AMD hexa-core or Intel Sandy Bridge quad-core for OpenVZ?
I am looking to build a couple of new servers that will host OpenVZ and KVM virtual machines. As our experience shows, usually disk IO wins over processor performance in the long run, so there is no point in buying Xeon or Opteron grade hardware for us at 3-4x the price per server, we are better off by assembling more cheap nodes.
Currently there are two contenders for these new nodes:
880G+SB850 motherboard, 16GB DDR3 RAM, PCIe RAID
Phenom II X6 hexa-core processor @ 3-3.2 GHz (1075T or 1090T)
Since I was unable to find any meaningful comparison benchmarks between the two, I'm asking for advice: given the typical VPS workload (mainly OpenVZ LAMP servers and some KVM machines), which one would be the better choice?
I'm also interested in any comparison benchmarks showcasing CPU or IO performance regarding these two architectures that others have seen.
The speed should not be much different until you get an i7-2600 instead of i5-2500. i7-2600 would be even faster than a 1100T on many workloads . Most likely Intel will need less energy when not idle. There are of course apps which really like amd cpus, so maybe get one of each system and benchmark with your apps.
These are all interesting points, however I'm still interested in seeing any benchmarks comparing Intel Sandy Bridge and AMD Thuban.
curaga: Of course I will be running more than 4 VM's at once on a given node, but I don't see how that's relevant. Any server running even a low number of virtual machines has hundreds of processes, so single core performance, inter-core communication and cache/memory access are probably much more important than the sheer number of cores.
Kano: Yes, we are thinking about the i7-2600 as well, HyperThreading looks promising but I would die to see some benchmarks that justify the cost increase.
chithanh: Since we are mainly running OpenVZ, maximum memory size is not a big consideration since it doesn't reserve private memory for each VM, rather it shares very efficiently. A single socket node with locally attached storage will usually reach it's IO or CPU limits well before running out of memory.