Single thread and thread-safe performance

[Janmajayamall]

Hey!

If I create a single Ntt instance and call NTTComputeForward on it from multiple threads, then will the performance increase by factor of CPU cores available ? Or will it remain the same ?

Does having a different Ntt instance per thread, affect the performance ?

Thanks!

[Janmajayamall]

I am using m6i.2xlarge ec2 instance.

For 5000 NTTComputeForward operations on vector size 2^15 it takes:
126 ms on 8 cores
127 ms on 7 cores
130 ms on 6 cores
135 ms on 5 cores
140 ms on 4 cores
188 ms on 3 cores
280 ms on 2 cores
560 ms on 1 core

I am curious about why the performance improvement isn't linear with num of cores in use ?

You can find the test code here. Try running with cargo test --release tests::rayon_ntt -- --nocapture and you can set the number of cores using export RAYON_NUM_THREADS=no_of_cores (where no_of_cores specifies no. of cores to use).

[Janmajayamall]

This is observed due to m6i.2xlarge only having 4 real cores. After switching to m6i.4xlarge with 8 cores, it works as expected