ORIGINALLY POSTED 15th September 2007
6,751 views on developerworks
Following on from Chris’ comments in “Pause for Thoughtput” I had a look around and there are some interesting posts being made.
Being one of those close to the “Virtualmania” scene I had a look through the second paper Gear6 have written, “Provisioning for Predictable Performance”. Jumping to the section discussing why storage is lagging behind and needs super-scalable caches to catch up. Reading this seems to imply that nothing other than caching is available today – and seems to miss a very big point. Caching is not the only solution to the problems outlined, and I’m not sure how caching solves problems like de-provisioning :
However, de-provisioning storage capacity is far more difficult. Of the layersdiscussed, storage is the only one that retains persistent data. Therefore, anytime a storage administrator chooses to reclaim capacity, all potential users ofthat capacity must be notified and confirm that their underlying storage will nolonger be available. While simple in concept, most workplace realities make thischallenging…to the point where any administrator looking to de-provisionstorage will likely need to make one or multiple backups.
It seems strange to me that a paper linked from an article discussing “Virtualmania” doesn’t discuss the fact that Storage Virtualization – when done well – solves just that problem using its ‘killer-app’ – online data migration. No need for backups, or to spread the pain to all users. Simple de-provision the virtual lun and re-provision – or migrate all virtual luns on the box to replace it.
Similarily it discusses provisioning for performance :
This often leads to grossly over-provisioned storage volumes made up of manydisk spindles creating a capacity management dilemma. If a development teamrequires a high-performance storage volume, it will likely take a large number ofspindles that will ultimately be difficult to reclaim.And even this level of provisioning can only achieve some benefits in terms ofthroughput or bandwidth. Access time, or latency, cannot be solved by relyingon mechanical disks, leaving current storage performance provisioning onlypartially capable of keeping up with the network and server layers.
So here again using Storage Virtualization means you already have your tiers of performance and you simply provision a virtual lun from an existing pool or group. The large numbers of luns can be shared between many applications, as long as the total performance requirement does not exceed the capabilities of the stripe set.
I’m not saying that caching is a bad idea, just that when you combine caching with virtualization its a double gain and it seems strange this paper does not discuss both in parallel.