sintharic/performance-tests

Collection of small scientific projects comparing different programming languages and impelementations with respect to performance and ease of use.

Collection of small scientific projects comparing different programming languages and impelementations with respect to performance and ease of use.

FFT results

Benchmark consists of two 2D double precision arrays (resolution N x N) and their Fourier transforms. A triangular wave in x- and y-direction is generated one wave vector after the other and transformed back and forth each time.

Computation time on MacBook Pro 2019 at N=1024:

• Julia (using Intel MKL): 7.6 s
• Julia (using FFTW): 9.0 s
• C++: 12.6 s
• Python: 15.5 s

Memory usage (according to top) on MacBook Pro 2019 at N=1024:

• C++: 50 MB
• Python: ~95 MB (32 MiB according to tracealloc)
• Julia (using FFTW): ~160 MB (352 MiB according to maxrss???)
• Julia (using Intel MKL): ~170 MB (352 MiB according to maxrss???)

Benchmarks involving different resolutions or other computational tasks may follow in the future.

Caveats

Figuring out memory usage appears to be a difficult task, since it does not necessarily take all the linked libraries into account. While Julia and Python both provide their own internal measurement tools, C++ does not and valgrind does not seem to work on Mac. With zero overhead, given that a double is 8 byte and a complex<double> is 16 byte, I would calculate an ideal memory usage of 2*8*1024*1024 + 2*16*1024*513 = 33.6 MiB, i.e. just above 2^25 bytes. Python's internal estimate was somehow below that value. And the estimate reported by my tracker function in Julia was an entire order of magnitude above...

Summary

• Julia's speed is very impressive.
• Julia's 1-based indexing is very inconvenient.
• The amount of C++ boilerplate necessary to perform such easy tasks makes me very sad.
• I need to figure out better ways to measure memory consumption before I continue with these benchmarks.