just a guess, 1. backplane (fabric oversubscription?) 2. resource contention when 2+ inbound if competing same outbound if other than interface ingress/egress qos, some products do fabric qos
My guess is ingress traffic can be put into queues of different priorities. The queue of absolute priority is of course taken care of faster than others, thus low latency.
but the problem is, if the interface/lien card is not oversubscribed, the low latency queue is actually not meanning too much, actually, in this situation, as long as we have enough queue depth, we even don't need to worry about the priorities
,
【在 L******t 的大作中提到】 : My guess is ingress traffic can be put into queues of different priorities. : The queue of absolute priority is of course taken care of faster than others, : thus low latency. : : 可 : 会
if the interface is not oversubscribed, in most cases(if no line card oversubscription), the queue depth would be always 0. we even don't need a very big buffer. the situation you mentioned is exactly what I am concerned. the low latency queue on port B would not mean much.
interfaces has
【在 L******t 的大作中提到】 : QoS is implemented by priority queues. That's why the earlier we classify : traffic into different queues, the better. : : 的
L*t
16 楼
My guess is ingress traffic can be put into queues of different priorities. The queue of absolute priority is of course taken care of faster than others, thus low latency.
可 会
【在 z**r 的大作中提到】 : if the interface is not oversubscribed, in most cases(if no line card : oversubscription), the queue depth would be always 0. we even don't need a : very big buffer. : the situation you mentioned is exactly what I am concerned. the low latency : queue on port B would not mean much. : : interfaces : has
z*r
17 楼
嗯,你这个说的是另外一个种可能,尤其是ingress line card oversubscription的情况
A 下 traffic 让 者
【在 L******t 的大作中提到】 : My guess is ingress traffic can be put into queues of different priorities. : The queue of absolute priority is of course taken care of faster than others, : thus low latency. : : 可 : 会
z*r
18 楼
第一个不应该是,如果到了fabric都oversubscribed了的话,任何traffic都没有办法保 证了。除非是一个特殊的traffic pattern。 can you explain more?
【在 c*a 的大作中提到】 : just a guess, : 1. backplane (fabric oversubscription?) : 2. resource contention when 2+ inbound if competing same outbound if : other than interface ingress/egress qos, some products do fabric qos : : 可 : 会
L*t
19 楼
我上上篇帖子没说清楚。假设一个controller/arbiter控制两个PHY port A & B;port A 满载,但全是垃圾traffic,port B undersubscribed,但是VOIP traffic。这种情况下 ,如果两个PHY port分别只有一个queue的话,arbiter作round robin,port B traffic will suffer even though it's undersubscribed. Port A的垃圾traffic会一路上去让 别人suffer直到classification point.不用说更经常的情况是一个arbiter控制4个或者 更多PHY ports. 所以说如果单个port不oversubscribed,不影响multiple queueing的必要性.如果都不 oversubscribed的话,共产主义就来到了.