A note on SCSI bus waits

Abstract

In the SCSI-2 standard, the unique IDs of devices on the bus define a fixed priority whenever several devices compete for the use of the bus. Although the more recent SCSI-3 standard specifies an additional fair arbitration mode, it leaves such fair mode an optional feature. Despite a number of allusions to potential unfairness of the traditional SCSI bus arbitration scattered in the trade literature, there seem to be few formal studies to quantify this unfairness.In this paper, we propose a simple model of SCSI bus acquisition in which devices on the bus are viewed as sources of requests with fixed non-preemptive priorities. We use the model to assess the expected extent of unfairness, as measured by the mean bus wait, under varying load conditions. Effects of tagged command queueing are not considered in this note. Numerical results obtained with our model show that there is little unfairness as long as the workload is balanced across devices and the bus utilization is relatively low. Interestingly, even for medium bus utilization a significant fraction of bus requests find the bus free which might correlate with the service rounds noted in a recent experimental study. For unbalanced loads and higher bus utilization, the expected wait for the bus experienced by lowest priority devices can become significantly larger than the one experienced by highest priority device. This appears to be especially true if the higher priority devices have higher I/O rates and occupy the bus for longer periods. As might be expected, even for balanced workloads, unfairness tends to increase with the number of devices on the bus.