>>10972677We have previously made a mechanism for decoupling PCIe devices from the hosts they reside in, allowing them to be temporarily reassigned to other hosts in a PCIe cluster. This paper was about how we have extended this functionality to virtual machines, allowing remote PCIe devices (such as GPUs) to be passed through to a local VM, and also enabled PCIe peer-to-peer between such passed through devices, even if they physically reside in different hosts. We've also done a bunch of performance evaluations, showing that there is minimal overhead using our approach.
