lyrm / testing_lf_ideas Goto Github PK

Trying to have a basic measure of false sharing impact in run time.

dune exec ./src/false_sharing.exe -- -d [NB_DOMAIN] -r [NB_RUN] -t [ID_TEST]

• [NB_DOMAIN] : number of domain
• [NB_RUN] : number of time the test is launched
• [ID_TEST] : between 1 to 4 (see below)

(median, average) run times in milliseconds.

• 1- with false sharing
• 2- just to see what happens: counters are int Atomic.t
• 3- trying to avoid false sharing with padding
• 4- trying to avoid false sharing with references

~20% in run time between test 1 (false sharing) and tests 3 and 4 (with 2 domains).

Playing with Domain.DLS to implement a counter and comparing performances with a basic Atomic.t counter.

The idea is that each domain has its local counter. Every time they accessed it, they try once to push their local value in the global counter. This way there are no CAS loops in the main function (incr, get). A flush function (a CAS loop) enables to get the final counter value.

dune exec ./src/counter.exe

There is a ~50% of improvement in run time with 4 domains with the counter using DLS versus an int Atomic.t counter.