[prev in list] [next in list] [prev in thread] [next in thread] 

List:       vchkpw
Subject:    Re: [vchkpw] Clustering
From:       Alex Borges <alex () co ! com ! mx>
Date:       2004-01-02 10:32:41
Message-ID: 1073039560.1611.19.camel () lextibook
[Download RAW message or body]


Well... i did some research for a proposed 120k maildirs on this setup,
although with squirrelmail.

Here are some big pointers:

A) The bottleneck of smtp is the storage (good storage bandwith), since
its such a simple protocol. Still, storage bandwith is the important
thing in the whole analysis cause if thats slow (you generate more IOPS
than what the storage can handle), everything will suck.

B) The bulk processing youll need, will be generated by email analysis

C) The second largest processing hog will be the webserver with the
webmail cgi's

The combo of this premises takes you down to this:

1) Take the average number of emails you think your users receive. 10-20
per day is a reasonable estimate for some uses (schools for example).
Then also an average email size. Multiply one by another, youll get your
kbps per mailbox.

2) Take into account if you are using pop or imap as the mailbox
protocol. Imap will charge more IOPS in a first analysis, but over time,
it actually takes less (cause users can delete email they dont want
before reading it). Also, work into the model the usage pattern of your
users... how about: 80% of the users check email from 9:00 to 9:30 AM.
Youll be able to calculate the IOPS generated from such an ugly,
although typical, peak. Also, mbox format will be easyer on the IOPS,
but a lot tougher to the imap server. So, choose mbox vs maildir
carefully.

3) As you want email analaysis (antispam, antiviri), youll want to
calculate how many emails can that really deliver (you can say you
receive 10000 emails per second, but the virus/spam farm wont be able to
actually deliver that many to the storage, unless you have an UberFast
bunch of servers doing the analysis).

Put all that into your opencalc, play with it, make a test farm, present
to boss.... ask for a raise....get a denial... go look for another
job...;-)


Now...techietalkin about the techs for this: Id go for ldap/sql based
user management (as your admin tools allow), consolidated, replicated
storage (SAN, DRBD or GFS), LVS for the load balancing, the smaller
protocol for internal email passing (i dont remember the name, its an
much smaller, simpler, smtp lookalike that suits delivery between
servers on a farm).
 
> I did some search on google and I have found clustering solution offered by
> HP "Proliant DL380 G3 Packaged Cluster" and a solution offered by Netapp
> using their Filers ( F710 for example) which I have found it here on this
> mailing list. I also checked LVS and I have found someone with the same
> setup, who mentioned that there are some modifications to be made on the
> sources.

I dont see why. 


["signature.asc" (application/pgp-signature)]

[prev in list] [next in list] [prev in thread] [next in thread]