So I've read the tuning guide, I've played around with
different options for SIZE and NUMBER of buffers and I
understand the formula of SIZE * NUMBER * drives *MPX
as it relates to shared memory.
You're going to get a lot of replies. Everyone is a buffer tuning
expert.
If all you really need is "how many buffers, of what size, muxed how
many ways to how many drives" can I possibly use, skip everything
after this paragraph. It was only six years ago that most NBU
platforms were 32-bit, media servers might have only a few GB of core
and have other limits on shared memory from the OS, so size
size*number*mpx*drives was a more pressing concern. Even today, it
takes a serious amount of buffering and streams to use up 32GB,
probably more than is useful. With, say MPX 3 and 32 256KB buffers,
that's over 1000 simultaneous streams. It would take one killer
network|media-server|storage combo to keep up with that.
The NUMBER of buffers and MPX level seem to be the two
variables
Makes more sense to me to think of SIZE and NUMBER as the two
variables for a backup stream. Then think separately about MPX as a
how-many-TunedWithBuffers-streams do I deliver to a tape drive to make
it stream. (Recognizing that you may mux different classes and
schedules in different ways for backup or restore performance
considerations.)
You didn't ask for tuning methodology, but since you're in the tuning
guide already... One of the points you may have gleaned from the
tuning guide is to look at the wait and delay counters, and whether
you're measuring a data supplier or a data consumer. Understanding
producer/consumer and wait/delay together gives you a sound basis for
making changes. As does gaining the numbers to see if it's 300
seconds of delay on a 10-minute backup or 300 seconds on a two-hour
backup. That's plan A.
Plan B is empirical. (My and others' methods will be in many posts
from the last ten years if you check the archives, so this'll be
relatively brief.) Define which path, under what conditions, with
what data, you want to investigate/optimize. Strongly recommend you
work initially with one server, one invariant chunk of data, one
stream, no conflicting activity to nail down the basic limits of that
client/network/STU combination. Only after that is optimized would I
throw multiplexing into the game.
Make a test policy, say TUNE-win, full schedule, no scheduling
windows, specifying the STU, client and path that will take a
non-trivial amount of time to minimize variables yet be short enough
that the testing doesn't take forever. Say, enough representative,
unchanging data for 10-15 minutes elapsed write time. Record number,
size of buffers, wait and delay values and write time. Double only
one of the parameters. Retest/record. Do that until there is no gain
and leave that value alone. Now repeat the cycle while changing the
other parameter. Retest until no gain. Then go back to the first
parameter and change it up and down (one doubling and one halving) to
see if that needs to be revisited.
BTW, NUMBER_DATA_BUFFERS is per backup stream. Just want to make sure
you are clear on that--not sure from your note.
You have a good start on setting size/number now. Extra credit for
trying other clients/STUs and other variables, of course. Use those
values and try controlled multiplexing (you've probably maxed out a
client, so generally this will be with multiple clients. Net
bandwidth might also rule here, of course.)
Since 6.x NetBackup, you are unlikely to run out of buffers. Status
89 if you do. Regarding number versus size, there's no point in
having a huge number of small buffers or a small number of huge
buffers in any environment I've seen. The methodology above usually
shows you an optimal combination that is somewhat balanced. Sooner or
later, the allocating of a ton of little spaces, or of a few huge
(contiguous) spaces will lead you to a reasonable balance. I probably
never had speed improvements over 1024 buffers, and often much less.
DON'T FORGET NET_BUFFER_SZ! Hugely important on both Windows (in the
GUI) and *nix clients to get the data out of the client.
Here's my question. Of the four parameters:
MPX level
# of drives (I have 12 drives)
NUMBER of buffers
SIZE of buffers (must be multiple of 1024 and can't exceed the block
size supported by your tape or HBA)
The NUMBER of buffers and MPX level seem to be the two variables
here. I have MPX set pretty low (2 or 3) and NUMBER of buffers set
to either 16 or 32. When I multiply it all out, I get a hit on my
shared memory of less than a GB. My media servers are dedicated
linux hosts that only function as media servers and that's it.
Furthermore, they each have somewhere around 35 - 50 GB of memory a
piece.
With my current configuration, I'm not even scratching the surface
of the amount of shared memory that's sitting idle in my system
while my backups run at night. Is there any reason I *shouldn't***
jack the NUMBER of data buffers up to... say... 500? 1000? I've seen
some people mention that they have the number of buffers set to 64,
but can we go higher?
I've searched around to see if there's a technote on the upper limit
of the NUMBER buffers parameter. If there is such a tech note, I
can't find it.
I'd bet a beer that the limit is something like the range of an
integer on that platform--I don't think anybody codes for short ints
and such any more.
_______________________________________________
Veritas-bu maillist - Veritas-bu < at > mailman.eng.auburn.edu
http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
