[Dovecot] Architecture for large Dovecot cluster

[Sven Hartge]

Stan Hoeppner <stan at hardwarefreak.com> wrote:

> Sven, why didn't you chime in?  Your setup is similar scale and I
> think your insights would be valuable here.  Or maybe you could repost
> your last on this topic.  Or was that discussion off list?  I can't
> recall.

Rather busy right now with a large scale Identity Management+AD rollout here, so
unfortunately not too much time to elaborate my setup in great detail.

But after testing the nothing-shared-6-node-cluster setup with imapc as
the backend for shared folders I concluded that this does not scale very
well (the imapc-part, that is) and changed my plans to an director-based
NFS-backed (Netapp 3240) setup, which is much more common.

I reckoned I'd be nearly the only one on this planet to be so crazy to
try to use a backwards-normal-user-as-master-user-for-imapc setup for
shared folders and that having anyone other than me understanding that
setup, let alone getting support for it, would be to big a hassle.

So I put the mdbox storage on two 15k-SAS-NetApp with 1TB FlashCache,
connected with 2x 10GBit to the SAN, using NFS to mount the volumes in
my 6 backend-dovecot servers, putting 2 director-dovecots in front,
which will sit behind a Linux IPVS loadbalancer. All systems are VMs on
ESX.

I recently added two more shelves with SATA drives to the NetApp to use
as storage for the alt-storage feature of dovecot to automatically
migrate mails older than 180 days to less expensive storage.

As of now, the system is not yet live (see IDM rollout above), I hope to
resume my migration in late spring, early summer.

But during initial synthetic benchmarks have show that this setup will
be more than sufficient to provide the needed oompf for my 15k users,
with enough room to grow.

Interesting datapoint: NetApp Deduplication did only recover about 1% of
storage space with mdbox-based mail storage, while on an maildir-based
mail storage, the rate was about 15%. (This was tested with a copy of
real user data, so is accurate for my workload.)

Grüße,
Sven.

-- 
Sigmentation fault. Core dumped.