[ovs-dev] [PATCH ovs resend v2] dpctl-netdev: Add the option "pmd" for dump-flows
xiangxia.m.yue at gmail.com
Tue Sep 8 01:39:16 UTC 2020
On Mon, Sep 7, 2020 at 9:01 PM Ilya Maximets <i.maximets at ovn.org> wrote:
> On 9/7/20 7:29 AM, Tonghao Zhang wrote:
> > On Thu, Sep 3, 2020 at 8:51 PM Ilya Maximets <i.maximets at ovn.org> wrote:
> >> On 8/24/20 4:44 AM, xiangxia.m.yue at gmail.com wrote:
> >>> From: Tonghao Zhang <xiangxia.m.yue at gmail.com>
> >>> "ovs-appctl dpctl/dump-flows" added the option
> >>> "pmd" which allow user to dump pmd specified.
> >>> That option is useful to dump rules of pmd
> >>> when we have a large number of rules in dp.
> >> Is this a performance concern or just a matter of output readability?
> > Yes, performance and the matter of output are considered. Think about
> > there is a lots of flows on every
> > pmd which may be offloaded.
> I understand that it will be faster to not dump flows that wasn't requested,
> however, dpctl interface is a debugging interface and should not be used
> frequently in normal conditions. And if you have so many flows that dumping
> takes significant amount of time, you might have much more important issues
> with inability to change them in time by revalidators.
Yes, you are right. In my branch, I change the logic of revalidator
which will not cleanup the flows which
installed by dpctl. and disable the upcall(via upcall-limit). In that
way, all flows are installed to HW, and packets
will be processed in the hardware.
> So, why the performance of this operation is such a concern for you?
If all flows are installed to hardware and there are multi-pmds(for
example 4), I don't want to dump flows from every pmd
even though they have same flows.
By the way, if we don't use the dpctl to install flow, To support ct
flow offload, there are a lots of flows in hardware ?
for example there are too many TCP connections.
> >> In case it's only to reduce the size of a printed data, we could just
> >> avoid printing flows from PMDs other than requested the same way as
> >> other filters work without flow_dump API modifications. This will be
> >> much more simple change.
> >> What do you think?
> >> Best regards, Ilya Maximets.
Best regards, Tonghao
More information about the dev