Amanda-Users

Re: Delaying tape flush ?

2007-03-30 16:21:30
Subject: Re: Delaying tape flush ?
From: Joshua Baker-LePain <jlb17 AT duke DOT edu>
To: Guy Dallaire <clepeterd AT gmail DOT com>
Date: Fri, 30 Mar 2007 15:53:54 -0400 (EDT)
On Fri, 30 Mar 2007 at 3:15pm, Guy Dallaire wrote

I noticed that my LTO-2 tape drive was "starving" for data and
shoe-shinning.

I was dumping to tape at 12Mb/sec and the minimum speed to prevent
shoe-shinning on an LTO-2 drive is 19Mb/sec.

So I did a check by timing a 'dd' of 10 Gb to the tape drive. It was indeed
topped at 12 Mb/sec. It took 13m51s

What exactly was the dd command you used to test with? Specifically, what blocksize did you use?

Now, I replaced the antedeluvian adaptec 2940 with an adaptec 29160 ultra
160 car and redid the exercice.

This time, it took 6m51s, so I was running at about 24 Mb/Sec. That was ok.
Not very fast, bot OK.

Now, when I looked at my amanda log this morning, I was very sad to see the
following:
*snip*
Avg Tp Write Rate (k/s) 17549.5    20734.5    12469.2
*snip*
Now, I figure the problem is that when I did the "dd" test, all my holding
disk had to do is read the files. Now, it has to read the files whil amanda
is pounding at it, wrinting dump files coming from the clients... and
unfortunately, it seems my single SATA disk can't keep up with all those
I/Os

Quite possible.

What are my options ? We can't afford to buy a beefy server just to do
backups. We are uning a centos beige commodity box. Even then, the server
seems to be I/O bound. What could I do anyway ? Buy a faster disk subsystem
? What kind of subsystem ?

The first thing to look into is if a larger blocksize will help you. On my LTO3, going to 2MB blocksize significantly sped things up. Test with larger blocksizes (via dd or tar) and, if they help, recompile amanda to allow you to use a bigger blocksize.

In terms of beefing up your disk system, if you could just add a drive or two and do a software RAID0, that would speed things up. But try the blocksize fix first.

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

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