Re: Robustness
We just went through a whole bunch of insanity trying to improve the
performance of our mail server... If you need 72GB of mail storage for pop
boxes and spool, a single set of spindles for RAID-1, would not be the best
idea. The bottleneck in our former server was the relentless mail-checking
habits of our customers being processed by a single spindle (even though it
was a Barracuda). It just simply could not handle all the random transactions
fast enough.
We ended up going with a Dual P3/500 Debian box with a Mylex eXtremeRAID 1100
controller with 5 9GB Barracudas in a RAID-5 configuration, with one more as
a hot spare. The spool and mail storage are on the same array, partitioned in
half. Even though RAID-5 is slower on writes, the massive speed of that
controller has made it hardly noticeable, though if write performance begins
to take a hit, it'll be changed to RAID-0+1.
If you need to support many users, you need some kind of striping across
multiple spindles... the mechanical nature of drives makes it mandatory.
On another note, our continuing problems with NFS on Linux (to our NetApp or
between machines) has been less than impressive. Processes are constantly
blocked by waiting NFS transactions. Does anyone have suggestions for
improving NFS performance? We've tried NFSv3... no help.
-Kevin Blackham
Xmission Internet
Salt Lake City, UT
On Fri, Nov 26, 1999 at 12:36:22PM +0100, Russell Coker wrote:
> On Fri, 26 Nov 1999, Chris Wagner wrote:
> >>I want to replicate a pop mailboxes but I don't know how.
> >
> >Set up POP service on both machines. Map both their IP addresses to the POP
> >server name. Then use rsync to make sure the mail spools match. This is a
> >simple solution and there are problems.
> >
> >1) Both machines have to have the same user accounts unless you go with qpop
> >or something
> >2) If one dies, the R-R DNS will give out the dead address 50% of the time
> >
> >A better solution is to setup two POP machines which mount the mail spool
> >off a third NFS machine. This NFS server should be RAID-5. Between your
> >users and these POP machines put a firewall/proxy. It is this proxy that
> >will split up the traffic between them and know if one dies.
>
> RAID-5 is good if you're on a budget. However performance for writes is poor
> (try running Bonnie or Bonnie++ on a RAID-5 filesystem and compare to RAID-10,
> or RAID-1). If you're in the situation where you need more than one server
> in case of a crash then you probably don't need to save money on RAID-5.
> Also as 72G SCSI drives are available you can have 72G of RAID-1 storage at
> minimal effort.
> For a mail spool writes will probably equal reads so write performance counts.
>
> >But that may be overkill. Really, your best bet is to just use rsync to
> >keep a backup copy of the mail spool on another harddrive and if the machine
> >dies, stick the backup mail spool in a "hot spare" POP server. Having the
> >kind of fault tolerant POP service you're thinking of is a non-trivial exercise.
>
> If you have a small amount of data then you probably don't need this. If you
> have a lot then rsync really won't cope. If you have 20G of data then rsync
> will probably take 20 minutes to run.
> The cheapest solution for redundancy of file servers is dual-attached SCSI.
> Of course making sure that machine A is dead before machine B mounts the
> drive is non-trivial...
>
> --
> Electronic information tampers with your soul.
>
>
> --
> To UNSUBSCRIBE, email to debian-isp-request@lists.debian.org
> with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org
>
Reply to: