[dpdk-dev] [PATCH v1 0/10] Policy Based Power Control for Guest

[Ananyev, Konstantin]

Hi Dave,

> This patchset adds the facility for a guest VM to send a policy down to
> the host that will allow the host to scale up/down cpu frequencies
> depending on the policy criteria independently of the DPDK app running in
> the guest.  This differs from the previous vm_power implementation where
> individual scale up/down requests were send from the guest to the host via
> virtio-serial.
> 
> It's a modification of the vm_power_manager app that runs in the host, and
> the guest_vm_power_app example app that runs in the guest. This allows the
> guest to send down a policy to the host via virtio-serial, which then allows
> the host to scale up/down based on the criteria in the policy, resulting in
> quicker scale up/down than individual requests coming from the guest.
> It also means that the DPDK application running in the guest does not need
> to be modified in any way, it is unaware that it's cores are being scaled
> up/down, reducing the effort in implementing a power-aware infrastructure.
> 
> The usage model is as follows:
> 1. Set up the VF's and assign to the guest in the usual way.
> 2. run vm_power_manager on the host, creating a channel to the guest.
> 3. Start the guest_vm_power_mgr app on the guest, which establishes
>    a virtio-serial channel to the host.
> 4. Send down the profile for the guest using the "send_profile now" command.
>    There is an example profile hard-coded into guest_vm_power_mgr.
> 5. Stop the guest_vm_power_mgr and run your normal power-unaware application.
> 6. Send traffic into the VFs at varying traffic rates.
>    Observe the frequency change on the host (turbostat -i 1)
> 
> The sequence of code changes are as follows:
> 
> Firstly, two new API calls are added to the ethdev layer
> 1. One to convert a VF id to a PF id. In the patchset
>    this id is a MAC address. This is needed so that the host can map the VFs
>    in the profile to PF so in can monitor the traffic on the relevant PF at the
>    host level.
> 2. The other function is to read the low-level traffic throughput on the NIC.
>    Currently this API reads a NIC register for speed, but we are looking at
>    using a more generic way to get these stats, suggestions welcome.

Why do you need a server (host) to collect RX/TX statistics for VM?
Such method seems to have a lot of limitations:
- no clear method to identify to which VM that VF belongs.
- rely on HW ability to provide such statistics for PF
  (limited HW support).
- wouldn't work if PF is not controlled  by the same DPDK app. 
Why not to make  it client(VM) responsibility to collect that statistics and
periodically send it to the server? 
Then server just will have to process that data and make decision.
Konstantin

> 
> Next we make an addition to librte_power that adds an extra command to allow
> the passing of a policy structure from the guest to the host. This struct
> contains information like busy/quiet hour, packet throughput thresholds, etc.
> 
> The next addition adds functionality to convert the virtual CPU (vcpU0 IDs to
> physical CPU (pcpu) IDs so that the host can scale up/down the cores used
> in the guest.
> 
> The remaining patches are functionality to process the policy, and take action
> when the relevant trigger occurs to cause a frequency change.
> 
> [01/10] net/i40e: add API to convert VF Id to PF Id
> [02/10] net/i40e: add API to get received packet count
> [03/10] lib/librte_power: add extra msg type for policies
> [04/10] examples/vm_power_mgr: add vcpu to pcpu mapping
> [05/10] examples/vm_power_mgr: add scale to medium freq fn
> [06/10] examples/vm_power_mgr: add policy to channels
> [07/10] examples/vm_power_mgr: add port initialisation
> [08/10] examples/guest_cli: add send policy to host
> [09/10] examples/vm_power_mgr: set MAC address of VF
> [10/10] net/i40e: set register for no drop