DimmWitted: A Study of Main-Memory Statistical Analytics
This provides insights for designers of analytics engines on optimizing hardware and statistical efficiency, though it is incremental as it focuses on tradeoff analysis rather than a new method.
The study explores tradeoffs in data access methods and replication for main-memory statistical analytics on NUMA machines, finding that their prototype engine can run at least one popular task at least 100x faster for each system considered.
We perform the first study of the tradeoff space of access methods and replication to support statistical analytics using first-order methods executed in the main memory of a Non-Uniform Memory Access (NUMA) machine. Statistical analytics systems differ from conventional SQL-analytics in the amount and types of memory incoherence they can tolerate. Our goal is to understand tradeoffs in accessing the data in row- or column-order and at what granularity one should share the model and data for a statistical task. We study this new tradeoff space, and discover there are tradeoffs between hardware and statistical efficiency. We argue that our tradeoff study may provide valuable information for designers of analytics engines: for each system we consider, our prototype engine can run at least one popular task at least 100x faster. We conduct our study across five architectures using popular models including SVMs, logistic regression, Gibbs sampling, and neural networks.