On 7/20/11, David Ehle <ehle@xxxxxxxxxxxx> wrote: > Would you be willing /able to elaborate more on this idea? I've been > looking into it a bit more and while your right that it sounds simpler on > paper, it looks like doing DRBD is packaged for Ubuntu, and there is > pretty good Cook Book documentation for how to do HA NFS on top of DRBD. > > I'm a real novice on the multipath topic so I'm having trouble comparing > apples to apples to see what the pros and cons of the two scenarios would > be. I'm a novice as well (if my nickname didn't make that obvious yet :D) It's all on paper since I haven't had the time to push for it internally nor was the original client eager to put out the budget for the necessary hardware. My basic idea was this Physically ======= [VM Host] NIC1 -> Switch 1 NIC2 -> Switch 2 [Storage Node] x 2 NIC1 -> Switch 1 NIC2 -> Switch 2 Export HDD x 4 (RAID 10 or could do with RAID 1) So for VM1 mp1: multipath NIC1 -> Storage1:HDD1, NIC2->Storage1:HDD1 mp2: multipath NIC2 -> Storage2:HDD1, NIC2->Storage2:HDD2 then md mirror using mp1 and mp2 This way, if one switch fails, multipath should keep the array working. if one node fails, md mirror should keep it alive. One question I haven't figured an answer to is whether it would be better to build the array in the host and simply show the guest a single drive (less overheads? more flexible? since I could change the setup on the host as long as the guest definition points to the correct block device), or do it within the guest itself (faster I/O since the kernel is aware of more drives?)
