Artifact description for PMBS'23 paper: Calculating Primes the Expensive Way: A Case Study in Write-Allocate Evasion on Intel Ice Lake SP

The article investigates a new hardware feature introduced with the Intel Ice Lake SP microarchitecture that automatically avoids write-allocate traffic for streaming write accesses. The analysis uses the CloverLeaf benchmark code from the SPEChpc2021 suite and micro-benchmarks to assess the effectiveness. The artifact description contains scripts to make CloverLeaf a stand-alone application without the SPEC benchmarking harness and patches for code changes applied to the CloverLeaf benchmark code for the in-depth analysis. Moreover, depending on the task, there are scripts that gather data for multiple parameters, commonly thread counts.

All data collecting scripts produce CSV files for further analysis or plotting. For reference, the CSV files with the results shown in the paper are part of the artifact description. Moreover, the artifact description contains the TeX sources to plot all figures presented in this work. They contain the transformations done to the data (scaling from GByte to Byte, etc.) before plotting.

With a working installation of the SPEChpc2021 suite, an Intel Ice Lake SP system and applied patches from this artifact description, all benchmarks can be re-run and reproduce the data presented in this work. It is important to note that the write-allocate avoidance feature behaves differently on different Ice Lake SP chips. For a full reproduction of the shown results, one has to use systems with Intel Xeon Platinum 8360Y as well. Furthermore, for disabling certain hardware features like the hardware prefetchers, SpecI2M or other undocumented prefetchers, one needs root access to the system.