Hi Bill,
Remember me? It's Steve; I teamed up with you in NBU training.
Anyway, regarding your question:
The problem is, that you can never stream your drives with 8x64k buffers. The
math, is that you want your total buffer size to equal the maximum throughput
of your drives. This actually diminishes your throughput. The drives *want*
to be streamed.
You are correct, in that 64K buffers are what's recommended by Veritas. Do not
let this be your gospel. Test!
I'm using DLT7K's, so I'm not sure what the throughput on 8K's is, but here's
how I'm configured:
I read this on this list a few months ago, and my own testing confirms:
I expect roughly 6MB/sec throughput.
divide 6291456 bytes (6MB/sec throughput) by 65536 bytes (64k buffer size) =
96 NUMBER_DATA_BUFFERS
Make sure that you are allocating enough shared memory to handle this.
HTH,
Steve
On Friday 27 September 2002 16:16, William Enestvedt wrote:
> I have read with interest recent messages (and the Veritas web page at
> http://seer.support.veritas.com/docs/183702.htm) about tuning
> NET_BUFFER_SZ, SIZE_DATA_BUFFERS and NUMBER_DATA_BUFFERS for better
> performance...but now I realize I'm confused about which way I need to
> focus my attention. My environment is NBU 3.4 on Solaris (the Media Servers
> run 2.6, and the Master Server runs 8); each Media Server has a Sun L9
> library, which holds a Quantum 8000 DLT and nine tapes. In my bptm logs, I
> find the following entries side-by-side:
> ---- begin log sample ----
> 21:51:46 [29568] <2> fill_buffer: [29567] socket is closed, waited
> for empty buffer 3 times, delayed 3 times, read 4738656 Kbytes
> 21:51:46 [29567] <2> write_data: waited for full buffer 12820 times,
> delayed 135496 times
> 22:25:26 [29971] <2> fill_buffer: [29970] socket is closed, waited
> for empty buffer 19242 times, delayed 22981 times, read 9905792
> Kbytes
> 22:25:26 [29970] <2> write_data: waited for full buffer 9057 times,
> delayed 30164 times
> ----- end log sample -----
> Here's my questions:
> 1. Have I changed things for the worse or the better by following
> Veritas' suggestions and making the following changes? NET_BUFFER_SZ = 64k
> SIZE_DATA_BUFFERS = 64 k
> NUMBER_DATA_BUFFERS is left empty, defaulting to 8
> 2. Is it odd that some jobs wait for empty buffers, while other don't?
> 3. What's the difference between the fill_buffer entries and the
> write_data entries? Do the write_data complaints about waiting for full
> buffers refer to the tape library going idle because NBU can't supply it
> with enough data? If so, then what do the fill_buffer complaints about
> waiting for empty buffers refer to? Thanks for any suggestions anyone can
> offer.
> -wde
|
