Multiple Sequential Streams: varying number of streams

Figure 9: We show the average achieved throughput over a period of two minutes as a function of the number of concurrent sequential streams reading from five SCSI disks with thinktime = $30$ ms, readsize = $8$ KB, and a $100$ MB cache.

In Figure 9, we study the aggregate streaming throughput of various prefetching algorithms when we increase the number of concurrent sequential streams while keeping the cache size constant. We observe that most algorithms saturate at some throughput beyond which increasing the number of streams does not improve the aggregate throughput. Algorithms that issue fewer but larger disk reads and at the same time waste little generally do better. We observe that no prefetching and OBL have the lowest throughput as they have a large number of small read requests. Somewhat better are the FA $_\textrm{8/3}$ and FS $_\textrm{8}$ algorithms as they create fewer read requests than OBL. Interestingly, FA $_\textrm{256/127}$ and FS $_\textrm{256}$ algorithms also have similarly low performance in spite of large prefetch degree. This is because the large prefetch degree leads to significant prefetch wastage. The FS $_\textrm{64}$ and FA $_\textrm{64/31}$ perform the best in their respective classes as they strike a balance and have large reads but do not waste as much. The AS $_\textrm{Linear}$ and AS $_\textrm{Exp}$ are generally good performers because they adapt the degree the prefetch. However, since these algorithms lack the ability to detect and avoid wastage, the more aggressive AS $_\textrm{Exp}$ fares worse than its linear counterpart. AMP being an asynchronous adaptive algorithm discovers the right prefetch degree for each stream thus avoiding wastage and achieving the best possible performance.

At the maximum number of streams, AMP outperforms the FA algorithms by $41$-$350$%, the AS algorithms by $18$-$60$%, the FS algorithms from nearly equal to $252$% and no prefetching and OBL by a factor of $12$.