I was fortunate to get my hands on a couple of FusionIO cards and put them into my lab. If you have never heard of FusionIO, they are Flash storage married to a PCIe card. You can see the size comparison to the EMC NS-120 disk array that is in the lab.
FusionIO claims that their cards can do 100,000 IOPS. Of course on their website they say that they achieve 100,000 IOPS using 512b IO segments. I wasn’t sure how useful this would be in the real world, but I used this as a starting point of comparison to my EMC NS-120. I decided I would leverage my VMware View 4.5 environment in these tests.
I created a VM that would run IOmeter as a service on boot and connect to a master IOmeter VM. I then used VMware View to scale to 50 worker VMs. All of the worker VMs are running on two Cisco UCS B200 M1 blades with 2 Quad Core Xeons and 48GB of RAM. The FusionIO cards are in a Cisco UCS C-210 rack mount server with 2 Six Core Xeons and 48GB of RAM.
As you can see in the results below, writing directly to the FusionIO card with 512b segments yielded ~90,000 IOPS.
Let’s compare that to a LUN of 5 200GB FC EFDs on the EMC NS-120 connected to UCS blades via Fibre Channel. Hmm… about ~30,000 IOPS.
Seeing these results made my eyes brighten as I thought about all of the potential they would have if I could use them for storage in a View 4.5 environment.
Josh,
Interesting. What is the setup on the C210 ? Are you running Linux/NFS to serve the Vsphere B200’s with Fusion IO storage ? In that case you must be running a blazing fast network to serve the 90000 IOPS to the B200’s.
Very much interested in this design as I hear that Fusion IO is going to be a supported option in UCS.
I should have clarified that the C210 is running VMware ESX 4.1. To get 90000 IOPS the VMs were running directly on the C210. I need to do another post with the rest of my results.
I tested various iSCSI Virtual Storage Appliances including Lefthand (HP) P4000, Openfiler, and Windows Storage Server. All of them topped out at about 30000 IOPS. My suspicion is that this was not a network limitation, but a single-threaded limitation in the software of converting IO to iSCSI as they all maxed out 1 virtual cpu.
I think a better test will be to test the max throughput instead of IO, however I was interested in validating FusionIO’s 100000 IOPS claim.
I don’t think that the UCS blades will be supporting FusionIO anytime soon, since that would require some kind of additional mezzanine port. However the rackmounts all have PCIe slots so multiple FusionIO cards can be supported.