My measurements

[buybackoff]

Salut Alexandre,

To avoid issues like I had with @nietras here is a preview of my measurements before I publish.

For the extended dataset that I prefer, you have almost made it to the top.

But with the default dataset I would expect at least 2x gain, bit not. I have two such datasets.

The methodology is described here.

[xoofx]

Thanks! I didn't look at other datasets but just used the default 1B one, so some defaults or optimizations might not work well with other cases. For example:

The extended dataset with 10K random weather station names

Definitely a case where I could see that the defaults might not be optimized for this case. But that's weird that it is the opposite. 😀 Will check

[xoofx]

To avoid issues like I had with @nietras here is a preview of my measurements before I publish.

Oh, what do you mean by issues? 😅

[xoofx]

And nice, didn't know that you wrote an entire article, congrats! ☺️

[buybackoff]

Thanks! I didn't look at other datasets but just used the default 1B

I will say again: your 10K results are great! But on my Linux machine your results on the default dataset are much slower than expected. Maybe you have some hardcoded assumptions about CPU count or something like that?

[buybackoff]

Oh, what do you mean by issues? 😅

Nothing worth mentioning ATM, we have sorted it out 😄 Haven't we, @nietras?

[buybackoff]

And nice, didn't know that you wrote an entire article, congrats! ☺️

Thanks!

[xoofx]

But on my Linux machine

Yeah, I would not be surprised of a different behavior. I only benchmarked on Windows x64 (and tried M1 ARM64 as well, seems it worked ok there)

[xoofx]

Double checked with 10k and regular dataset and the numbers on my machine feel normal:

Also, I checked on Linux (WSL2) and it seems even doing slightly better than on Windows.

Trying with your repository with latest version on the same dataset:

Something seems off on your machine. I would believe that it is more IO related, but can't be certain. For example, on my machine, I'm running almost 2 times faster on all datasets but you seem to indicate that it's not the case on your machine. You should probably try to run your benchmarks on different machine specs.

[buybackoff]

When you say your Ryzen is at 60W what does that mean? I see that its typical TPD is 170W. Is your CPU frequency fixed? For a cross-check it would be nice to calculate the sum of frequencies for all cores and scale the results across machines. (BTW is there a term for it? E.g. for my machine it's 2.5x6 = 15 GHz hyperthreaded)

I restarted the machine and now see meaningful numbers. Do not know what happened yesterday, I'm running the same commands.

[xoofx]

When you say your Ryzen is at 60W what does that mean? I see that its typical TPD is 170W. Is your CPU frequency fixed?

Yeah, I limit the power consumption of my machine as suggested by @nietras a while back, I discovered that it's only 10% slower at 60W (benchmarks here). CPU frequency is not fixed (because it adapts to thermal power requirement), but when you run things with hyperfine, it should not a big deal.

[buybackoff]

But what is the average frequency during the runs? No need for an exact number, just your guess, please.

[xoofx]

But what is the average frequency during the runs? No need for an exact number, just your guess, please.

It's clocking at 4.05 GHz when running 10 times the 10k (Base speed is 4.50 GHz)

[buybackoff]

Oh, it looks like WSL uses only half of the cores. The scaling makes sense then for both Windows and WSL

[xoofx]

Oh, it looks like WSL uses only half of the cores. The scaling makes sense then for both Windows and WSL

You mean WSL, not WSL2? Not sure I noticed that on WSL2, same machine specs...

Btw, looking at the accuracy of the results, It seems that your version is off on 2 lines with my 10k run:

I compared it with other versions and it seems consistent (well, the one that are working with 10k 😅 ), so there might an issue somewhere in your code.