Use atomic types to speedup variable difficulty alogirthm #3

Merged

realaravinth merged 9 commits from optimize-libmcaptha into master

2023-12-31 01:31:22 +05:30

realaravinth commented

2023-12-30 14:44:12 +05:30

Owner

The actor model currently in use in libmCaptcha is causing a bottleneck that was previously not visible. Protobuf implementation from #1 was yielding very low throughput.

Generating [a flamegraph of that version showed that the AddVisitor message handler was taking up ~59% of the CPU time while the node was hit with the locust benchmark. Keeping track of visitor count is a critical path in the variable difficulty algorithm, but this particular test also involved message serializing/deserializing, request routing and networking, so 59% CPU time seemed odd.

Using CPU concurrency primitives in the place of actors, optimizing memory allocation, and concurrent memory access significantly reduce the footprint of the adding visitor path. The new implementation takes only 0.4% of CPU time on the same benchmark run.

The actor model currently in use in libmCaptcha is causing a bottleneck that was previously not visible. Protobuf implementation from #1 was yielding very low throughput. Generating [a flamegraph of that version showed that the [`AddVisitor` message handler](https://github.com/mCaptcha/libmcaptcha/blob/e3f456f35b2c9e55e0475b01b3e05d48b21fd51f/src/master/embedded/counter.rs#L124) was taking up ~59% of the CPU time while the node was hit with the locust benchmark. Keeping track of visitor count is a critical path in the variable difficulty algorithm, but this particular test also involved message serializing/deserializing, request routing and networking, so 59% CPU time seemed odd. ![flamegraph](https://git.batsense.net/mCaptcha/dcache/raw/branch/optimize-libmcaptha/bench/results/v2/libmcaptcha-bottleneck/problem/flamegraph.svg) Using CPU concurrency primitives in the place of actors, optimizing memory allocation, and concurrent memory access significantly reduce the footprint of the adding visitor path. [The new implementation takes only 0.4% of CPU time](https://git.batsense.net/mCaptcha/dcache/raw/branch/optimize-libmcaptha/bench/results/v2/libmcaptcha-bottleneck/solution/flamegraph.svg) on the same benchmark run. ![flamegraph of the new solution](https://git.batsense.net/mCaptcha/dcache/raw/branch/optimize-libmcaptha/bench/results/v2/libmcaptcha-bottleneck/solution/flamegraph.svg)