Hi folks,
I am currently working on a Power Management example application for a Virtual
Machine environment running on qemu/KVM and would appreciate any feedback(with
code to share shortly).
The basic idea is to provide librte_power functionality from within a VM to
address the lack(for good reason) of MSRs to facilitate frequency changes from
within a VM.
For those unfamiliar, librte_power affects frequency changes via the
"acpi-cpufreq" userspace power governor, accessed via sysfs.
The VM implementation allows for DPDK applications to request frequency changes
via the librte_power API, however requests are forwarded over a message bus to
a host monitor daemon which manages frequency changes for any number of VMs,
the daemon itself uses librte_power then to honour the VM requests.
VM: rte_power_freq_max ----> guest_channel_send_msg(pkt) ----> HOST
HOST: epoll_wait() ----> read(pkt) ----> validate_and_process_request()
----> get_pcpus_mask(vCPU) ----> power_manager_scale_core_max(pCPU_mask);
The architecture requires a number of components to achieve this:
Message Bus:
A means of forwarding frequency change requests to the host. I am using
Virtio-Serial, it gives us a secure channel that can be configured in a number
of ways. Each lcore in the VM has exclusive access to a channel. Each channel
is configured as a serial device on the VM and as an AF_UNIX socket on the
host. Both endpoints support poll/select/epoll. More information on
Virtio-Serial is here: http://fedoraproject.org/wiki/Features/VirtioSerial
VM Application:
For each lcore, a channel is opened in non-blocking mode and frequency changes
are just packets send via "write" to the channel. The existing l3fwd-power
application be reused. Each packet has format of command(Power), resource(core)
and amount(min/max/up/down).
Host Monitor:
Epoll based monitor to manage channel requests: frequency changes(after
conversion of vCPU to pCPU), VM shutdown and error events
Management CLI:
For channel management, adding channels to host monitor, disabling/re-enabling
VM requests to allow for manual core frequency management(via CLI) and
inspecting vCPU to physical CPU pinning.
Power Management:
A wrapper around librte_power to enable frequency changes for a mask of cores,
however running a virtual CPU on multiple physical CPUs is not ideal, but is
supported. The sharing of a physical CPU with multiple VMs is not supported,
while it can be attempted there is no coordination of requests from different
VMs.
Thanks,
Alan
