1. Networking Issues in LAN Telephony Brian Yang 5-3-00

2. Topics Some background Flow Based QoS Class Based QoS and popular algorithms –Strict Priority (SP) –Round-Robin (RR), Weighted Round Robin (WRR) and Weighted Fair Queuing (WFQ) –Random Early Drop (RED) and Weighted RED (WRED) –Earlier Due Date and Weighted Fair Discard Research Plan

3. Convergence LAN used to transport one type of traffic: bulk data Voice was transported on a separate network Merging the 2 networks promises huge cost savings. An integrated network might carry –interactive multimedia, e.g. phone calls –multimedia, e.g. training video –interactive data, e.g. Web business, mission critical data –traditional bulk data, e.g. ftp –best effort, e.g. private web browsing These traffic types have diverse delay/bandwidth requirements

4. Flow Based QoS A flow is a sequence of packets belonging to a session, e.g. a phone call Advantage: capable of precise guarantee to each flow Disadvantage: –requires per flow classification/scheduling by network –difficult to implement in hardware switches Current proposal not for LAN

5. Class Based QoS Flows with same QoS requirement are aggregated into a class. Each packet carries its class label in its header Class labels can be given by –the operating system, e.g. Win 2000 –given by a classifier at the edge of the network The network deals only with classes. The network provides a pre-negotiated per-hop behavior (bandwidth allocation and/or latency) to each class. Reduces implementation complexity

6. Example –Class 1: time sensitive traffic, e.g. phone calls –Class 2: interactive data traffic, e.g. web business, financial apps –Class 3: bulk data, e.g. email, ftp –Class 4: best effort, e.g. private web traffic From a 10Mbps WAN connection,

7. Examples of QoS-aware LAN Switches –Source: data.com 05/99

8. Class Based Queuing At each output port, packets of the same class are queued at distinct queues. Non-class based queuing: Class 1 Class 2 Class 3 Class 4 Class based scheduling First come first serve

9. Strict Priority Lower priority packets are served only if all higher priority queues are empty Provides low latency for the highest class Problems: –No bandwidth partition. A higher class can take over the entire link, at the expense of lower priority packets. Not fair to lower classes. –No QoS guarantee for lower priority traffic, unless higher priority classes are strictly controlled by some other method.

10. Round Robin and its Variants All these algorithms aim at bandwidth partition. Round Robin (RR): each queue is served in a round robin fashion, one packet at a time –In our example, each class is guaranteed 25% of the link bandwidth. Weighted Round Robin (WRR): the bandwidth partition is adjustable. Weighted Fair Queuing (WFQ) –A refinement of WRR, accounting for variable packet size

11. Random Early Discard Dropping policies Traditional: Drop when the queue is full (drop tail) RED Discard probability Queue length 100% Discard probability Average queue length 100% min_thres max_thres

12. Weighted RED Different dropping thresholds for different classes RED was designed to avoid global synchronization of TCP connections, thus achieve high link utilization. Average queue length 100%

13. Problem with strict priority: no bandwidth guarantee Problem with round robin: lack of delay guarantee Problem with RED: not sure if suitable for voice Originally developed for WANs, are these desirable for integrated services LANs? Propose an alternative that –attempts to transmit a packet before its delay bound –achieves bandwidth partition

14. Earliest Due Date Work done while I was with Vertex Networks Due Date(DD) = arrival_time + delay target All packets are timestamped with its arrival time Scheduling: –examine head-of-line packets of class based queues –serve the packet with earliest (smallest) due date

15. Class 1, voice, delay target = 5ms 3536 Current time = 38 ms 37 1824 11 352 Class 2, interactive data delay target = 20ms Class 3, bulk data delay target = 40ms Class 4, best effort delay target = infinity DD = 39 DD = 38 DD = 51 34 DD = inf Recall our earlier example

16. Congestion Plane If the queue length triple (Q1, Q2, Q3) is inside the congestion plane, arrivals are not discarded. If the queue length triple is outside the congestion plane, classes that occupy disproportionate amount of space are identified, and their new arrivals are discarded. Q1 Q2 Q3 50kbits 200kbits 400kbits

17. Weighted Fair Discard Upon a packet arrival –if (Q1, Q2, Q3) are inside congestion plane, don’t drop –else if arrival is class 1 and Q1>25kbits, drop if arrival is class 2 and Q2>62.5Kbits, drop if arrival is class 3 and Q3>50Kbits, drop; else don’t drop The dropped packets would violate its delay target even if they were admitted.

18. Research Plan Investigate the current QoS algorithms –Look into their performance on a mix of voice and data traffic –Quantify the behavior of EDD and Weighted Fair Discard Develop solution for legacy LAN equipment –Hubs accounts for almost half of all Ethernet ports sold last year –Majority of equipment are QoS-unaware in existing networks Extend our results to the LAN/WAN interface –LAN/WAN interfaces are often bottlenecks in intranets Integrate our solution with WAN QoS standards –RSVP, Diffserv and their interaction with LAN QoS algorithms