Peppermint and Sled: tools for evaluating SMP systems based on IA-64 (IPF) processors

In this paper, we describe Peppermint and Sled: tools developed for evaluations of computer systems based on IA-64 processors. Sled generates trace from applications running on IA-64 processors, while Peppermint models the complete system using cycle-accurate, trace-driven simulation. Peppermint is based on Augmint, which leaves open the possibility of doing execution-driven simulations in future. Peppermint and Sled allow us to perform a trace-based evaluation of 4 applications running on SMP systems based on Itanium and McKinley processors. We find that the improvement in IPC of McKinley relative to Itanium ranges from 7% to over 100% for our different applications. The improvement can be attributed to a variety of factors. These range from the availability of additional functional units and issue ports in the McKinley processor to our assumption of a better memory system. While the improvement in performance remains valid in SMP systems in some cases, higher contention for system bus and memory reduces the performance gain in other cases. Increasing the system bits bandwidth and size of queues for pending requests in the memory controller are identified as first steps for optimizing SMP performance.