Open voximity opened 8 months ago
The fact that these days std
locks are faster is true indeed; I'm assuming this (performance degradation) hasn't been observed yet? FWIW, no lock contention has popped up in perf
profiles I've run before, but perhaps I wasn't testing in extreme enough conditions. Also, I tested std
locks against parking_lot
ones in a stress test for an unrelated network app in the past, but wasn't able to detect a difference in performance (though perhaps that app had fewer lock-related operations involved).
how can we monitor this @ljedrz ?
with periodic (but not necessarily long) perf
runs; high lock contention should be apparent there
We have been considering the possibility of mutex contention causing performance degradation in snarkOS and snarkVM.
The two ecosystems make use of parking_lot for its
Mutex
es andRwLock
s. It is known thatparking_lot
suffers from performance degradations when under heavy contention. We could confirm this when running the benchmark on a machine that runs a validator on the Canarynet.Steps to Reproduce
N/A. We are looking into analyzing the prevalence of contention, as well as experimenting with replacing the usage of
parking_lot
withstd::sync::Mutex
.Expected Behavior
N/A.
Your Environment
N/A. We tested the benchmark above in Ubuntu on WSL and Ubuntu on bare metal.