FT investigation reveals popular scientific benchmarks do not accurately represent TLB behaviors of real applications

FT group members Collin McCurdy and Jeff Vetter, along with Alan Cox of Rice University, will be presenting their paper "Investigating the TLB Behavior of High-end Scientific Applications on Commodity Microprocessors," in Austin, TX this April at the 2008 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS’08). The paper is the culmination of work, undertaken as part of the PetaSSI FastOS project, seeking to understand the TLB behavior of scientific applications. The analysis shows that two benchmark suites that are understood to represent scientific application behavior (SPEC CPU and HPCC) are not representative of the TLB behavior of important full scale applications. Furthermore, the paper demonstrates that false conclusions drawn from benchmark TLB performance can have significant ramifications for application performance.