[BBLISA] 10+ TB RAID experiences?

[Bill Rucker]

Somewhere around Thu, Oct 04, 2007 at 04:18:48PM -0400, a message
from Dean Anderson went like this:

I'll answer both of these together, as best I can...

> > For my own knowledge, why is having snapshots on their own volume better 
> > than cluttered?    Any particular scenarios where a separate volume for 
> > the snapshopts works better than cluttered?

It all depends on what your set up is. I won't say there is a "benefit" of one
over the other, just that its possible. In fact, the Celerra can do either or.
The savevol if placed in the same array as the production filesystem takes up
space in that same array, but not necessarily within the filesystem. It creates
another storage area within that array. This is the default. The nice thing is
the checkpoint storage can be placed on a slower (ATA) storage array. It does
need to be within the same frame though. The iSCSI checkpoints are much the same
unless you are using replication. 

> Probably should let Bill respond, but I'll try to take a stab at it:
> 
> Removing the snapshop to another volume allows most likely allows the
> chain of copy-on-write updates on a modified block to be reduced. This
> length of this chain reduces the filesystem performance similar to
> filesystem fragmentation (I'd assume the chain goes most recent back, so
> the 'most recent' change is fastest in that respect).  Keeping those
> altered copies of blocks in the middle of the file probably kills disk
> performance on sequential reads, in the same way the disk fragmentation
> does.
> 
> So I suppose at worst they just they keep only 2 copies of the unchanged
> blocks, (one on each volume), with a long chain of updates on one volume
> and and short chain of updates on the 'fast' volume.  Maybe, if they are
> really clever, they can keep only one copy of the unchanged blocks. I
> don't know...

Okay, I am not an expert with the Celerra, but I do know they only need to keep
one copy of a saved block. If that block is changed in multiple checkpoints,
there is only one actual block on disk in the savevol. There are however,
multiple pointers to it within each diskmap that belongs to each checkpoint.

The other point I can speak to is that there is no apparent performance
degradation regardless of the number of checkpoints. We currently have a pretty
fair number of checkpoints, some as old as 6 months, against a number of heavily
used filesystems. Our performance is exceptional and we have not had any
indication of declining performance after having the Celerra in production for
almost 18 months. 

I'll qualify this with the statement that this is my first experience with EMC.
When we purchased the Celerra, we were warned by some that EMC equipment,
especially the Celerra, was not as advanced as the NetApp. I know NetApp
equipment is very good, and if I had my way 18 months ago, we'd have NetApp. I
was overruled and I have no regrets in the decision we made. Take that for what
its worth.

Bill

> 
> The procedure for updating a block of a file is probably like this: 
> 
>   The slow volume gets a copy-on-write update as usual. 
> 
>   The 'fast' volume, just overwrites the original block of that volume,
> which is always "latest".
> 
> This makes the fast volume stay as sequential on disk as possible. But
> I'm just guessing that this is how it works.

I believe this is correct.

> 
> 
> > Also, any other configuration and/or performance differences between the 
> > netapp and emc?
> 
> I've only used the NetApp.

And I've only used the EMC equipment. Sorry. 

Bill