Messages in this thread | | | Subject | Re: [PATCH net] openvswitch: meter: remove rate from the bucket size calculation | From | Ilya Maximets <> | Date | Thu, 29 Apr 2021 23:15:06 +0200 |
| |
On 4/28/21 8:59 PM, Ilya Maximets wrote: > On 4/28/21 6:31 PM, Jean Tourrilhes wrote: >> On Wed, Apr 28, 2021 at 01:22:12PM +0200, Ilya Maximets wrote: >>> >>> I didn't test it, but I looked at the implementation in >>> net/sched/act_police.c and net/sched/sch_tbf.c, and they should work >>> in a same way as this patch, i.e. it's a classic token bucket where >>> burst is a burst and nothing else. >> >> Actually, act_police.c and sch_tbf.c will behave completely >> differently, even if they are both based on the token bucket >> algorithm. >> The reason is that sch_tbf.c is applied to a queue, and the >> queue will smooth out traffic and avoid drops. The token bucket is >> used to dequeue the queue, this is sometime called leaky bucket. I've >> personally used sch_tbf.c with burst size barely bigger than the MTU, >> and it works fine. > > Makes sense. Thanks for the clarification! > >> This is why I was suggesting to compare to act_police.c, which >> does not have a queue to smooth out traffic and can only drop >> packets. > > I see. Unfortunately, due to the fact that act_police.c uses time > instead of bytes as a measure for tokens, we will still see a difference > in behavior. Probably, not so big, but it will be there and it will > depend on a line rate.
I found my mistake in calculations (see another reply). So, there should not be significant difference with OVS meters, indeed.
> >> I believe OVS meters are similar to policers, so that's why >> they are suprising for people used to queues such as TBF and HTB. >> >> Regards, >> >> Jean >>
| |