On 22/06/2015 09:10, Igor Mammedov wrote: > So far HVA is unusable even if we will make this assumption and let guest crash. > virt_net doesn't work with it anyway, > translation of GPA to HVA for descriptors works as expected (correctly) > but vhost+HVA hack backed virtio still can't send/received packets. > > That's why I prefer to merge kernel solution first as a stable and > not introducing any issues solution. And work on userspace approach on > top of that. Also, let's do some math. Let's assume 3 network devices per VM, one vhost device per queue, one queue per VCPU per network device. Let's assume the host is overcommitted 3:1. Thus we have 3*3=9 times vhost devices as we have physical CPUs. We're thus talking about 108K per physical CPU. >From a relative point of view, and assuming 1 GB of memory per physical CPU (pretty low amount if you're overcommitting CPU 3:1), this is 0.01% of the total memory. >From an absolute point of view, it takes a system with 60 physical CPUs to reach the same memory usage as the vmlinuz binary of a typical distro kernel (not counting the modules). Paolo > Hopefully it could be done but we still would need time > to iron out side effects/issues it causes or could cause so that > fix became stable enough for production. > -- To unsubscribe from this list: send the line "unsubscribe kvm" in
