Comments on: Aside – That old spindle problem

By: Preston

Preston — Mon, 20 Jul 2009 21:52:48 +0000

That technique is certainly becoming more common – I remember a few years ago attending a Compellent web demo, where the notion of seamless automated tier migration of data depending on access was available in their arrays: i.e., RAID levels as appropriate across spindle speeds as appropriate; as data was accessed more frequently it would be moved up the access chain to increase responsiveness. (Data could also be forcibly assigned to particular levels as well.)

I think most storage systems – either array or OS/filesystem – are going to need to (no pun intended) migrate to this style of storage, so as to take advantage of the various hardware storage units that are available on a system or SAN.

By: David Magda

David Magda — Mon, 20 Jul 2009 21:42:56 +0000

It depends on how you use SSDs. Currently I think ZFS’ “hybrid storage” idea is the best thing going right now:

http://blogs.sun.com/studler/entry/zfs_and_the_hybrid_storage

You use regular spinning rust for your bulk storage needs, and slot in SSDs transparently to improve performance. It can improve write performance with SLC SSDs:

http://blogs.sun.com/brendan/entry/slog_screenshots

And optimize reads with MLC SSDs:

http://blogs.sun.com/brendan/entry/l2arc_screenshots
http://blogs.sun.com/brendan/entry/test

Most (all?) other file systems don’t allow for the inserting of a transparent shim layer.

By: Preston

Preston — Mon, 20 Jul 2009 21:12:14 +0000

Indeed, SSDs are not the magic silver bullet that some would have us believe (yet). However, regardless of whether you find SSD write performance great/good/mediocre/terrible, the article remains relevant in as much as it helps to point out the IOPS issue caused by ever increasing numbers of applications running on fewer physical systems.

By: Don

Don — Mon, 20 Jul 2009 18:05:18 +0000

One thing to take into consideration for a ‘networker’ environment.

While everyone talks about the great READ performance, no one ever mentions much about the almost 50% LESS WRITE performance of an SSD vs a hard drive.

And I would think that in the Networker world write performance is king since most of the IO’s is writing…not just the backup data but all the indexes is writes writes writes.

So I would not be surprised if you replaced your disks with SSD in a Networker use case you would get WORSE performance.