This PR tracks the work on plugging the memory leak in both decision algorithms, discovered in #17. The current strategy is to rewrite the whole thing using mutexes (with a crate-internal wrapper).
I've also made a test (using getrusage) to check if a lot of calls leak more memory than they should. The bounds are correct on my laptop, I hope they'll be ok on Travis too.
This change improves single-threaded performance and makes multi-threaded performance a bit worse (which was tainted by the memory issue though, so who's to say).
Benchmarks before:
Running target/release/deps/multi_threaded-d8bd7867e9908eda
running 2 tests
test bench_gcra_20threads ... bench: 138 ns/iter (+/- 2,129)
test bench_leaky_bucket_20threads ... bench: 567 ns/iter (+/- 11,141)
test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured; 0 filtered out
Running target/release/deps/single_threaded-b5dced51ae82c67e
running 6 tests
test bench_allower ... bench: 22 ns/iter (+/- 11)
test bench_gcra ... bench: 145 ns/iter (+/- 64)
test bench_gcra_bulk ... bench: 249 ns/iter (+/- 539)
test bench_leaky_bucket ... bench: 164 ns/iter (+/- 47)
test bench_leaky_bucket_bulk ... bench: 178 ns/iter (+/- 100)
test bench_threadsafe_leaky_bucket ... bench: 200 ns/iter (+/- 134)
Benchmarks after:
Running target/release/deps/multi_threaded-bab1ac060974eb03
running 2 tests
test bench_gcra_20threads ... bench: 29,887 ns/iter (+/- 42,982)
test bench_leaky_bucket_20threads ... bench: 108,383 ns/iter (+/- 33,229)
test result: ok. 0 passed; 0 failed; 0 ignored; 2 measured; 0 filtered out
Running target/release/deps/single_threaded-c89da25a5fa0b4d2
running 6 tests
test bench_allower ... bench: 20 ns/iter (+/- 2)
test bench_gcra ... bench: 80 ns/iter (+/- 23)
test bench_gcra_bulk ... bench: 102 ns/iter (+/- 13)
test bench_leaky_bucket ... bench: 75 ns/iter (+/- 8)
test bench_leaky_bucket_bulk ... bench: 77 ns/iter (+/- 24)
test bench_threadsafe_leaky_bucket ... bench: 76 ns/iter (+/- 16)
test result: ok. 0 passed; 0 failed; 0 ignored; 6 measured; 0 filtered out
This PR tracks the work on plugging the memory leak in both decision algorithms, discovered in #17. The current strategy is to rewrite the whole thing using mutexes (with a crate-internal wrapper).
I've also made a test (using
getrusage) to check if a lot of calls leak more memory than they should. The bounds are correct on my laptop, I hope they'll be ok on Travis too.This change improves single-threaded performance and makes multi-threaded performance a bit worse (which was tainted by the memory issue though, so who's to say).
Benchmarks before:
Benchmarks after: