Elastic Translations MICRO'24 Artifact

This repository contains scripts and other supplementary material for the artifact evaluation of the paper Elastic Translations: Fast virtual memory with multiple translation sizes (DOI link TBD).

ET extend Linux to optimally and transparently take advantage of ARMv8-A's OS-assisted TLB coalesecing. ET transparently supports 64KiB and 32MiB translations, both at fault time and via asynchronous khugepaged migrations. The Leshy userspace profiler, based on ARMv8-A's Statistical Profiling Extension (SPE), samples the TLB misses of applications, either at runtime or offline, and provides translation size guidance to the kernel. We plan to add RISC-V Svnapot support to ET in the future.

Authors

• Stratos Psomadakis (National Technical University of Athens)
• Chloe Alverti (University of Illinois at Urbana-Champaign)
• Vasileios Karakostas (University of Athens)
• Christos Katsakioris (National Technical University of Athens)
• Dimitrios Siakavaras (National Technical University of Athens)
• Konstantinos Nikas (National Technical University of Athens)
• Georgios Goumas (National Technical University of Athens)
• Nectarios Koziris (National Technical University of Athens)

Paper Citation

TBD

Directory Structure

• bin/ Scripts and binaries required to run the experiments
• benchmarks/ Benchmark binaries and datasets
• env/ Bash scripts to configure the environment for the experiments
• hints/ Precomputed Leshy hints for the benchmarks, for both TLB miss sampling and accessbit sampling traces
• lib/ Libraries needed to run the experiments (gperftools tcmalloc, mpich)
• scripts/ Artifact bash scripts to prepare the host and build, install and run the artifact
• src/ Source code for the ET kernel, the ET userspace tools and utilities and the benchmarks.

Hardware Dependencies

ET requires a machine with ARMv8-A CPUs with support for the contig-bit in their TLBs (cf. ARMv8-A architecture reference manual D8.6.1). Leshy also requires support for the ARMv8.2-A Statistical Profiline Extension (SPE) (ARMv8-A architecture reference manual A2.14).

The benchmarks have a maximum memory footprint of 122GiB.

For the paper, we used a 2-socket Ampere Altra Mt.Jade server, with 80 Neoverse N1 (ARMv8.2+-A) CPUs and 256GiB memory in each socket.

We've also verified that ET run on NVIDIAs Grace (ARMv9 Neoverse V2).

Software Dependencies

For our evaluation, we used Ubuntu Jammy (22.04) for both native and virtualized execution. We list and install the required packages for building and running the artifact in scripts/prepare.sh.

For more information see the artifact appendix.