On Thu, Apr 01, 2021 at 07:53:25PM +0800, Tonghao Zhang wrote:
>
> Hi Ben, Ilya
> Try to explain this patch again. Now OvS has supported the burst_size,
> as one user case,
> if users don't use the burst_size feature, we should set burst_size to
> rate or 0. This patch set this to 0.
'0' is definitely not an "implementation defined optimal
value" as the spec requires. Actually, most token buckets
implementations do not work or work very poorly with a bucket size of
zero.
The first hit on a Google search :
https://www.juniper.net/documentation/us/en/software/junos/routing-policy/topics/concept/policer-mx-m120-m320-burstsize-determining.html
I don't fully agree with their recommendations, 5ms is way too
large for high speed networks, but they expose the problem properly.
> As Ilya said, we should check the OFPMF13_BURST in userspace datapath,
> I think it's right.
I don't have enough context to comment on the patch and
I don't know where that flag should be tested. It would seem that all
datapaths will suffer from the same issue, you need to figure out an
optimal burst_size for user space *and* for kernel space if none is
given via OpenFlow.
I have a strong suspicion that the patch does not do a great
job if OFPMF13_BURST is not set, but I may be wrong.
Regards,
Jean
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
