On Thu, 23 Jun 2005 at 11:46am, Mitch Collinsworth wrote

> On Thu, 23 Jun 2005, Joshua Baker-LePain wrote:
> 
> > Something else I heard on another mailing list:
> >
> > "However AIT-4 (unlike AIT-1 til -3) appears to write fill
> >  bytes onto the tape if it's not fed with data quickly enough,
> >  thus wasting lots of capacity."
> >
> > The person said they heard it somewhere and hadn't seen it confirmed.  If
> > true, though... yuck.
> 
> Well, maybe, maybe not.  DLT8000 did that, too.  Lots of people griped
> bitterly about "poor performance" with their DLT4000's and 7000's.
> Problem was usually that their 1-pass backup software kept starving the
> write buffer and the drives had to shoe-shine in order to deal with it.
> The DLT8000 had variable speed write, which meant the tape kept streaming
> and the data was laid down as fast as it came in, even if that was slower
> than what the tape could handle.

Indeed -- the other person referred to the AIT-4 behavior as "DLT 
syndrome".  But isn't "variable speed write" different than "writing fill 
bytes"?  Does DLT8000 lose capacity when not writing as fast as it can?

> The good news for amanda users is that when you stage your dumps to
> holding disk, you eliminate the most frequent cause of the data
> starvation problem, which is the backup program scouring the partition
> looking for which files to backup today.  Once you have your data on
> the holding disk you're unlikely to starve the tape drive and it can
> stream the data at full speed onto the tape.  If it can't, you have a
> h/w problem that is typically easy to fix.

Actually, with LTO3, I'm a bit worried about keeping the tape streaming 
when writing from staged dumps.  Native transfer rate of LTO-3 is stated 
as 288 GB/hr, which is about 76 MiB/s.  That's more than most single 
spindles can handle, *especially* if you're trying to write to tape while 
other dumps are coming in to the holding disk.  Looks like I'll have to do 
some work on the server end to really get this thing going.

-- 
Joshua Baker-LePain
Department of Biomedical Engineering
Duke University