Resolution of Hardware Performance Counters

2022-03-24 | Tags: #research #perf #perf-counters #linux

in this post i will try to estimate resolution and correctness of performance counters in intel processors. i should remark that these test are conducted in a single threaded application. we shouldn't forget multi-threaded programs can led very different results.

perf event open

perf is a powerful tool to measure hardware performance counters in linux. perf also provides a syscall interface to interact with various counters like llc miss.

syscall(__NR_perf_event_open, ....)

creating llc miss

creating a last level cache miss is simple. first we should create a variable, after we should flush (with using clflush instruction) from cache and access it again. trying to access this variable causes a cache miss, and cpu needs to fetch from main-memory. we can simplify this process as follow:


to measure correctness of performance counters we will create n number of cache misses and compare with the expected results. results are shown like that :

sample count expected observed (mean) standart deviation correctness (%)
3000 1 1.0026 0.051579 99,9331551
3000 3 3.0001 0.143783 98,1333333
3000 10 10.0023 0.14945 98,1666667
3000 20 20.013 0.231334 96,4666667

This was the end of the blog post. You can reach me via email umusasadik at gmail com