Chuck Yerkes

Sumit Malhotra wrote:
> i am running sendmail on solaris 7.0. the sendmail being used as

SunOS 5.7 = Solaris 2.7 or marketng speak: "solaris 7" (neither here
nor there).

> internet gateway in our organisation handles ll incoming/outgoing
> mails of the organization .
> With around 70 mails being receivd per sec and sendmail process
> generally in range of 100-500 the handling is getting a bit
> tough(Connection getting timed out, mailq reaching 1000 etc.)
>
> will tweaking in the sendmail timouts help in this case ?

Not usually.

you want disk IO. Multiple spindles, multiple spool directories.

There's a book written by a former compatriot I refer to as The Stephen
King of sendmail authors...
Sendmail Performance Tuning, by Nick Christenson.
It would seem appropriate.

(I suppose I'm alone in calling him that, but it starts somewhere ;).


Also Solaris 9 has better file system than Solaris 7.
Veritas is MUCH better.

You can also run FreeBSD 5.2 (with FFS2) on SPARCS.
ftp.freebsd.org:/pub/FreeBSD/releases/sparc64/ISO-IMAGES/5.2/

Nicely fast file system.
Sun's file systems just aren't really impressive. fortunately, they
ship slow disks, so you don't notice too quickly.

15kRPM disks and groups of them.


Seriously, ponder many disk spindles and an OS upgrade or Veritas FS
if you can.

I came to love Baydel RAID arrays - the fastest RAID I've found for
a reasonable cost. it regularly smokes RAID boxes costing 3x more.
What's nice is that it has a big (mirrored, battery backed) RAM write
cache and a really fast RAID algorithm.

With mail, you write the queue file and delete it 5 seconds later. With
the Baydels, this usually means that the files never hit actual disk.

I brought a loaded E250 that the customer insisted on using (450Mhz x 2
- wow!) that handled 40k messages/hr and was often at a load average of
15+ down to 3 with a new SCSI card (it was put: "Oh, not Sun SCSI, fast
SCSI") and an older 50GB Baydel (we needed the RAID controller and
cache, not lots of disk).



Claus, what's the 8.12 feature that lets me not fsync if the machine
behind (next hop) gets the message and accepts it? I learnt it well
over 20 months ago and forget...

Chuck Yerkes

Sumit Malhotra wrote:
> Thanks for the quick responses.
> I have already started tweaking the sednmail configuration with
> timeouts.
>
> In addition to the above , as all mentioned, the problem could be in
> disk I/O or with options while mounting the filesystem where queue
> resides.
>
> But i have SunOS 2.6(Apoligies for writing solaris 7 previously)
> which doesn't have the logging and noatime options available.
>
> Is there ay other way of achieving the same ?

Hmmm, thought noatime was really old (sunos 4 or 3 even).

Be careful on timeouts, you may not want to lower some of them much.

That said, for volume SENDING, I've tighten the timeouts on the main
machine - and using a fallbackMX to get it to another machine if I
can't send it out quickly from the main machine.

(layers of FallBacks can be used in a bucket brigade. I was
sorry this never made it into queuegroups (if > 1 day, move it to
"slowmail" queuegroup).


Multiple queue directories will help immediately.
1) If you're not running 8.10+, then do (8.11.$last or 8.12).
2) if you can get more SPINDLES involved (2GB disks are fine),
then do so.

making your queue dirs (or queuegroups which have MANY advantages)
/var/spool/mqueue/q*

and having /var/spool/mqueue/q01 /var/spool/mqueue/q02
/var/spool/mqueueq03, etc. you keep number of files/queue lower.

I'd written this up on sendmail.net but the site seems to be gone
(replace with sendmail.com).


VERY important with UFS/FFS file systems (sunos - all).

okay, now you have more disks. Mount it under /mq/disk1 (for example).

/var/spool/mqueue/q03 is a sym link to /mq/disk1/q03/.


We can play games that underneath where ..../q03/xf .../q03/qf
..../q03/df are directories.

Fancy, you make them each links to a separate spindle.
That way when writes are happening, they use multiple disks.

Nick, afair, has a good writeup on that.

I make xf/ subdirectories actually point to a memory file system.
Locks don't survive reboots.
/var/spool/mqueue/MFS/q01-xf
/var/spool/mqueue/MFS/q02-xf

etc. MFS is fast :)


You'll benefit from a platform upgrade. New OS (solaris 9 or perhaps
BSD) or just an Intel box.

Often people run mirroring on Suns for robustness. On Sun disks, which
are usually slow (eg. not 10,000 RPM or 15,000 RPM (preferred)).

Mail is *ALL* about I/O. Give me a 500MHz CPU and really fast disk
and I'll do better than a 4-way box with crappy disk.

Solaris 2.6. 1997? '96? Solid. Liked it more than 2.7. But 2.7
is pretty old too ('98 or '99).


queue groups can be REALLY REALLY helpful in separating mail.

I shove mail FROM daemon/root/other 'bounce' type things into
a queuegroup with 20 directories.

I compiled it with FFR_QUEUERETURN_DSN so I can bounce DSN's in less
than 5 days.

Inbound mail hits one queuegroup, outbound another.

My main outbound machine, when I arrived, was usually 20,000 messages
in 1 queue (40,000 files in a dir on Sun takes 5-6 minutes just to run
'ls'.

It's now got ~5000-10000 queued messages 30 directories. It's a bit
faster :) (most of those are spam bounces).


That's my second long answer. two is my limit for free advice.

Sumit Malhotra

Hi Chuck,

Thanks for response.

I already have queue groups. but i am not using it for the
prioritizing based on the fallback MX or slow response. It's sort of
random. Can we achieve the same with queugroups (in any way ???)

I will not prefer the OS upgrade or change of harddisk to the
spindles.

regards,
Sumit

Chuck Yerkes <Newsboy@Jan2004.NOSPAMsnew.com> wrote in message news:<mSqNb.69862$8H.111225@attbi_s03>...
> Sumit Malhotra wrote:
> > Thanks for the quick responses.
> > I have already started tweaking the sednmail configuration with
> > timeouts.
> >
> > In addition to the above , as all mentioned, the problem could be in
> > disk I/O or with options while mounting the filesystem where queue
> > resides.
> >
> > But i have SunOS 2.6(Apoligies for writing solaris 7 previously)
> > which doesn't have the logging and noatime options available.
> >
> > Is there ay other way of achieving the same ?
>
> Hmmm, thought noatime was really old (sunos 4 or 3 even).
>
> Be careful on timeouts, you may not want to lower some of them much.
>
> That said, for volume SENDING, I've tighten the timeouts on the main
> machine - and using a fallbackMX to get it to another machine if I
> can't send it out quickly from the main machine.
>
> (layers of FallBacks can be used in a bucket brigade. I was
> sorry this never made it into queuegroups (if > 1 day, move it to
> "slowmail" queuegroup).
>
>
> Multiple queue directories will help immediately.
> 1) If you're not running 8.10+, then do (8.11.$last or 8.12).
> 2) if you can get more SPINDLES involved (2GB disks are fine),
> then do so.
>
> making your queue dirs (or queuegroups which have MANY advantages)
> /var/spool/mqueue/q*
>
> and having /var/spool/mqueue/q01 /var/spool/mqueue/q02
> /var/spool/mqueueq03, etc. you keep number of files/queue lower.
>
> I'd written this up on sendmail.net but the site seems to be gone
> (replace with sendmail.com).
>
>
> VERY important with UFS/FFS file systems (sunos - all).
>
> okay, now you have more disks. Mount it under /mq/disk1 (for example).
>
> /var/spool/mqueue/q03 is a sym link to /mq/disk1/q03/.
>
>
> We can play games that underneath where ..../q03/xf .../q03/qf
> .../q03/df are directories.
>
> Fancy, you make them each links to a separate spindle.
> That way when writes are happening, they use multiple disks.
>
> Nick, afair, has a good writeup on that.
>
> I make xf/ subdirectories actually point to a memory file system.
> Locks don't survive reboots.
> /var/spool/mqueue/MFS/q01-xf
> /var/spool/mqueue/MFS/q02-xf
>
> etc. MFS is fast :)
>
>
> You'll benefit from a platform upgrade. New OS (solaris 9 or perhaps
> BSD) or just an Intel box.
>
> Often people run mirroring on Suns for robustness. On Sun disks, which
> are usually slow (eg. not 10,000 RPM or 15,000 RPM (preferred)).
>
> Mail is *ALL* about I/O. Give me a 500MHz CPU and really fast disk
> and I'll do better than a 4-way box with crappy disk.
>
> Solaris 2.6. 1997? '96? Solid. Liked it more than 2.7. But 2.7
> is pretty old too ('98 or '99).
>
>
> queue groups can be REALLY REALLY helpful in separating mail.
>
> I shove mail FROM daemon/root/other 'bounce' type things into
> a queuegroup with 20 directories.
>
> I compiled it with FFR_QUEUERETURN_DSN so I can bounce DSN's in less
> than 5 days.
>
> Inbound mail hits one queuegroup, outbound another.
>
> My main outbound machine, when I arrived, was usually 20,000 messages
> in 1 queue (40,000 files in a dir on Sun takes 5-6 minutes just to run
> 'ls'.
>
> It's now got ~5000-10000 queued messages 30 directories. It's a bit
> faster :) (most of those are spam bounces).
>
>
> That's my second long answer. two is my limit for free advice.