Veritas-bu

[Veritas-bu] Preemptive scheduling/ Performance tuning

2000-10-17 19:25:42
Subject: [Veritas-bu] Preemptive scheduling/ Performance tuning
From: Jonathan Geibel Jonathan.Geibel AT disney DOT com
Date: Tue, 17 Oct 2000 16:25:42 -0700
you can adjust these netbackup parameters to your liking:

/usr/openv/netbackup/NET_BUFFER_SZ
/usr/openv/netbackup/db/config/SIZE_DATA_BUFFERS
/usr/openv/netbackup/db/config/NUMBER_DATA_BUFFERS

these are plain text files including only a single number which will
change the default value of that parameter

NET_BUFFER_SZ (set on both the client and the server) is the amount of
memory you will designate for, well, buffering the data going over to and
from the client across the network.  the client sets the receive buffer
size and the server's value sets the send buffer size (in a backup).
Making these equal is a good idea.  default value is 32k, we have is set
to 512k.  The number to place in the file is in bytes, so ours is set to
524288.

SIZE_DATA_BUFFERS sets how large the buffer is when streaming data to and
from of the tape.  default is 32k, we have it at the default at the
moment. This is only set on the media servers.  Be careful changing this
value, though! it can give some adverse reactions..  ie, if you don't have
variable-length writing set on you will encounter 174 errors whenever you
reach the end of the tape..  also, changing this value could screw up
restoring off of tapes that were written to with a different block size.

NUMBER_DATA_BUFFERS sets the number of data buffers (which are all set by
the SIZE_DATA_BUFFERS) that you will have during backups and restores.

watch out for setting these too high if you are low on memory..  every
process will suck up SIZE_DATA_BUFFERS*NUMBER_DATA_BUFFERS of memory even
if it does not use it.  If you run out of memory, you will start to get
error 11s on backups and restores..

there are a bunch of TCP parameters to tweek as well that can speed up
your backups..

these are set using ndd on a Sun.  

to see the list of parameters that are available to you to tweak, type:

ndd /dev/tcp \?

some of the obvious ones to change are tcp_xmit_hiwat and tcp_recv_hiwat

these change the window sizes on your tcp connections -- ie, the amount of
data the server will send to the client (or vise versa) without having to
receive an ACK back.  

this is set to 8k by default on the Suns..  IMHO, this is rather low..

on some of our servers we have set this as high as 512k

note that ndd changes these values dynamically, so you don't have to
reboot your server for these to take effect.

also, BE CAREFUL changing any tcp parameters as they will affect any other
application running on that server as well.

also, DO NOT make your TCP window sizes larger than your netbackup
NET_BUFFER_SZ..  this will cause some highly adverse behavior by your
server causing really slow performance.  basically, your TCP stack will
try to shove more data than your netbackup buffer has allocated and will
cause long, extended timeouts which are highly undesirable.  

anyone else have any performance tuning tips and tricks?

On Tue, 17 Oct 2000, Chandra Kalle wrote:

> 
> Is there any configuration parameters involved here?
> 
> We have an E4500 that sends data to an ATL P1000 (4/30 DLT7000)
> via Differential SCSI 12 meter fast wide cable 
> and the thruput we're getting is 2.61 - 3.71 MB/Sec.
> 
> The DLT7000 drive thruput according to vendor specs is
> 18 GB/hour (thats 5 mb/sec).
> 
> ?
> 
> 
> Chandra
> 
> On Tue, 17 Oct 2000, Rob Worman wrote:
> 
> > hmmmm
> > 280 GB / 25 hrs = 11.2 GB/hr = 3.1 MB/sec
> > 
> > your DLT 7000 should probably be going a bit faster than that, my friend. 
> > :-\
> > 
> > Under normal circumstances I'm feeling good if I average 25GB (or more) per 
> > hour per drive.
> > 
> > rob
> > 
> > At 3:32 PM -0700 10/17/00, Richard C Bond wrote:
> > >?? what/how many tape drives ? I am getting 280 gig in 25 hours, 
> > >on a DLT 7000. 
> > >
> > >-- 
> > >-
> > >Richard Bond (rbond AT mbt.washington DOT edu  (206) 605-3561
> > >System Administrator                          K-351, Health Sciences Center
> > >Department of Molecular Biotechnology         Box 357730
> > >University of Washington                      Seattle, WA 98195
> > >
> > >
> > >_______________________________________________
> > >Veritas-bu maillist  -  Veritas-bu AT mailman.eng.auburn DOT edu
> > >http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
> > 
> > _______________________________________________
> > Veritas-bu maillist  -  Veritas-bu AT mailman.eng.auburn DOT edu
> > http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
> > 
> 
> _______________________________________________
> Veritas-bu maillist  -  Veritas-bu AT mailman.eng.auburn DOT edu
> http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
> 




<Prev in Thread] Current Thread [Next in Thread>
  • [Veritas-bu] Preemptive scheduling/ Performance tuning, Jonathan Geibel Jonathan . Geibel <=