This repo contains the code used to produce the results reported in our ISMM 2021 paper "Understanding and Utilizing Hardware Transactional Memory Capacity". You can watch the talk on YouTube.
We extended the original work by Carl Ritson.
Simply run make.
To reproduce our results reported in the paper, run ./run_ismm21.py Baseline Reuse Invalidation Warmup WarmupIter.
To enable huge page support, run ./run_ismm21.py HugeTLB Baseline Reuse Invalidation Warmup WarmupIter.
You can select which tests to run by only passing certain configuration names to run_ismm21.py.
Add hugepagesz=1G hugepages=<n> to your kernel's command-line parameters, where <n> is a number that is big enough to meet the benchmarks allocation requirement.
The amount of memory to be allocated is printed at the beginning of the benchmark.
Install kernel module wbinvd and make sure that /proc/wbinvd is readable by the user running the benchmark.
The some of the parameters are machine dependent. For example, the warmup iteration used for our Haswell machine is 128.
The run_ismm21.py will choose the right parameters for each machine by reading the CPU model names from /proc/cpuinfo.
If you use a different CPU than what we evaluated, you are very welcome to use this tool.
You will first need to run experiments to figure out the correct parameters for your particular machine, and then change the run_ismm21.py script accordingly.
For technical questions directly related to the code, please contact Zixian Cai directly.