Quantcast
Channel: VMware Communities : Blog List - All Communities
Viewing all articles
Browse latest Browse all 3135

Testing Virtual Flash Read Cache

$
0
0

So I have been playing with vFlash on an ESXi 5.5 host. My host is a Dell R630 running Intel Xeon E5-2680 with 2 sockets, 12 cores per socket and 512GB RAM.

I have a 800GB MLC SATA SSD disk configured as a Virtual Flash Resource.

My SAN is a Dell Compellent with SSD for Tier1 and 10K disks for Tier3, we don’t have any Tier2. Now, I am not going to go over how you configure VFRC as there are a lot of information on that out there already. This is more about the results.

I setup a VM running Windows 2012 with 2vCPU, 4GB RAM, then cloned it and configured the second one with vFlash.

At first I gave it a 8GB vFlash file with 4K blocks and ran the iometer tests you can see in the screen shots (middle results).

Then on the second pass, I gave it a 8GB vFlash file with 64K blocks (right most results).

Both tests were done with a 8GB iometer file.


Now, I am just learning how vFlash works, but from what I read you will achieve the best performance when your workload IO size matches that of the vFlash file, when the file is big enough to hold your entire workload, and when your workload is more read-heavy then write-heavy. Which, if I am reading this right, seems to be what I’m seeing.

It’s interesting that with the 4K test, even with the 4K block vFlash, you get more improvement on the random workload then in the sequential. With the 4K sequential, you almost get no improvement at all.


Is anybody taking advantage of VFRC? What type of workload are you using it for? How do you have it configured? I’d be curious to see how other people are using it.

The other thing I found out is that, since in my case ESXi is installed on an SD card, I would like to use this local SSD disk to store the scratch partition and coredumps, and the information I got from tech support is that putting that on a VFFS volume is not supported.

Also, I am seeing some errors on the vmkernel log during the tests (when using the cache) that according to support, can cause performance issues, but not always. They don't have any published information on this yet, in case anybody has seen it this is what I am seeing:

 

2015-09-24T20:23:11.451Z cpu14:33115)WARNING: VFC: VFC_TreePrepPage:1170: Can not insert key 3133622 into policy Busy 195887108


So for now I am reverting this back to a VMFS datastore, putting the scratch partition, coredumps and using it for old-fashioned host swapping.


 



Viewing all articles
Browse latest Browse all 3135

Trending Articles