A road named NoNameUno

If you know the concept behind the song "Band on the run...", you might have already guessed how I'm going start this blog, otherwise just forget about the name :). If you are not aware of them, sr-iov and mr-iov are specifically to enhance the virtualization to a higher level of performance. Virtualization at the commodity hardware started back about a decade ago. x86 family being widely accepted, the first wave of virtualization on this family of hardware took the emulation path. A basic tenet of it is that any instruction that have a visible effect to processor state needs to be trappable. There are about 17 instructions in x86 that were not in this class, hence emulation was the first approach to it. With that in mind, the next step was the para-virtualized I/O implementations to share a hardware I/O device among multiple virtual machines. Para-virtualization came at a cost on I/O performance. Virtual machines saw around 50% I/O thru-put of real device capacity. So there is currently ( as of 2010 ) very extensive research and development going on in this area. The basic problem is how to boost the I/O performance in virtualized environment. And the first approach is to tear-out the I/O processing. Basically the idea is to think that I/O consist of two main parts: Data payload path; Control and Management path. If data payload path can be optimized using suitable hardware and software architecture modification or enhancement then we could have a chance to realize thru-put comparable to real device capacity. SR-IOV, MR-IOV, IOAT, and others are some efforts toward that direction to achieve this goal.