Analysis Services and Solid State Disks

There’s a lot of discussion going on in the SQL Server relational world about solid state disks and their benefits, but for some reason very little has been said about how useful they could be for Analysis Services. And yet, with their impressive performance for random reads, SSAS and SSDs would seem to be the perfect combination. I think the reason for the silence is that there’s less general understanding of SSAS performance issues out there in the blogosphere, and probably also since there are fewer really large installations of SSAS not many people have had the chance to test SSDs with SSAS properly.

Anyway, the other week I was lucky enough to spend some time testing SSDs with SSAS for a customer of mine. For reasons of privacy I can’t share any significant details about what I found, and before I go on I need to point out that what I was doing was comparing the performance of a particular SSAS cube on a particular existing hard disk configuration with the same SSAS cube on a particular SSD configuration. So while it’s certainly possible to say that, in lab tests, SSDs can perform extremely well for certain tasks, in the real world you have to ask the following question: what benefit will an SSD give me for my cube, over my existing disk configuration? So in fact even if I could share the details of my tests they wouldn’t be all that useful.

That said, here are a few high-level findings:

  • Backup on the SSD was approximately 50% faster than on their existing, very respectable, hard disk configuration; Restore was 65% faster. These were the clearest examples of the SSD showing massive performance improvements.
  • For various reasons I couldn’t do full tests on processing performance, but Process Index operations were about 10-15% faster, as was Synchronisation.
  • Performance of individual queries was not significantly faster, maybe 1-2%; there were some cases where, with large numbers of concurrent users, I saw a bigger improvement of around maybe 10%. This didn’t surprise me at all because I knew that any query performance issues the customer has are related to the formula engine and not the storage engine, so IO wasn’t much of a problem in the first place.

These tests also led to some interesting behind-the-scenes discussions about how you can actually measure the impact of IO on SSAS storage engine operations, which Teo has already blogged about here and which are well worth reading:
http://prologika.com/CS/blogs/blog/archive/2010/04/13/analysis-services-i-o-time.aspx
http://prologika.com/CS/blogs/blog/archive/2010/04/16/using-xperf-to-test-analysis-services.aspx

For this particular customer, my recommendation was that it wasn’t worth the time, money and effort to move to SSDs, at least from a query performance point of view. In fact, based on several years experience of tuning SSAS implementations I find that in most cases IO is not one of the most important factors in determining query performance – it’s more often the case that the problem lies in the formula engine, and even when it doesn’t, an appropriate partitioning strategy and aggregation design can work wonders.

However, let me be clear: I am not making some general statement that SSDs are not useful for SSAS. What I am doing is giving the usual ‘your mileage may vary’ and ‘it depends’ answer. There are definitely going to be SSAS implementations where SSDs will make a massive difference to query performance, and where no other tuning technique can have the same impact. For example, I’ve heard that with distinct count measures (especially when there are a large number of concurrent users) SSDs can lead to very impressive improvements in query performance; I would assume that where there are large many-to-many relationships you would also benefit accordingly. What you need to do before you spend money on SSDs is to understand the causes of any performance issues you currently have and make sure you tune your cube to the best of you abilities.

Has anyone else had any experience with SSAS and SSDs they’d like to share?

UPDATE: the SQLCat team have published this technical note on the subject: http://sqlcat.com/technicalnotes/archive/2010/09/20/analysis-services-distinct-count-optimization-using-solid-state-devices.aspx

%d bloggers like this: