Probably best to compare different levels of virtualization on the gpu with the counterparts on the cpu, just as a rough idea.
virgl + llvmpipe on the host = qemu with software emulation for different architecture between host and emulated guest. Dead slow.
virgl + hw gl driver on the host = qemu, but with a magic JIT that translate the emulated guest's instructions into hw instructions for the host. Not sure this exists as open-source. The JIT tends to seriously hurt compared to native performance.
gvt = paravirtualization like Xen, guest code from userspace runs like native, but there's quite a bit of glue and overhead involved still in both host and guest kernel, but performance tends to be nice.
sr-iov (nvidia has that in hw afaik) = hardware virtualization, guest kernel talks directly to the gpu with only very rare exceptions that need special handling in the host. think kvm. Upside is mostly better isolation and less need for adjustements in the guest kernel (it still needs some changes to the gpu driver usually, but much more minor compared to something like gvt).
virgl + llvmpipe on the host = qemu with software emulation for different architecture between host and emulated guest. Dead slow.
virgl + hw gl driver on the host = qemu, but with a magic JIT that translate the emulated guest's instructions into hw instructions for the host. Not sure this exists as open-source. The JIT tends to seriously hurt compared to native performance.
gvt = paravirtualization like Xen, guest code from userspace runs like native, but there's quite a bit of glue and overhead involved still in both host and guest kernel, but performance tends to be nice.
sr-iov (nvidia has that in hw afaik) = hardware virtualization, guest kernel talks directly to the gpu with only very rare exceptions that need special handling in the host. think kvm. Upside is mostly better isolation and less need for adjustements in the guest kernel (it still needs some changes to the gpu driver usually, but much more minor compared to something like gvt).
Comment