mojo-sort
https://github.com/mzaks/mojo-sort
By @Maxim
GitHub
GitHub - mzaks/mojo-sort
Contribute to mzaks/mojo-sort development by creating an account on GitHub.
37 Replies
Added parallel tim sort. On large lists it seems to be up to 3x faster than Tim sort (tried on m1 and Intel i7 with 8 cores), but still slower than Radix sort. I guess on a machine with 64 Cores that might be a different story though.
If it can be useful and there is a benchmark script. I can run it on one of our quiet servers (96 cores Threadripper - Ubuntu server 22.04)
That would be awesome, the benchmark script can be found here: https://github.com/mzaks/mojo-sort/blob/main/benchmark.mojo if you give me a few minutes I can also make it output proper CSV.
GitHub
mojo-sort/benchmark.mojo at main · mzaks/mojo-sort
Contribute to mzaks/mojo-sort development by creating an account on GitHub.
No worries, just ping me when ever you are happy with the script.
it's ready 🙂
BTW benchmark results for M1 are her https://github.com/mzaks/mojo-sort/blob/main/benchmark_result_m1.csv
GitHub
mojo-sort/benchmark_result_m1.csv at main · mzaks/mojo-sort
Contribute to mzaks/mojo-sort development by creating an account on GitHub.
GitHub
mojo-sort/benchmark_result_i7.csv at main · mzaks/mojo-sort
Contribute to mzaks/mojo-sort development by creating an account on GitHub.
The
0 values in m1 benchmark for smaller lists are due to the fact that time function on mac does not give nanosecond precision it rounds to microsecondsShould I run it on Mojo nightly or the main branch?
Nightly please.
Speaking of, I think it is possible to implement a parallel Radix sort, which is kind of cool 🙂
Here is the results.
I ran the benchmark using
The server that I ran it on has dual AMD EPYC 7401 24 cores, 2 threads/core.
All the cores were working by they didn't reach 100%
Hm, interesting
generally the single core algorithms seems to be much slower, then m1 (I guess the machine has lots of CPU but they are not as powerful), you know what, I think I am doing the parallelisation wrong 🙂 need to rewrite the algorithm
ok, will continue on it tomorrow 🙂
afaik LLVM struggles to interleave vector reads on ARM
might be a factor
Updated the lib after the
ComparableCollectionElement landed in nightly. Benchmarks for string sorting are quite interesting.Benchmarks for string sorting are quite interesting.Interesting bad or interesting good? Or just weird?