System Bottlenecks

[Mike Kercher]

Julian would have to confirm, but I believe that sendmail keeps a copy of the
email until it has been successfully delivered (assuming you don't have
supersafe disabled).  I refer quite often to my Sendmail Performance Tuning book
and still learn new stuff all the time.  Although the newer IDE drives are close
to SCSI in speed, the performance of SCSI drives comes from the on-board CPU of
the SCSI controller.  On IDE systems, the disk subsystem has to use the system
CPU for processing.  All of my mail servers are PIII-800 and below with at least
a gig of RAM.  I host email for several law firms and have never lost an email
yet and I do use a tmpfs for the workdir.

Mike


> -----Original Message-----
> From: MailScanner mailing list 
> [mailto:MAILSCANNER at JISCMAIL.AC.UK] On Behalf Of Errol Neal
> Sent: Tuesday, September 02, 2003 11:11 AM
> To: MAILSCANNER at JISCMAIL.AC.UK
> Subject: Re: System Bottlenecks
> 
> 
> At 04:47 PM 9/2/2003 +0100, you wrote:
> >The best I/O improvement is by making sure you've got plenty 
> of RAM and 
> >putting the MailScanner work directory in tmpfs (not either 
> of the mail 
> >queues though!)
> 
> That is a bit scary for us. Unpacking messages in a memory 
> based file  system could be catastrophic. *Shudders*. Too 
> scary to even think about it if for example, MailScanner dies 
> and leaves a bunch of mail in the tmpfs and we unknowingly 
> reboot the system... for us.. instant law suit. Can anyone 
> explain how this works? Does MailScanner unpack messages 1 at 
> a time, does it unpack all the messages bulky in this directory?
> 
> >Personally speaking ufs sucks and anything FS intensive struggles on 
> >Solaris (in fairness my experience is with low end machines, 
> E250 and 
> >lower).  You'll get more bang-per-buck using linux on Intel.  Where 
> >Solaris excels is at the high end and I can't see why anyone 
> would need 
> >a high end server for a mail load of only 15-20k.
> 
> We are using the lower end Netra T-1 and V Fire 100 (I 
> think). Turning on logging increases performance 
> dramatically. Compared against linux using XFS logging on 
> ultra 160 drives, the performance is almost equal.
> 
> >If you've got the money for Sun hardware buy Intel and get 
> an extra box 
> >for redundancy/ load balencing!
> 
> Lower end sun models are actually quite inexpensive these 
> days. 550MHZ cpu, 512 RAM, two nics and 40GB ide for less and 
> 1K US is not too bad.
> 
> We actually have 3 systems deployed at the moment, each 
> system handles about 15-20K messages a day, and that varies. 
> I guess what I am trying to achieve as I said earlier is a 
> strategic investment of dollars into what will make the 
> difference most dramatically. For example, if 1 gig of ram 
> will improve the systems performance over our current 512MB 
> Ram in a much greater way than deploying SCSI based /var/ 
> slices, I will put my money in the RAM and stick to my IDE 
> disks. This is what I need to know.
> 
> 
> Errol Neal, Systems/Network Administrator
> eneal at enhtech.com
> Enhanced Technologies Inc.
> http://www.enhtech.com
> 703-924-0301 or 800-368-3249
> 703-924-0302 Fax
>