How many options does ECS need?

Finally, given that EC scheduling cannot consider the full range of options, we investigate how many options such a scheduler requires to obtain good performance. To answer this question, we present a simple workload which repeatedly writes to the same track, and vary the number of target options it is given. Figure 8 presents the results.

Figure: The Diminishing Benefits of More Choice. The figure plots the performance of successive write requests to the same track. Along the x-axis, we increase the number of choices available for write targets, and the y-axis plots average write time.

0.9

In this experiment, we assume that if there exists only a single option, it is to the same block of the track; thus, successive writes to the same block incur a full rotational delay. As further options are made available to the scheduler, they are equally spaced around the track, maximizing their performance benefit.

From this figure, we can conclude that ECS does not necessarily need to consider all possible options within a range to achieve most of the performance benefit, as expected. By expanding a entire-track range to just eight choices that are properly spaced out across the track, most of the performance benefits can be achieved.

However, this example simplifies that problem quite a bit. For ranges that are larger than a single track, the expansion becomes more challenging; exactly how this should be done remains an open problem.