and a prediction that does not take interference into account.
The prediction closely matches the actual execution time, but
shows minor anomalies for the 2, 7 and 8 slots cases; these
are caused by a small change getting exaggerated by the
rounding. Most importantly, the prediction correctly captures
the actual trend allowing us to predict how this workload
will scale when per-node parallelism is changed for the map
stage. When interference is not considered the prediction is
far too optimistic, indicating the importance of considering
the effects of I/O interference in the model.
The reduce stage results are shown in Figure 4(b). The
interference prediction over-estimates the execution time to a
larger degree than for map tasks, particularly for three slots,
which is primarily due to caching; some of the intermediate
data is still cached when it is read by the shuffle or reduce
phases. When the number of slots increases, the amount of
data per wave becomes big enough for caching to be less
effective.