2014-10-14 12:37, Carew, Alan: > > > The following patches add two DPDK sample applications and an alternate > > > implementation of librte_power for use in virtualized environments. > > > The idea is to provide librte_power functionality from within a VM to > > > address > > > the lack of MSRs to facilitate frequency changes from within a VM. > > > It is ideally suited for Haswell which provides per core frequency > > > scaling. > > > > > > The current librte_power affects frequency changes via the acpi-cpufreq > > > 'userspace' power governor, accessed via sysfs. > > > > Something was preventing me from looking deeper in this big codebase, > > but I didn't know what sounds weird. > > Now I realize: the real problem is that virtualization transparency is > > broken for power management. So the right thing to do is to fix it in > > KVM. I think all this patchset is a huge workaround. > > > > Did you try to fix it with Qemu/KVM? > > When looking at the libvirt API it would seem to be a natural fit to have > power management sitting there, so in essence I would agree. > > However with a DPDK solution it would be possible to re-use the message bus > to pass information like device stats, application state, D-state requests > etc. to the host and allow for management layer(e.g. OpenStack) to make > informed decisions.
I think that management informations should be transmitted in a management channel. Such solution should exist in OpenStack. > Also, the scope of adding power management to qemu/KVM would be huge; > while the easier path is not always the best and the problem of power > management in VMs is both a DPDK problem (given that librte_power only > worked on the host) and a general virtualization problem that would be > better solved by those with direct knowledge of Qemu/KVM architecture > and influence on the direction of the Qemu project. Being a huge effort is not an argument. Please check with Qemu community, they'll welcome it. > As it stands, the host backend is simply an example application that can > be replaced by a VMM or Orchestration layer, by using Virtio-Serial it has > obvious leanings to Qemu, but even this could be easily swapped out for > XenBus, IVSHMEM, IP etc. > > If power management is to be eventually supported by Hypervisors directly > then we could also enable to option to switch to that environment, currently > the librte_power implementations (VM or Host) can be selected dynamically > (environment auto-detection) or explicitly via rte_power_set_env(), adding > an arbitrary number of environments is relatively easy. Yes, you are adding a new layer to workaround hypervisor lacks. And this layer will handle native support when it will exist. But if you implement native support now, we don't need this extra layer. > I hope this helps to clarify the approach. Thanks for your explanation. -- Thomas
