Information Technology Reference
In-Depth Information
Fig. 5. Time to process 1 million data points vs the dataset size and the number of
iterations
Fig. 6. Number of cache PUTS (writes) and HITS (reads) per dataset size
using cache with local only partitions but, interestingly enough, as we increase
dataset size, this reuse drops, probably to due to RAM memory exhaustion on
each computing node. This behavior is a subject for further experimentation and
understanding.
Finally, it is worthwhile mentioning that our experiments produced an average
train accuracy of 86.77% of successful digit recognition (with 0.59 standard de-
viation), and an average 86.91% accuracy on the test data (with 0.76 standard
deviation). This figures fall within the expected accuracy of similar methods
reported for the MNIST dataset, including the method stability (very low stan-
dard deviation, under 1%) and generalization capabilities (very low difference
between train and test data), and guarantees the well behavior of the algorithms
throughout all experiments.
5 Conclusions
Caching strategies are key to enable scaling machine learning methods of iterative
nature. This work shows that even different strategies yield to different scalability
properties and, thus, caching arquitectures for distributed data must be taken into
account when devising scalable algorithms. Our results evidence that strategies
favoring cache reuse throughout the different iterations over the data outperform
simpler strategies, but this requires the algorithms (or the frameworks used) to
keep track and exploit data locality, combining different levels of caching (disk
and memory). This supports the convergence towards models where “computing
