An Investigation of Integration and Performance Issues Related to the Use of Extended Page Sizes in Computationally Intensive Applications

The combination of increasing fabrication density and corresponding decrease in price has resulted in the ability of commodity platforms to support large memory capacities. Processor designers have introduced support for extended hardware page sizes to assist operating systems with efficiently scaling to these memory capacities. This paper will explore integration strategies the designers of the Linux operating system have used to access this hardware support and the practical performance impact of using this support. This paper also provides a review of common strategies for adding support for this functionality at the application level. These strategies are applied to a sampling representative of common scientific applications to support a practical evaluation of the expected performance impact of extended page size support. An analysis of these results support a finding that a 5% performance improvement can be expected by adding support for extended page sizes to memory intensive scientific applications