On service guarantees of fair-queueing schedulers in real systems

Abstract In most systems, fair-queueing packet schedulers are the algorithms of choice for providing bandwidth and delay guarantees. These guarantees are computed assuming that the scheduler is directly attached to the transmit unit with no interposed buffering, and, for timestamp-based schedulers, that the exact number of bits transmitted is known when timestamps need to be updated. Unfortunately, both assumptions are unrealistic. In particular, real communication devices normally include FIFO queues (possibly very deep ones) between the scheduler and the transmit unit. And the presence of these queues does invalidate the proofs of the service guarantees of existing timestamp-based fair-queueing schedulers. In this paper we address these issues with the following two contributions. First, we show how to modify timestamp-based, worst-case optimal and quasi-optimal fair-queueing schedulers so as to comply with the presence of FIFO queues, and with uncertainty on the number of bits transmitted. Second, we provide analytical bounds of the actual guarantees provided, in these real-world conditions, both by modified timestamp-based fair-queueing schedulers and by basic round-robin schedulers. These results should help designers to make informed decisions and sound tradeoffs when building systems.