Tweaking Cyrus IMAP

Storage Tiering

Various opportunities exist to tier storage. With storage tiering, we mean to distribute payload over different levels of storage, where levels mean to refer to storage cost, performance and capacity, in a way that makes optimal use of the storage solution(s) -- with an eye on excellent performance yet remain cost-effective.

Imagine that you have the following levels of storage:

  • A low-cost solution with large capacity,

    such as many consumer-grade 4 TB HDDs,

  • A high-cost solution with little capacity,

    such as a few business-grade 512 GB SSDs.

The following options are available to split up the payload for a Cyrus IMAP server:

  1. Partitioning the mail spool;

    where you divide a single backend server's mail spool in to multiple partitions, presumably at least two, and make sure that mail folders end up on the correct partition.

    You would have one partition mounted off of the fast storage contain the day-to-day payload, such as user's INBOX folders, while the partition mounted off of the slower, cheap storage holds less frequently used data, such as archive folders.

  2. Distributing the mail spool;

    where you would have one server hold day-to-day payload, and other servers hold less frequently used data, such as archive folders.

    Note that distribution of the mail spool requires a Cyrus IMAP Murder topology, so that access to the mailboxes remains transparent.

    Also note that such archive folders may be hosted using not only a low-cost storage tier, but perhaps also using a less resourceful compute node.

  3. Caching I/O using dm-cache;

    This mechanism is a feature of Linux distributions that ship a kernel version of 3.9 or later, called dm-cache, and allows multiple disk volumes to be used as tiered caching storage.

    Using this technology, frequently accessed data can automatically be promoted to the fast storage, while less frequently accessed data can automatically be demoted to the cheaper storage.

  4. Using metadata partitions;

    While the access patterns for data and metadata are different, so could their storage locations be.

    In this scenario, you would distinguish between a mail folder's message content and metadata -- the metadata consists of mail headers, indexes created for the purpose of searching, and such, while mail messages are stored in separate files.

    When a mail folder is opened, and the messages in the folder are listed, no mail message content is touched -- the results come entirely from metadata.

    It is only when the individual mail message is being fetched, that the message file is opened.

Partitioning the Mail Spool

Distributing the Mail Spool

Using Metadata Partitions

Synchronous File Operations

A default Kolab Groupware installation comes with a set of settings suitable for the vast majority of our users -- mostly home users and other small(er) deployments.

It is typical for these installations to not have battery-backed I/O controllers, and/or some other form of enterprise-level storage.

To prevent data from being corrupted during a power outage, the default for our Cyrus IMAP packages is to ensure the mail spool (/var/spool/cyrus/) and configuration directory (/var/lib/imap/) and all files contained therein have the synchronous filesystem flag set.

To gain performance, execute the following:

  1. Remove the synchronous flag from the directories and files:

    chattr -RV -S /var/lib/imap/ /var/spool/cyrus/
  2. In /etc/sysconfig/cyrus-imapd (or /etc/default/cyrus-imapd), change the following:

    CHATTRSYNC=1
    

    to:

    CHATTRSYNC=0
    

Mailbox locking

Cyrus IMAP uses fcntl(2) based file locking for mailboxes, for example during SELECT commands. To mitigate race conditions it locks mailbox names even for non-existing mailboxes. For example, if user foo issued the following command

SELECT INBOX.x

for non-existing mailbox x, it creates a lock file

$CYRUS_CONFDIR/lock/user/foo/x.lock

which might be left on the filesystem after completion of the command.

This has in practice not shown to be an issue. If this is a concern however, keeping the lock file directory in a tmpfs allows for both fast locking and to purge stale locks during controlled Cyrus downtimes.