As workloads have grown and computing has become more data intensive there has been a move to improve the data crunching capabilities of modern processors. Traditional scalar instruction sets have been augmented by extensions borrowing techniques from super computers in the form of SIMD or Vector processing instructions.
In the world of virtualisation these new instructions offer a number of new challenges. From efficient emulation for code generating JIT's such as QEMU to increased latency when multiplexing multiple machines under KVM.
This talk will briefly examine the history of SIMD instructions before introducing ARM's innovative Scalable Vector Extensions. We will then discuss the challenges involved in virtualising them both in software and hardware alike.
