Flash Storage – My take.
Everyone is taking about Flash these days, and I suspect the coming #VMworld will be swarming with new startups entering the Flash storage or Hybrid Storage market.
So I though I’d give you my ideas about Flash in a normal sized virtualized environment.
The setup I work with often consists of 2 EMC VNX5700 in a Metro-cluster setup with EMC VPLEX’s in front of them, also known as a Stretched Datacentre.
Having recently upgraded to this from an aging HP EVA 6400 the performance boost from the VNX’s is quite huge. And the Tiering actually moved alot of dead data to SATA, which helped that databases gain a nice boost.
But at the moment we’re nowhere near the capacity of the VNX’s. And actually it seems to be overperforming somewhat. I’ve created a 10 disk, disk group with 2xraid 5 for a small exchange environment. All 10 disks are 15.000 RPM FC drives. If we set the max IOPS per disk to around 200, that should max give me 2000 IOPS from that disk group. However during backups I see it consistently doing 6000 IOPS for around 2 hours. At first I found this odd, but verified that the VM also saw 6000 IOPS during that period.
This is due to caching, and a nice read-ahead algorithm. Both the disk itself has a small Cache, the VNX’s has a cache + some of the SSD’s are used for the VNX FAST cache. On top of that the VPLEX’s also has a Cache. So I’m getting more IOPS from the physical disks than I should :).
So at the moment I have nothing running that makes the VNX’s come anywhere near their performance limits. However we’re in the process of building a few Datawarehouses, and I suspect/hope that these will be able to put a bigger load on the VNX’s.
Should we ever hit the limits of the VNX’s there are currently quite a few options.
And this is at the moment the road I would take:
1. The Datawarehouses are currently mostly read intensive, so sticking a few SSD’s into the ESXi Hosts and enabling vSphere Flash Read Cache.
2. The ETL machines are a bit more write intensive so to help out on that I would definitely go for PernixData’s FVP solution, being both a read and write cache. And I’ve only heard good things about them from fellow vmug’ers who had this in their production environment.
3. If SSD’s alone isn’t enough, sticking Fusion IO cards or the equivalent in the machines should give another big boost. And the good thing about these is, that they also work with PernixData’s FVP solution.
4. If all else fails and we can’t get enough IO’s into the hosts, something like All Flash Arrays, might be an option. But that would really have to be tested performance wise. And these are very easy to put behind the VPLEX’s and presenting those to the ESXi hosts.
So I’m fairly certain our performance scale up options are quite good.
Looking forward to VMworld starting next week in San Francisco!.