Ben Rockwood wrote:
> I'm trying to tune the buffer settings for my tape system and am very
> curious what other people might have found as optimal.
> I'm using an ADIC Scalar 480 AIT, with 4 drives on a Sun 420R using gigether.
>
> Right now NET_BUFFER_SZ, SIZE_DATA_BUFFERS and NUMBER_DATA_BUFFERS are all
> default. I definately need to increase the number of buffers, but I'm not
> sure what the optimal buffer size might be, ...
You should look at the Veritas support site articles on this subject.
Don't worry if you're on 4.5 and the article was written for 3.x.
One of these is http://seer.support.veritas.com/docs/183702.htm
I'm relatively new at this, but there seem to be at least 2 camps. One
would have you raise buffers until messages you see in logs tend to
balance out waiting at tape and waiting for communication ... with
enough buffers to keep waiting down. There's an excellent Veritas
article describing this method.
I'm in another camp. First, maximize NET_BUFFER_SZ for your network
(maybe 32k-64K). Next, maximize SIZE_DATA_BUFFERS for your tape
drives/adaptors (maybe 32K-256K). Note that any one tape must be written
entirely with one SIZE_DATA_BUFFERS. Finally, calculate the maximum
NUMBER_DATA_BUFFERS for your system (My Solaris 4G memory system is
limited to 1G shared memory and data buffers come out of shared memory.
Max shared memory request may need to be configured in the OS). Other
stuff uses shared memory, so you must not use all. (I use 32 buffers
and would raise if I could. Restore buffers is another entry.
I like to maximize the number of data buffers, because that allows my
backup mix to change over time without undo degradation or action on my
part.
I'm not at all sure how fragment size for MPX backups enters into these
numbers. Should it be limited to the amount of data awaiting in data
buffers or something much less? I've seen Veritas documents suggesting
fragment size is not performance related, but if during an MPX write to
tape an entire fragment must be written before going on to the next
client fragment, fragment size is very important to performance!
As always, I'm interested in hearing other opinions and fact ...
cheers, wayne
--
Wayne T. Smith -- WTSmith AT Maine DOT edu -- Systems Software Analyst
University of Maine System (UNET)
|