On Wed, Sep 24, 2003 at 10:48:39AM -0400, Scott Adkins wrote: > What version of Cyrus are you using?
CVS HEAD from yesterday, which is called 2.1.15. > We are using 2.2b1 here, and are > experiencing something similar with memory issues for IMAP. I brought up > the discussion a few weeks ago about it and it was characterized as some > kind of mmap() weirdness on our Tru64 platform, which I don't really think > is the case. > > I will look in our logs to see if we are seeing a similar issue with a > lot of accepted connections as you are. I actually only had 1 connection - I should have made it clear that those were *all* the accepts in imapd.log (from my ... it looks as though the line is repeated - sorry) - it was just to show that nothing much was going on apart from that 1 imapd running for about 35mins until that one imapd exhausted virtual memory. > My experiences are as follows with the problem: > > 1) It seems that the RSS footprint of the process increases when a user > makes a connection (from observation). Yes - and for me, essentially the more messages I append, the greater the increase. It doesn't matter which mailbox, nor whether the selected mailbox changes. It seems to be a per message increase. > 2) Most of the time, we may see a process jump up to 26 or 27MB in size > (and lately, we are now seeing them jump up to 30-32MB in size), but > typically go back down to several hundred KB shortly thereafter. If > the system starts to slow down for whatever reason, the time for a > process to shrink increases, and we have more processes at the bigger > RSS size and run the risk of exhausting memory and swap. > > 3) The problem was characterized as an mmap() problem on Tru64 because > our mailboxes.db file is about 27MB in size. However, we are seeing > the sizes jump to 30-32MB in size, and our mailboxes.db file is still > only about 27MB in size. *shrugs* Hmm, I'm using NetBSD, and my mailboxes.db is tiny (20k). Remember, I was just testing with 1 user. There are 2 users total in the system. > 4) Restarting the Cyrus server has a dramatic effect on memory usage by > putting all the processes back at the base several hundred KB size. > It takes quite a while before we see the issue #2 become a problem > again (even on a busy server). > > Anyways, I have been looking at this problem over that last several days > and my gut feeling has been leaning towards the on-connection route... > which certainly jives with your message below... For me it seems there is a memory leak in append :/ Maybe for you each imap connection receives another few messages => they grow? So memory isn't even freed between connections? Cheers, Patrick
