In search of a fair scheduling policy for serving multi-queue systems

Abstract

A Weighted Round Robin Scheduling (WRRS) scheme was proposed in [SI to serve a system of TWO queues by a single server. The weights --which determine the time duration for which a non-empty queue is served --are determined on the basis of monitoring the traffic mix and characteristics at the entry point of the node. This note proposes incorporation (optional) of one more factor in the process of determination of the weights. This being the setting up of an adaptive threshold in each of the buffers in order to determine the number of cellslpackets which are served during every visit (by the server) to each of the queue. The technique can be easily extended to multi-queue systems. .O The Original TechniqueThe Weighted Round Robin Scheduler (WRRS) can be implemented using TWO different techniques. The first one is as follows. The High Priority Buffer (HPB) is served in a gated exhaustive fashion, i.e., all the cells which are there in the HPB at the instant of the server's (i.e., the link controller's) arrival are served. Arrival can still continue to occur in the HPB, but these will be served in the next service epoch. The Low Priority Buffer (LPB) is served in a limited fashion if the HPB is non-empty at the instant when the server switches to the LPB, otherwise it is served in an exhaustive fashion, i.e., all the cells which are there at the instant of the server's arrival to the LPB are served. The second one calls for the use of a table. The entries in the table are updated on the basis of scanning the activities of the HPB and LPB, and monitoring traffic at the input of the bucket bank. These entries are periodically updated, where the period is the time horizon over which the control is exercised. The Proposed Option and Its EvaluationThe proposed modification is as follows. When the server arrives at the queue, e.g., the HPB or LPB (Figure l), it serves a maximum of KH where K , I H1 (or KL where KL I L1 ) cells/packets only if the number of cells/packets in the queue exceeds the adaptive threshold H1 for HPB (or L1 for LPB), otherwise the cells/packets have to wait till the server scans that queue in the next cycle. Note that the length of the cycle is also adaptive, since both H1 and L1 vary in every cycle. The thresholds H1 and L1 are highly adaptive, H2 and L2 are less adaptive, and H3 and L3 are least adaptive. The second and third thresholds (i.e., H2 and H3 in HPB, and L2 and L3in LPB) are for triggering congestion notifications.Two possible actions can be executed when the server (i.e., the Link Controller) scans a queue. Action-1 (al) is performed when the backlog (i.e., the number of cells in the queue) equals or exceeds a pre-specified threshold, and hence this queue is allowed to transmit. Action-2 (a2) is executed otherwise, and hence the queue is not served during this scanning cycle. This is shown in Figure 2.Note that the arrival rates to the queues may be affected by the type of action (i.e., a1 or a2) executed during one scanning cycle. This may not be the case when the size of the buffer s...