Others have referenced some of the Veritas tuning docs. You'll find a
lot of information and disinformation in old postings on this list.
My recollections are that, as you might expect, you want as much
buffering as your resources allow. This starts with the largest tape
buffers (SIZE_DATA_BUFFERS) your system can stand. Important: Any one
tape should be written with just one tape buffer size. Also, test to
make sure your system can successfully use the buffer size you specify
... sometimes hardware can limit your choices.
Next, you want to provide as many buffers (NUMBER_DATA_BUFFERS) to each
writing process as possible. Important: the tuning docs give a formula
to help you determine this number. As I recall, the buffers sit in
shared memory and your system limits shared memory to 1/4 of total
memory. Also, there are probably other uses of shared memory, so don't
tie all shared memory up with NetBackup buffers.
Finally, I'll offer that Veritas probably doesn't change the defaults,
because it is scared to. They don't want an upgrade to break a site
because they've chosen a different default SIZE_DATA_BUFFERS, for
example. IBM's ITSM (aka ADSM) has done the same thing. IBM addressed
the situation by adding a self-tune feature (by default off). The
self-tuning works well for almost all shops, so that solved the problem
of a badly tuned product out of the box. Perhaps veritas will do
something in the area in the future?
NetBackup probably has a "pretty big difference" from ufsdump in most
cases, but you should be able to increase perofrmance significantly with
some tuning. Hope this helps!
cheers, wayne
Mark Sandrock wrote:
> I am struggling to grasp the use of NetBackup DC 4.5 on Solaris 8.
> The only server is a V480 with a stand-alone SDLT 320 and an L8 tape
> library having an LTO-2 drive.
>
> I went through the wizard and configured things for a simple test,
> but it's really hard to tell what's going on and when.
>
> I eventually got a manual backup to run, and was surprised to see
> the i/o to tape running about 5000KB/sec. Our current backup method,
> ufsdump, gets about 15000KB/sec. Pretty big difference.
>
> So a Google search finds a posting from 2001 talking about an
> undocumented method of improving NBU i/o performance by creating
> files called "NUMBER_DATA_BUFFERS", "SIZE_DATA_BUFFERS", and
> "NET_BUFFER_SZ".
>
> Are these files the way to go? And if so, why is this information
> not explained and configured right up front?
>
> Thanks for any help.
> Mark
|