1) As everybody has already mentioned doing 100% I/O of a workload (reads/writes/random/seq) is not representative of any real environment

2) Using a single ESX host is not an implementation customers typically deploy. How many customers with a single ESX server do you know?

3) Using RDM only lends itself to speculation as to why VMFS was absent from the tests given that most VMware customers deploy VMFS

If you gonna put out a paper comparing protocols, then it ought to be done correctly and in a manner representative of a typical VMWare customer deployment. Also, since ESX enables scale out clustering, and VMFS is the primary mechanism in doing so, a good test would be to compare the scaling aspect of it and what does performance look like in that scenario.

If you create a virtual machine on NFS it is thin provisioned. If you convert it to a template the template is already thick provisioned and so every new virtual machine you create from this template becomes thick.

It’s the normal behavior on NFS storage as far as I’ve seen it.

Robert

iometer tests using 100% read and write setting are not a good indicator for vmware performance. Random IO is the key.. We have NFS storage systems that are very fast at sequential IO and poor random IO and the VM’s are very slow.

Also, a 256Mb VM may have been an issue in their numbers…

I believe there is a larger issue going here… I think vmware wants to get out of the filesystem business and NFS systems are their way out.

That VDI demo is pretty neat. I’ll be glad when the technology to actually do that is available to the public.

As for the VMware whitepaper, I agree that the performance measurements should try to be more applicable to real-world VM activity. That would give us a more realistic comparison of Fibre Channel, iSCSI, and NFS.

Thanks for your feedback!

Dmitry

Regarding NetApp storage performace whitepaper, IMHO it’s useless. They compare sequential read and write and measure bandwidth of the data chennel (FC, iSCSI and NFS) and CPU utilization which is NOT the typical VM workload. Realistically, they should measure latency and RANDOM read/write performance of different protocols on different arrays (of the same class).
Let’s say there’s IBM FC array, Equallogic iSCSI array and NetApp NFS filer (not that NetApp can’t do any other protocols, but IMHO it’s the best in doing NAS). They cost around the same, but which one will give better performance/managebility with real world workloads - that’s the question. And we all know that bandwidth/throughput it NOT, or NOT always, a deciding factor and how much managebility, reliable backup and quick restore could matter.
From my experience running VM over NFS (on appropriately sized NetApp filer) always been a best choice, specially over 4GB Ethernet trunk or now, when ESX supports 10GB, over 10GB link.

Anyway, that’s my 2c

Dmitry

I use the same general method to prepare the host for cloning; then copy the host files using an NFS-attached host. This is not fast - templates seem to be faster here, unlike with our Linux Kickstarts - but it’s worth the cost trade-off for us. I use a couple shell scripts to rename files and correct paths with the VM’s config.

Similar to your method, I use “vmware-cmd -s register” to register the virtual.

The same caveats about SID duplication apply; I use NewSID from Sysinternals/Microsoft to address that.

Andy

Andy