WOW, the talk internally at EMC is that we might have become the market leader in all flash arrays last quarter! What’s so impressive? The product went GA half way through the quarter!! What’s so Xtrem’ly hot about this product?
Consistency… it simply does what it advertises to do.. and it does it all the time!! Let me explain, it’s like looking at the fuel consumption of cars! You read the specification and it show the absolute best consumption, which in reality is unachievable, (or at least for me). Well it turns out that conservative EMC publishes the ‘on-road/real’ numbers, where the others play the theoretical specs. Like a car the only way to work out the real numbers is to fill it up and run it for a while, (perhaps putting your foot down occasionally)! That is what you should do when you test an all flash array:-
– Fill it up: to get an indication of the true usable capacity
– Run it for a while: to see what happens to performance over time.
– Put the foot down: to see what happens!
On the last point, I’m guessing you are looking at all flash for performance, so you have to stress the box to see what happens. It’s not easy because you aren’t used to your servers being the bottleneck, so get some beefy servers and load it up! I’m am warning you that you might be disappointed at what happens, on some arrays as it gets busy, services get shut down and it goes into a catastrophic spiral bleeding capacity and/or performance.
At SNIA we spent a great deal of time working out ways to test and classify the performance of Flash devices to enable you to compare. The reason this work had to be done is due to the way Flash as a media works, essentially its page oriented and needs some sort of garbage collection, it requires write levelling to improve durability, etc. If you are interested SNIA has all the information in the Solid State Storage Special interest Group, (here),
We’ve seen this behaviour before when I started at EMC I was responsible to introducing a new way to connect to storage called Storage Area Networks or SAN for short. (Yes I’ve been here for that long!) EMC as always tested out the full configuration and did the eLab job and published the real-world numbers. We got a shock when we saw the competitor’s numbers, a factor of about 100 higher!! An interesting trick, they had worked out that the fibre channel chips had a small buffer in them, so their test wrote a small piece of data to the chip and then read it back again. Fantastic wasn’t it: absolutely valid as they were writing and reading from the array, while as the same time being absolutely useless as a measure of what would happen on your site.
I use to work with a guy, George Z some of you will know him, who had a way of giving you an absolutely accurate answer, which was at the same time completely useless. Don’t get caught by this, it could be costly!