Fixed Mojo + benchmark script

[uberkael]

This is incorrect. The benchmarks only measure the Python startup overhead and the Mojo compilation overhead; the test programs are small compared to those times.

I'm not an expert, but out of curiosity, I did update the Mojo code and compiled it before running the hyperfine call with a shell script bench.sh.

As you can see all the times are mostly the same, meaning all the Leetcode examples are too small for a benchmark, even if mojo compiled is 9x faster in this case.

❯ bench.sh
Benchmarking problem-1
Benchmark 1: python problem-1.py
  Time (mean ± σ):      71.0 ms ±   3.9 ms    [User: 39.1 ms, System: 32.3 ms]
  Range (min … max):    62.7 ms …  82.3 ms    45 runs
Benchmark 2: problem-1
  Time (mean ± σ):       9.3 ms ±   5.2 ms    [User: 2.2 ms, System: 8.4 ms]
  Range (min … max):     5.9 ms …  53.1 ms    417 runs
Summary
  problem-1 ran
    7.63 ± 4.28 times faster than python problem-1.py
Benchmarking problem-2
Benchmark 1: python problem-2.py
  Time (mean ± σ):      70.8 ms ±   3.4 ms    [User: 40.0 ms, System: 31.4 ms]
  Range (min … max):    63.5 ms …  78.1 ms    43 runs
Benchmark 2: problem-2
  Time (mean ± σ):       8.1 ms ±   1.8 ms    [User: 2.0 ms, System: 7.9 ms]
  Range (min … max):     6.0 ms …  18.2 ms    350 runs
Summary
  problem-2 ran
    8.76 ± 2.00 times faster than python problem-2.py
Benchmarking problem-3
Benchmark 1: python problem-3.py
  Time (mean ± σ):      70.8 ms ±   4.6 ms    [User: 37.4 ms, System: 34.2 ms]
  Range (min … max):    61.9 ms …  83.7 ms    45 runs
Benchmark 2: problem-3
  Time (mean ± σ):       8.5 ms ±   2.1 ms    [User: 2.3 ms, System: 7.8 ms]
  Range (min … max):     6.1 ms …  20.4 ms    429 runs
Summary
  problem-3 ran
    8.33 ± 2.16 times faster than python problem-3.py
Benchmarking problem-4
Benchmark 1: python problem-4.py
  Time (mean ± σ):      73.9 ms ±   5.7 ms    [User: 41.9 ms, System: 32.5 ms]
  Range (min … max):    64.7 ms …  87.4 ms    44 runs
Benchmark 2: problem-4
  Time (mean ± σ):       8.0 ms ±   1.7 ms    [User: 2.1 ms, System: 7.7 ms]
  Range (min … max):     6.1 ms …  19.0 ms    420 runs
Summary
  problem-4 ran
    9.24 ± 2.10 times faster than python problem-4.py
Benchmarking problem-5
Benchmark 1: python problem-5.py
  Time (mean ± σ):      72.0 ms ±   3.7 ms    [User: 38.5 ms, System: 33.9 ms]
  Range (min … max):    64.1 ms …  81.2 ms    41 runs
Benchmark 2: problem-5
  Time (mean ± σ):       8.6 ms ±   2.1 ms    [User: 2.1 ms, System: 8.0 ms]
  Range (min … max):     6.0 ms …  19.7 ms    370 runs
Summary
  problem-5 ran
    8.37 ± 2.13 times faster than python problem-5.py```

[SaadBazaz]

Hey @uberkael ,
Thank you for the Pull Request. Reviewing it, would love to improve the benchmarks for the community.

[SohaibBazaz]

Hello @uberkael, Thanks for the review!
As a beginner, based on my readings, it appears that Mojo must be compiled before it can be executed. Therefore, the issue i think that I encountered in measuring only the "compilation overhead" might be due to the way I conducted the benchmarking, correct?

[uberkael]

Yes,
But also those examples are below the millisecond of runtime so the benchmark is almost invisible.
Try to execute them multiple times, I will update the PR.

[SohaibBazaz]

I could see why importing from python was not necessary in my other codes but here this should be working just fine, no?

[SohaibBazaz]

I'm trying to understand why you changed the defining keyword here from "def" to "fn". I mean it worked fine with "def", any reason for that?

[uberkael]

fn in mojo allow to use types, which enable the compiler to optimize more effectively.

[ayghri]

Thanks for the PR. I found it strange that Mojo is ~ x20 slower. I was just googling benchmarks and stumbled upon this. I've never used Mojo and these results almost swayed me from doing so.

These examples are definitely misleading with the compilation overhead and lack of optimizations.

[uberkael]

I updated the code and removed all print() statements.
It appears that the Mojo print function is still not well optimized and makes the code quite slow.
modular/max#975 (comment)

The results:

Summary
  problem-1.bin ran
    4.08 ± 0.25 times faster than mojo problem-1.mojo
    9.01 ± 0.58 times faster than python problem-1.py

  problem-2.bin ran
    4.72 ± 0.47 times faster than mojo problem-2.mojo
   10.30 ± 0.97 times faster than python problem-2.py

  problem-3.bin ran
    2.22 ± 0.08 times faster than python problem-3.py
    3.27 ± 0.10 times faster than mojo problem-3.mojo

  problem-4.bin ran
    3.53 ± 0.26 times faster than mojo problem-4.mojo
    4.97 ± 0.35 times faster than python problem-4.py

  problem-5.bin ran
    3.95 ± 0.46 times faster than mojo problem-5.mojo
    7.20 ± 0.64 times faster than python problem-5.py

@ayghri @SaadBazaz @SohaibBazaz

[ayghri]

print statement are always going to slow down the program due to the I/O operations. The alternative is to use a I/O buffer, but it seems Mojo doesn't have them yet.

[SaadBazaz]

We can exclude print statements for the raw benchmark and find a way to display them only before/after the benchmark timing function. @ayghri @uberkael

[soraros]

You will need to write keep(lengthOfLastWord("Hello World"), otherwise lenghOfLastWord gets DCE'd.

from benchmark import keep

...