Information Technology Reference
In-Depth Information
4Rsus
Metrics collected in experiments focused on understanding (1) how different
cache configurations affected the total elapsed time of the experiments and (2)
how the processing time per data item (or 1 million data items) evolved through-
out the different run configurations. Elapsed time is understood as the wall time
of each experiment from beginning to end. As each experiment was run twice,
results show averages of the two runs. Deviations in times between the two runs
were insignificant.
Figure 4 (left) shows the experimental elapsed time as a function of the number
of data items processed. Here we averaged runs processing the same number of
data items, regardless the number of iterations and dataset size. For instance, a
dataset with 30K data items through 40 iterations processes a total of 1.2M data
items, whereas a dataset with 60K data items through 25 iterations processes a
total of 1.5M data items.
As it can be seen any cache strategy largely outperforms the lack of cache
and, furthermore, caching strategies using only local data tend to perform even
better that caching strategies that use no criteria to select the data to work on.
Figure 5 shows the average time to process 1 million data items as a function
of the number of iterations (left) and the dataset size (right). In all cases, as the
amount of data processed grows (either through iterations or through dataset
size) the time to processes 1 million data items tends to decrease, probably as
the first data items loads in other low level caches (processor, OS, Java virtual
machine, etc.). Anyhow, again we see that exploiting local data outperforms all
other strategies.
When coming to compare the two caching strategies used we can observe in
Figure 6, as expected, that using only local data results in a reduced number of
writes (PUTs) to the cache whereas the number of reads is similar. This signals
data reuse within each cache and we interpret this as the root cause of the
observed improvement of caching strategies using data locality information.
Figure 4 (right) shows the number of cache HITs per PUT with respect to
dataset size where one can observe some indication of the degree of data reuse in
each configuration. In general each data item is re-used about 5 times more often
Fig. 4. Total elapsed time of experiments (left) and data reuse (right)
