inspired by this repoistory, here I conducted a basic performance comparison betweenand C and Python with the same task. the task is to find first 1e+7 primes.
| code | Machine-1 Runtime | Machine-2 Runtime |
|---|---|---|
| prime-int32.c | 36s (24s with -O3) | 56s (37s with -O3) |
| prime-int64.c | 50s (42s with -O3) | ~2 min |
| dummy-prime.py | > 24 hour | > 24 hour |
| simple-prime.py | ~30 min | ~25 min |
| numpy-prime.py | 77s (65s int32) | ~3.5 min |
| numba-prime.py | 30s | 51s |
| code | Machine-1 Runtime | Machine-2 Runtime |
|---|---|---|
| parallel-prime.py | 9.7s | 3.1s |
on both machines namba code was able to beat c/c++ (g++ compiled) in 64-bit calculations (with or without -O3 option), and it was also able to beat g++ without -O3 option in the 32-bit calculations
1- what makes the diffrent of -O3 option?
2- why in64 calculations affect c++ that much but it barely affect numba runtime (kind of 1% at most)?
3- why my economic core-i5 is faster than the Hi-end 64-core Xeon in this benchmark (when its pluged in)
4- what makes the diffrent of LLVM compiled (I mean numba) and g++ compiled runtime