From: Kevin Thorpe <kevin.thorpe@pibenchmark.com<mailto:kevin.thorpe@pibenchmark.com>> Date: Friday, June 21, 2013 8:14 AM To: "hpdd-discuss@lists.01.org<mailto:hpdd-discuss@lists.01.org>" <hpdd-discuss@lists.01.org<mailto:hpdd-discuss@lists.01.org>> Subject: [HPDD-discuss] Wondering about running OSTs on my cloud platform. Hi, we've been looking at private (and maybe hybrid) cloud provisioning. We've got an architecture for the traditional SAN + hosts model but we're interested in the SANless SAN devices that Nutanix offer. We realise that a hardware SAN device is more likely to deliver performance but we think we could get away with using distributed local storage on the VM host servers. Do you think that running an OST VM on each of our VM hosts utilising local disks would be a sensible thing to do? I realise this would eat into RAM and processor on each host, but new hosts would be far cheaper than a SAN device. I'm particularly interested in the ZFS option so se get SSD cache on each node. The only thing that I can see that Nutanix do that Lustre can't is migrate data to the nodes using that data, or am I wrong on that? -- Kevin Thorpe Chief Technical Officer PI Benchmark Sorry, bad mail options, resending with proper format Running an OST on a VM, critical item is how much the VM throttles IO/network bandwidth. Lustre itself is happy on VM but getting real-iron IO bandwidth can be an issue. You should be able to test that without Lustre, or with a fairly simple Lustre setup. You could compare sgpdd-survey results with/without the VM. Second question is what is the delta between 'local disk' and the storage solution? That's totally dependent on the iron you have, and what you consider to be an acceptable cost/performance trade. Again, you can benchmark the local disk without Lustre and have a fairly good estimate. Lustre does not migrate data, you are correct in that. cliffw