The Competitors

We implemented prefetching algorithms that represent the FS, FA, and AS classes.

Within the FS and FA classes we pick members with small, medium and large $p$. In Figure 6, we observe that for the FA algorithms there is no optimal fixed value for $g$ that works for all workloads. We have chosen $g$ to be half of $p$ as that works best for the widest variety of workloads. For AS algorithms, we chose two popular variants, which adapt $p$ linearly (AS $_\textrm{Linear}$ ) and exponentially (AS $_\textrm{Exp}$ ). We also compare with OBL and the case with no prefetching.

Figure 6: On x-axis we vary the trigger distance ($g$) in an FA algorithm with $p=256$. On the y-axis we show the throughput when using a $120$ MB cache. When using five fast streams we get higher throughput with higher values of $g$, whereas, with a hundred slower streams, a smaller $g$ performs better.