peterjeremy at optushome.com.au
Wed May 31 12:23:20 PDT 2006
On Wed, 2006-May-31 08:05:21 -0700, Eugene M. Kim wrote:
>While watching the output of iostat -dxz -w10 -n100 to monitor the
>progress/performance of a dump(8) process straight to a tape, I found
>out something interesting and disappointing at the same time: The disk
>read throughput was exactly twice as high as the tape write throughput,
>throughout the entire dump phases 4 and 5, i.e. dumping actual inodes.
I looked into dump performance many years ago (as part of a thread
which resulted in Matt Dillon adding the existing caching code) but
can't find my notes right now. From memory, the main problems were
that dump would re-read the inode multiple times and the physical read
sizes were (pretty much) limited to the filesystem blocksize. The
caching code increases the read blocksize but this also means that
data read from the disk may not be needed. Dumping small files is
the worst case.
If you remove the '-C' parameter, you'll probably find that your disk
throughput drops to only slightly more than the tape throughput,
though the tape utilisation will probably drop further.
If your concern is tape drive utilisation (because you want it for
other tasks) or the tape drive dropping out of streaming mode, your
only real solution is staging to disk. I wrote a tool similar to
ports/misc/buffer that supported hysteresis to minimise the number
of start/stop cycles but it only marginally speeds up the total
dump time (sometimes dump runs faster than the tape and so a buffer
There probably is more scope for enhancing dump throughput.
More information about the freebsd-hackers