ORIGINALLY POSTED 13th April 2010
15,787 views on developerworks
So after all the EMC hype, and all the claims that the DS8000 was “pushing up daises” it seems that IBM have managed to announce the GA of ship our sub-lun hot-spot management, aka Easy Tier before EMC have shipped their FAST function.
Now, before you all jump and say EMC have been shipping FAST since last year, I don’t count version 1 as anything more than LUN migration, something we’ve been shipping in SVC since 2003. The only real version of FAST that matters is version 2 where it can work at a sub-lun level.
Anyway, as you may have seen, and there has been much discussion amongst us storage twits out there, that today IBM announced Easy Tier functionality on the DS8700. Yet another sign that by no ways or means is the DS8K platform being run down.
Easy Tier provides the necessary mechanisms to detect hot extents within a system and if you want it to automatically migrate those extents to a higher tier, it will.
The devil is in the detail and while some people look as this function as a cache, its not really – certainly not in the way we think of processor or storage RAM based caches. Conceptually I guess you can think of it in a similar way, but the “age-out” period of a RAM cache is seconds. If you were to start migrating extents from HDD to SSD or back again with a period of seconds, your whole storage system would spend its time (bandwidth) migrating and never servicing host I/O. You need to take a longer time to look at the data, days to weeks and make decisions based on longer term hotness.
Now all of this I/O analysis means lots of meta-data. With the same function being added to SVC as we speak (as I’ve said numerous times before) that leaves a few interesting issues. SVC supports 4 million extents, and of course each node can be accessing that extent, so 8x 4M, and now you are keeping that data for days and weeks… that’s A LOT of meta-data – unless you come up with some cunning meta-data compression techniques that require little or no processing overhead and of course are loss-less… So the DS8000 and SVC teams have been working closely with the Almaden storage research team to come up with techniques to gather, analyse and store this meta-data long term, within the subsystem, without loss of system performance. The same source code is being used for both DS8000 and SVC. With SVC’s variable extent size, you will be able to tune the granularity.
The whole idea here is to improve response time. SSD while great at providing huge IOPs at low latency are expensive (compared to the same capacity of magnetic media) so unless you have an application that requires huge IOPs from a small amount of capacity, todays traditional arrays as I’ve discussed before either can’t handle the IOPs, or with SVC can, but only to a relatively small amount of capacity.
Easy Tier allows the system itself to determine which extents will get the most benefit from being placed on SSD, with the primary goal to reduce latency. This means a small number of SSD can be used to accelerate a very large environment, and with SVC an investment in SSD now can be spread even further with Easy Tier is available.
Its interesting some of the discussions, EMC instantly try to start pick holes in it – storagebod is confused, and most of all, people want more information.
There are a couple of papers covering some early analyst thoughts on the launch pages, for further information, contact your IBM account team, or business partner.
Edited 14th – to clarify as per comments