Patrick Ben Koetter put forth on 8/10/2010 6:37 AM:
> * Bjorn Mork <[email protected]>:
>> i have tried to answer your queris, (Please correct, if I am wrong in
>> understanding your question...)
>>
>> We do have multiple IBM Blade server with 2.4 Xeon + 16GB + NAS over iSCSI
>> protocol......
>>
>> How many blades will be involved for such load....???

16GB should be able to handle an insane number of simultaneous SMTPd processes
especially if you use Postscreen, which I'd highly recommend with the user
count you're talking about.  This is highly dependent on your Postfix
restrictions, anti spam daemons and virus scanners.  Postfix, and SMTP in
general is not CPU bound, so with dual 2.4 GHz CPUs, even if they are the
"NetBurst" architecture, you'll run out of other resources before CPU
bandwidth.  The main one being...

Storage, the most important aspect of your inquiry.  You mentioned iSCSI and
NFS, which at face value doesn't make sense.  You don't connect an NFS client
to an NFS server over iSCSI.  iSCSI is for block level disk access and
functions almost identically to a fiber channel SAN.  iSCSI works at the block
device level underneath the filesystem.  NFS works _at_ the filesystem level.
 I'll assume you're actually using NFS, not iSCSI.

Given your load, you'll need at least 2 GbE links on the Postfix servers--1
for inbound SMTP connections, and other for access to the NFS server, and
these should be on separate ethernet and IP networks (different switches and
subnets from the incoming smtp traffic) for security and other reasons.

SMTP mail is primarily a disk bound application.  You need a very high seek
rate--high IOPs--as the mail flow is going to be lots of small files,
including queue files and files permanently written to maildir on the NFS 
server.

Would you please post the NFS/NAS server manufacturer, NFS/NAS server model
number, number of disks and their RAID (stripe) level, and what types of
disks--i.e. fiber channel, SCSI, SAS, SATA, etc?

The random IOPs performance of your NFS/NAS server will be the overall
performance limiting factor, if it's not sufficiently beefy.

-- 
Stan

Reply via email to