Initial Results

We ran the Andrew benchmark under kernels modified to prefetch based on PCM and last successor modeling. Figure 6 shows the elapsed time and read latency reductions for several tests. From these tests we saw reductions of up to 26% in total I/O latency and 15% in read latency. The simple last successor based prefetching did better than some settings of the more complex PCM based prefetching. PCM based prefetching improved as the partition size increases from 16 to 32, but the increase to 64 offered no further improvements.

However, the compute times for our benchmark tests increased 0.05 seconds, apparently due to modeling and prefetching overhead. Compute time for the last successor test increased by as much as, and in some cases more than, those for PCM based prefetching, even though last successor is a much simpler model. This indicates that the prefetching engine is most likely the dominant factor in the increased computational overhead. Latencies for both open and exec events also increased. Despite these increases, predictive prefetching reduced both the total I/O latency and read latency.

  

Figure: Reductions in elapsed times and read latencies for the Andrew benchmark with the last successor, PCM and hot cache tests. Bars marked with P represent PCM tests. Partition sizes (ps) and model order (mo) are labeled as ps/mo.