I started with a calculation like that, but I also
incorporate destaging speed. I've never used anything higher than 16 concurrent
writes, and that is for incremental backups. With incremental backups, I don’t
care as much because even if the destaging speed is poor, the amount of data is
so small that it will still be written to tape quickly.
However for full backups, I’ve found that the more
streams you have, the slower your destaging will be. I created the
graphic below to help explain it to management. The issue isn’t
fragmentation as much as it is sequential blocks of data. If you were to write
one sequential stream to a raid group then reading it back would be a
sequential operation. And in my case, SATA disks read sequential data
very quickly. But as soon as you multistream, the data “fragments” or becomes
non-sequential (really, the file sizes are so large every file is fragmented).
Image removed due to size limitation.
So to find your “sweet” spot you will have to test.
I started by building my environment and then testing with a single stream,
from client to disk, then disk to tape. This was my best case scenario,
and I was able to drive LTO3 with it.
My configuration: Dell MD1000 w/ 12 500GB DATA disks in a
Raid 5. PERC 5/e card with Adaptive Read Ahead and Write Back enabled. GPT
Partition table w/ 64KB block sizes.
The 64KB block size and adaptive read ahead settings were
key to my performance. I was going to test the same hardware with a linux ext2
file system and 4KB block sizes for comparison, but I never got the chance.
Once you have a baseline you can run more streams. You
also need to consider job length. For example, I backup an NFS server
with 16 streams to a DSSU that allows all 16 streams because 12 of those
streams take less than an hour each. I’ve also tested with 1TB SATA Disks
and 300GB 15K SAS Disks. I still rely on the 500GB SATA as my work horse
because I can drive LTO3 if I need to. Most of my DSSUs are 8 streams and push
70MB/sec to tape.
-Jonathan
-----Original Message-----
From: Victor Engle [mailto:victor.engle AT gmail DOT com]
Sent: Monday, April 26, 2010 3:39 PM
To: Martin, Jonathan
Cc: veritas-bu AT mailman.eng.auburn DOT edu
Subject: Re: [Veritas-bu] Architectural question (staging)
Thanks Martin. Good info. What criteria do you use to
determine the
number of concurrent jobs for a DSSU? Is it reasonable to
determine
the concurrent DSSU sessions based on the speed of the
clients? For
example, If I have a disk to which I can write 100MB/s
and clients
that can write 5MB/s then I would use 20 concurrent
sessions?
Thanks all for your responses!
Vic
On Mon, Apr 26, 2010 at 10:14 AM, Martin, Jonathan
<JMARTI05 AT intersil DOT com> wrote:
> We use Disk Stage Storage Units (DSSU) almost
exclusively for our backups.
> As someone has already mentioned, you can stream
many slow clients to a DSSU
> without impacting your speed significantly. To do
this with tape you would
> have to use multiplexing, which is a real
performance killer come restore
> time. The DSSUs essentially allow you to
“multiplex” to disk, then single
> stream to tape. Regarding Ed’s speed issue below,
I’ve got data here that
> directly correlates number of concurrent streams
written to a DSSU with the
> performance to tape. You’ll need to balance
this when setting this on the
> DSSU, but we get away with 8-12 streams on 14x500GB
SATA Disks in a Raid-5
> and still drive LTO3.
>
>
>
> Here we were also able to purchase a much smaller
tape library. I replaced
> a library with 8 x SDLT220 drives with a library with
2 x LTO3 drives (and
> disk storage.) The disk storage is far more
reliable than tapes and
> libraries. As Ed noted below you also benefit from
having your backups on
> disk which makes restores very snappy.
>
>
>
> -Jonathan
>
>
>
> From: veritas-bu-bounces AT mailman.eng.auburn DOT edu
> [mailto:veritas-bu-bounces AT mailman.eng.auburn DOT edu]
On Behalf Of Ed Wilts
> Sent: Sunday, April 25, 2010 3:45 PM
> To: Victor Engle
> Cc: veritas-bu AT mailman.eng.auburn DOT edu
> Subject: Re: [Veritas-bu] Architectural question (staging)
>
>
>
> On Sun, Apr 25, 2010 at 2:14 PM, Victor Engle
<victor.engle AT gmail DOT com>
> wrote:
>
> Just wanted to get some opinions about whether disk
staging units are
> worthwhile. My backup server has two BasicDisk
staging units with the
> storage units configured such that the data goes to
disk and is then
> moved to tape. I have a tape library with four LTO-3
drives connected
> via FC. So what I'm wondering is, since the LTO
drives are reasonably
> fast, and since I'm writing the data ultimately to tape
anyway, would
> it be better to just write directly to tape. The
disk is just old
> fashioned spinning disk with no de-duplication so
there are
> operational costs for the disks. All tape and disk
storage units are
> local to the backup server. I'm thinking it would be
better to add LTO
> drives and eliminate the disk for now and maybe
later add a
> de-duplicating disk unit.
>
> We worked with disk staging units for at least a
year before we mostly gave
> up. The biggest challenge we ran into was that
destaging was too slow. Even
> though we proved to Symantec that we could read from
those disk drives at
> over 100MB/sec, we could never destage even half
that fast. We had an open
> case with Symantec for a VERY long time before we
agreed that it wasn't
> going to get fixed.
>
> Under what circumstances does it make sense to stage
data on disk. I
> would appreciate hearing what your thoughts and
experiences are with
> regard to disk staging.
>
> There are times when DSSUs make sense.
1. If you don't have a tape drive
> free but want to do a backup anyway - we still use
DSSUs for things like
> small backups of Oracle archive logs. 2. If
you need to throttle your
> backups, especially across things like a bunch of
virtual servers on the
> same physical server. NBU only allows you to
set the maximum jobs per
> client name, not per client. DSSUs make an
acceptable choke point for
> clusters. 3. If you have small backups,
but don't have a lot of them at
> once, multiplexing may not buy you enough performance
boosts. Use DSSUs to
> write those little jobs to disk and then destage
them at once.
>
> If you currently multiplex, realize that your
restores are going to be
> slower than if you don't multiplex. All tapes
created from a DSSU destage
> are non-multiplexed so your restores can go faster.
>
> DSSUs also give you a staging area for
restores. If your tapes go offsite,
> you may still be able to do a restore from the
staging unit the next day (or
> longer) depending on how big your stagig units
are. NBU is smart enough to
> realize that if the same data is on both disk and
tape and you kick off a
> restore, the restore will automatically come from
disk.
>
> In general, I'd say that there is a place for DSSUs
but it's not the great
> benefit we thought it was going to be.
>
> .../Ed
>
> _______________________________________________
> Veritas-bu maillist -
Veritas-bu AT mailman.eng.auburn DOT edu
>
http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
>
>