Hi --
In doing some simple testing with my new AIT-3 based tape libraries,
I note that they really seem to prefer larger block sizes, despite
the admonition in Sony's manual to use 512 byte blocks for best
performance. Here's what I'm seeing using "dd" to write to the tape
device (/dev/nst0) under RedHat Linux 7.3:
bs= 512 5.4 mb/s
1024 6.0 mb/s
2048 6.1 mb/s
4096 6.7 mb/s
bs= 32k 7.5 mb/s
64k 7.3 mb/s
128k 10.6 mb/s
256k 10.2 mb/s
bs= 2mb 8.6 mb/s
The vendor-cited maximum native data transfer rate is 12 mb/s.
The "st" driver man page notes that it uses up to 16 x 128k internal
buffers, which is why I tried 2mb for a top-end blocksize. I've got
the drive's dip switches set to manually disallow compression, and
not let the driver change it. The manual for the AIT-3 drive seems
to indicate their are two buffers on the drive, an "interface" buffer
(2 MB) and a "group" buffer (8 MB), although I'm not sure what each
one is used for.
>From watching "iostat" during the test runs, the lower block sizes
look like they're starving the drive, i.e. you can see the kb/s drop
to zero every few seconds, then come back up to peak (11-12 mb/s) for
5-6 seconds, then slide back to zero, etc. At the higher blocksizes
(128k, 256k) the drive pegs 11-12 mb/s and stays there. For some
reason 2mb doesn't work as well, I'm not sure why.
>From the Amanda docs and code, it looks like the Amanda taper
allocates "tapebufs" number of 32k buffers as shared memory. Data
from the holding disk files gets pushed into the buffers, and then the
buffers flush to the tape device. It also looks like there's a
"blocksize" config directive in the tapetypes section, but the man
page says that, "The minimum blocksize value is 32 KBytes. The
maximum blocksize value is 32 KBytes." (?)
My question is this: given the better performance under the higher
blocksizes, would increasing the buffer size from 32k up to something
like 128k give me better performance? Can I obtain the same objective
(keeping the drive streaming) by just increasing "tapebufs" to some
fairly large number?
Thanks for any thoughts,
-mgs
|