
Most of them are max 500MB, but average use is much less. It's a little difficult to calculate because almost all POP users empty their mailbox. We did some reports a few months ago where it showed that average mailbox size for POP users is less than 5MB, while average mailbox usage for IMAP was over 50MB.
Cor
On Tue, Apr 29, 2008 at 11:49:43AM +0200, Cor Bosman wrote:
How large are the (individual) mailboxes you're hosting there?
Most of them are max 500MB, but average use is much less. It's a little difficult to calculate because almost all POP users empty their mailbox. We did some reports a few months ago where it showed that average mailbox size for POP users is less than 5MB, while average mailbox usage for IMAP was over 50MB.
Ok, so you're probably not facing any filesystem bottlenecks w.r.t. number of files in a Maildir folder?
GeertJust in case I understand you wrong: You're serving 20k concurrent users with 1 (one) server?
Wait, I think I misunderstood you. We do not have just 1 imap server. We have 30 imap servers (a little overdimensioned at this time). I was just showing the graph of one of them. The others look similar.
Cor
On Tue, Apr 29, 2008 at 10:31 AM, Cor Bosman cor@xs4all.nl wrote:
Just in case I understand you wrong: You're serving 20k concurrent users with 1 (one) server?
Wait, I think I misunderstood you. We do not have just 1 imap server. We have 30 imap servers (a little overdimensioned at this time). I was just showing the graph of one of them. The others look similar.
Thanks for the info.
I conclude that it is no problem to serve about 1000 concurrent users per server (with dovecot 1.1 probably even lots more!).
Unfortunately I cannot compare this with our setup as we do not have dedicated NFS-boxes (Netapp). Our servers have to do local hard disk IO as well. I think this probably decreases performance (?).
Chris
Very persuasive. So what is the conversion process like tp go to 1.1?
Cor Bosman wrote:
We've been running 1.1 on about half of our servers for about a week now. Ive mailed before that I was pleasantly surprised by its better use of resources. Here's a graph showing that fact. Server load in the last 10 days.
http://uwimages.smugmug.com/photos/286355874_9FNp2-L.png
Cor
participants (5)
-
Anders
-
Chris Laif
-
Cor Bosman
-
Geert Hendrickx
-
Marc Perkel