Scalability of the SAS/STAT HPGENSELECT High-Performance Analytical Procedure: A comparison with RevoScaleR

Effectively implementing high-performance analytics software solutions in the insurance industry Executive Summary At the Strata Conference on October 25, 2012, the research and planning division of a large insurance corporation (hereafter " insurer ") presented various methods that they used to model 150 million observations of insurance data. A summary of their presentation is available at:  The GENMOD procedure in SAS/STAT ® software  Custom MapReduce code on a Hadoop cluster  Open-source R  Revolution R Enterprise using RevoScaleR The insurer reported that PROC GENMOD took more than five hours to fit a Poisson regression model, whereas Revolution Analytics used the RevoScaleR package to fit the model in 5.7 minutes. However, this is not an " apples to apples " comparison because RevoScaleR was run on a cluster of computers, whereas the GENMOD procedure executes only on a single server. A more informative comparison can be made by using the HPGENSELECT procedure, which had not yet been released at the time of the Strata comparison. Introduced in SAS/STAT 12.3 in June 2013, PROC HPGENSELECT runs in either single-machine mode (multiple threads on a single machine) or distributed mode (multiple threads on multiple machines). Distributed mode requires SAS ® High-Performance Statistics. Purpose This paper compares the performance of the HPGENSELECT procedure with results cited for the RevoScaleR package by using data that are similar to the insurer's data. The paper also demonstrates the scalability of the HPGENSELECT procedure by using two sizes of data sets and three different computing environments. Results On a small grid with two nodes, the HPGENSELECT procedure fits a Poisson regression model with 150 million observations in 159 seconds, which is less than half the time that RevoScaleR required on a somewhat larger grid. On a grid with 140 nodes, the HPGENSELECT procedure solves the problem in 22 seconds. The scalability of the HPGENSELECT procedure is demonstrated by increasing the size of the data set. For a data set that has the same variables and one billion observations, the procedure executes in less than one minute. These results, which are summarized graphically in Figure 1, show that the HPGENSELECT procedure provides a faster alternative.