Amanda-Users

Re: Writing to tape performance issue?

2007-01-05 17:11:58
Subject: Re: Writing to tape performance issue?
From: Jon LaBadie <jon AT jgcomp DOT com>
To: amanda-users AT amanda DOT org
Date: Fri, 5 Jan 2007 17:01:41 -0500
On Fri, Jan 05, 2007 at 03:53:32PM -0500, Mark Hennessy wrote:
> Many thanks for the answers to my previous questions.  I'm just trying to
> iron out a few minor issues with my backup cycle at this point.
> 
> It would appear that based on the "Tape Time" value provided in the report
> that the rate at which data are written out to the tape is 3.7 MB/s.  The
> drive is rated by the manufacturer at 11 MB/s or 22 MB/s, so it would seem

Assuming you are using software compression, the 11 MB/sec is the
figure of interest.


> that 3.7 MB/s is a bit slow, yes?  Is "Tape Time" solely the time that Amanda
> is writing dumps directly to tape?  Does that include other interactions that
> would add a substantial amount of time as well?
> 
> I'm using Amanda 2.5.1p2 on FreeBSD 6.1 and a Quantum SDLT220 drive.
> 
> STATISTICS:
>                           Total       Full      Incr.
...
> 
> Tape Time (hrs:min)        5:24       1:07       4:17
> Tape Size (meg)         72005.8    36674.1    35331.8
> Tape Used (%)              71.0       36.2       34.9
> Filesystems Taped            60         11         49
> Avg Tp Write Rate (k/s)  3793.3     9408.1     2342.3
> 

As I understand amanda's working, two processes are started
that server as "masters" for the dumpers and the taper procs.
The taper waits til there is something to tape (a completed
DLE on the holding disk for example) and forks of a child
taper process to actually do the taping.  The time that
child exists is what is recorded as "taping time".

Assuming the DLE's being taped are coming from a holding
disk, I would expect the speed of taping to be similar
for full and incremental dumps.  There is some overhead,
and that would be more significant for the smaller
incrementals, thus a small difference is to be expected.
But your difference in rate 9.4 vs 2.3 MB/sec is very
large.  Thus it seems no holding disk is involved.
Do you expect you are using a holding disk?

If not, here is what I think is happening.  Your dumps
are skipping the holding disk, going "direct to tape".
I.e. the dumper for each DLE connects to a child taper
process via a pipe.  Now the taping rate is dependent
upon the dumping rate as well as other factors.

Incremental dumping rates are generally slower than
level 0 dump rates.  Perhaps more overhead in jumping
around getting individual files than in getting all
the files in a directory.  That extra overhead probably
causes the pipe feeding the child taper to go empty
sometimes and thus taping stops.  But time continues
to accumulate and the total tapeing rate drops.

Note your level 0 taping rate, 9.4 MB/sec is reasonably
close to your tape drive's rated 11 MB/sec.  Less dumper
overhead keeps the pipe to the taper filled better.
Thus fewer or no stops in taping.

Why aren't you using a holding disk?

-- 
Jon H. LaBadie                  jon AT jgcomp DOT com
 JG Computing
 4455 Province Line Road        (609) 252-0159
 Princeton, NJ  08540-4322      (609) 683-7220 (fax)

<Prev in Thread] Current Thread [Next in Thread>