Re: Request for help, implementing a new network scheduler

[Zoltan Kiss <zoltan.kiss@linaro.org>]

Hi,

On 03/02/15 19:55, ronald pina wrote:
> Thanks Zoli for your clear and very helpfull answer, my idea is to
> prioritize traffic that comes from different guest, for example if
> guest_1 is used for a voip server than it would be reasonable to
> schedule first the vif guest_1 before other guests. The diagram atached
> below explains better my idea. The goal is to improve the latency and
> jitter for real-time traffic which have critical requirements to QoS
> especially on heavy network i/o hosts, and if the vifs of different
> guest are scheduled on round robin manner in dom0 backend than we can
> improve this scheduling algorithm like using weighted round robin which
> can give more scheduling time to a vif that was used by a  voip-server
> guest. I would like to know that which module or function is responsible
> for scheduling vif on dom0 ?

As I said, it is a NAPI polling function:

http://www.linuxfoundation.org/collaborate/workgroups/networking/napi

So it is scheduled by NAPI, which doesn't have a clue about what kind of 
packets the device dealing with, therefore can't make a decision based 
on that. You can try manually configure the weight of a device through 
netif_napi_add(), but I strongly recommend to look for other ways I 
mentioned in my previous letter.

>
> Thanks
> Ronald
>
> On Tue, Feb 3, 2015 at 12:54 PM, Zoltan Kiss <zoltan.kiss@linaro.org
> <mailto:zoltan.kiss@linaro.org>> wrote:
>
>
>
>     On 03/02/15 00:42, Ronald PIna wrote:
>
>         Hi
>
>         I am working for the msc thesis to improve the performance on
>         network
>         for guest domains that uses real-time services like voip or video
>         streaming servers , i have an idea to implement a network
>         scheduler on
>         network backend, the schedulers may be weighted fair queuing or
>         weighted
>         round robin,  the idea is to schedule first the packet coming from
>         real-time guest services, one of those schedulers could make the
>         job and
>         can prioritize the network traffic.
>
>         As far as i have studied from previous works explain that the
>         outgoing
>         network traffic in xen is scheduled in round robin manner inside the
>         function net_tx_action(). Later on the last version of xen this
>         function
>         has changed xenvif_tx_action() and have changed his primarily
>         structure.
>         My primarily goal is to modify the round robin to more advanced
>         scheduler  which introduce priorities. The clear question is
>         that where
>         is located the function which schedule in round robin and if
>         possible
>         are there any concern about using another network scheduling
>         method in
>         netback ?
>
>     xenvif_tx_action is called from NAPI polling context, per queue. It
>     copies & maps the packets from the guest to Dom0 (at which point it
>     doesn't have the slightest idea about the packet content), formats
>     them in xenvif_tx_submit to be a well formed skb, and calls
>     netif_receive_skb to hand it over to the stack.
>     You could try doing some reordering in xenvif_tx_submit or
>     introducing priorities in the ring buffer, but I would strongly
>     discourage you to do that, for similar reasons mentioned by Ian. You
>     would be better off using standard Linux tools for doing that, see
>     'man tc', and I would try on the sending side, in the guest, before
>     netfront gets the packet.
>     Or, if your idea is to prioritize between traffic coming from
>     separate guests, then xenvif_tx_action is definitely the wrong place
>     to do that: as said above, it deals only with the traffic of one
>     particular queue of a netback device. You might want to look into
>     the entity switching/routing between the guests and NICs: Linux
>     bridge module, openvswitch, the IP routing engine of the backend
>     etc. I think the first two could be the reasonable choices, bridge
>     might lacks such functionality though.
>
>     Zoli
>
>
>
>
>         Thanks in advance
>
>         _________________________________________________
>         Xen-devel mailing list
>         Xen-devel@lists.xen.org <mailto:Xen-devel@lists.xen.org>
>         http://lists.xen.org/xen-devel
>
>