Multiple Queue Priority Routing

Brad Hedlund fabrics built from fixed configuration pizza box switches with 'big end' chassis switches. Comments made by various different readers were all over place but surprisingly nobody addressed the queuing problems. Lowercost devices in general use unsophisticated internal queuing mechanisms. You should take it into account. Assuming there's no QoS configured on output forwarding, port and even queuing hardware works along this kind of lines.

While resulting in substantially latency for all other traffic streams, not surprisingly, a really great traffic stream going toward an output port saturates output port queue. Each queue works as a FIFO queue -once hardware finds out which queue to use for a specific packet, packet is stuck in that queue.

Order in which output interface hardware serves the queues determines service actual quality. Packets from a priority queue will be sent 1-st, and hardware probably support multiple priority levels.

Straightforward round robin algorithms use 'perqueue' byte count quotas. Remember, the hardware sends at least ten. This kind of algorithms are probably obviously not precise, as they usually send a bit more than the queue's quota worth of facts. This deficiency is fixed in the Weighted Deficit Round Robin algorithms that reduce the per queue byte next count round robin cycle by excess amount traffic sent in current cycle.

As well, imagine a scenario where a great file transfer lands in same path across internal switching fabric as a requestresponse protocol handling shorter transactions. Then once more, it generates continuous stream of facts that fills all output queues in the path, once the file transfer gets going. Every time the transactional protocol sends some record, it encounters massive queues at every hop, notably increasing 'end to end' latency and deteriorating the response time.

Cisco solved this queuingonoutputinterface partition trouble with Weighted Fair Queuing, an intriguing solution that uses a separate FIFO output queue for every flow. This hardware implementation solution is probably rather over-priced and is rarely attainable in switching silicon. Now look. Lofty end' switches solve at least some 'head of straight line' blocking scenarios with virtual output queues. With that said, the hardware implements 'perclass' virtual output queue on input ports, while not having a single 'perclass' queue on an output port.

Loads of information can be found on the internet. packets stay in virtual output queue on input linecard till the output port was usually prepared to accept another packet, at which time the hardware requires a packet from amongst the virtual output queues, in general in roundrobin fitness, the packet forwarding and queuing mechanisms work as before. Cannot solve HoL blocking troubles betwixt same flows traffic class entering the switch thru the same input port, virtual output queues solve threshold head blocking between input ports.

Lofty bandwidth' chassis switches mostly use multistage forwarding thing. Doesn't it sound familiar? Transport across internal fabric may cause more delays. Even if switch uses virtual output queues, a jumbo frame transferred across the fabric delays shorter transaction requests traversing the same fabric lane, in the end.

Now look. Cellbased fabrics solve this poser when slicing the packets in smaller cells. Want to understand what hardware and application features data center switching vendors introduced in last 12 months? Register for data Center Fabrics Update webinar.