-------- Original-Nachricht --------
Datum: Thu, 22 Oct 2009 12:57:56 +0100 Von: Ed W lists@wildgooses.com An: Dovecot Mailing List dovecot@dovecot.org Betreff: Re: [Dovecot] HA Dovecot Config?
Steve wrote:
Hallo Ed,
I have never used FileReplicationPro but looking at what it offers it reminds me of GlusterFS. I use GlusterFS for all www data of the domains I host. I don't use jet GlusterFS for IMAP/POP storage for all domains I host. Only a small subset of the domains I host have GlusterFS for IMAP/POP but so far it works without issues and I will soon or later migrate the other domains to be on GlusterFS as well.
The setup of GlusterFS is ultra easy (compared to other clustering FS solution I have seen) and it offers some very nice functions.
In the past I have burned my fingers with older GlusterFS releases when I have tried to use it as storage for IMAP/POP but the later 2.0.x releases of GlusterFS are more stable.
The thread moved on and no one seemed to bite, but I also have watched glusterfs for a long while now and been very attracted by the basic principle. I would be very interested to hear more about how it's worked out for you?
I use GlusterFS since long time. For mail hosting I waited for release 2 and then when it was out I switched all mail domains to use GlusterFS 2.0.1. It was a ultra big failure for me. The process took so much CPU that I could barely run anything else on the system and stability was very, very bad. That forced me to switch back to NFS for all the domains. Then later around 2.0.4 I looked again at it and things where more stable. I started then with a bunch of domains to run on top of GlusterFS 2.0.4 and moved then to GlusterFS GIT and somewhere around 2.0.7 the GIT version broke horribly in my setup that I switched to 2.0.7. That has been some weeks ago and since then I run around 1/3 of my domains on top of GlusterFS 2.0.7 with two active nodes using server side replicate, io-threads, write-behind and io-cache. On the client side I have no performance translators or anything such. Just bare client. I used server side replication because I wanted the shared storage to behave like a SAN and not use the server part as dumb bricks where the client is responsible for the replication. So far both nodes have 2 x 1TB disks in RAID 1 mode and those disks are then exported as a GlusterFS brick doing replicate. Setup in Dovecot is +/- like you would do if you would use NFS.
Got any benchmarks, perhaps comparing against local storage and vs NFS?
I have benchmarked but have nothing made to show to others. In general I can say that using GlusterFS on gigabit network is +/- 1/3 to 1/2 of the raw disk speed. If you add performance translators then the speed is somewhere between 1/2 to 1/1 of raw disk speed (depending what block size I use and depending if I use Booster or not). With the performance translators you can easy saturate a gigabit connection. Have not tried to use anything faster then gigabit. Off course local storage is faster since I don't have that theoretical 125MB/s limit I have when using gigabit. The newer releases of GlusterFS are comparable to NFS in terms of speed. In terms of CPU usage GlusterFS is way behind NFS and local storage. In terms of flexibility GlusterFS is better then anything else.
It seems intuitively very appealing to use something heavily distributed in the style of gluster/mogile for the file data, and something fast (like local storage) for the indexes. Perhaps even dovecot proxying could be used to help force users onto a persistent server to avoid re-creating index data...
Have never tried MoglieFS. What I don't like about it is that it uses HTTP for transport and operations.
For smaller setups it would be appealing to find something simpler than DRBD for a two server active/active setup (eg both servers acting as both storage & imap and failover for each other)
That is easy done with GlusterFS. Very easy.
Cheers
Ed W
// Steve
GRATIS für alle GMX-Mitglieder: Die maxdome Movie-FLAT! Jetzt freischalten unter http://portal.gmx.net/de/go/maxdome01