Alternative results
#103
Replies: 12 comments 57 replies
-
|
Thank you very much, @paweld. It is great to see that FPC can produce such good results from my new pure-Pascal implementation. Hope FPC team adds support for them soon. |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
Date: 2024-04-14 all stations, 1 000 000 000 measurements, 32 threads json results: results mt b2-120.zip |
Beta Was this translation helpful? Give feedback.
2 replies
-
|
Thank you very much @paweld. |
Beta Was this translation helpful? Give feedback.
-
|
You're welcome. |
Beta Was this translation helpful? Give feedback.
-
|
Date: 2024-04-18 Run - threads: 32 | records: 1_000_000_000 | stations: ALL Run - threads: 32 | records: 1_000_000_000 | stations: 400 Run - threads: 32 | records: 5_000_000_000 | stations: ALL json results: results_20240418.zip @OttoCoddo Sorry, but in your case only the results of the last test were saved, because I had an error in the script. Next time they will be saved all |
Beta Was this translation helpful? Give feedback.
5 replies
-
|
@paweld that's awesome, thanks a lot! |
Beta Was this translation helpful? Give feedback.
-
|
Thanks a lot for all those numbers! AFAICT OVH Public Cloud b3-128 have 64 vCore. Anyway, your result tables are very informative. Our good It is also interesting to see that the Last but not least, it seems that cpu/core affinity has no negative impact on this HW. Perhaps because they are virtual machines, cpu affinity is just ignored. I don't know. Do you have any clue? I have ideas for a new version, which I will implement this week end. ;) |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse),
Care to explain? And how could I make things better on that front? Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
@gcarreno On the contrary, @paweld tests are run on a clean idle machine. Perhaps if you include all .lpr on a run, it could help to make accurate comparisons. |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Gotcha !!! I, finally, had a run on a very quiet machine. Only graphical thing running was the GNOME login and ran it on Here's the Next Saturday run will be made in the same fashion. And the results of that will be used instead of a less quiet machine. I was very surprised that the Desktop alone introduced so much noise to the results. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Hi @synopse
I set up an account on "ovh public cloud" only because of testing, so I have limits on the account as to the parameters of the machines. one of them is the number of 34 cores per location. From the panel I can not unlock this, but I will try to write an email to support, maybe they will unlock.
I do not know what specific model it is, but it is the |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
Thanks a lot for the feedback! I will implement some new algorithms and some asm tricks this week-end, if I have time. |
Beta Was this translation helpful? Give feedback.
-
|
Now should work also with 8 threads. |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
Hi everyone, Run - threads: 32 | records: 1_000_000_000 | stations: ALL Run - threads: 32 | records: 1_000_000_000 | stations: 400 Run - threads: 32 | records: 5_000_000_000 | stations: ALL Run - threads: 8 | records: 1_000_000_000 | stations: 400 Unfortunately, I forgot to copy the json files from this test, so I had to create the instance from scratch and run the test :-) but this time the results are a little worse than those presented above. All tests with the exception of the |
Beta Was this translation helpful? Give feedback.
5 replies
-
|
So here is my new best: Now we understand why the previous answers were almost identical, whereas they should show a difference, as it does with your latest results. Your testing is now consistent with what I got on my Intel Core i5-13500 PC - only faster because of a better CPU. Thanks a lot! |
Beta Was this translation helpful? Give feedback.
-
|
Again, I apologize for the mistake in the previous tests. As for performance, unfortunately, such is the charm of cloud computing. I also ran some tests on Google machines (same AMD CPU series, number of vCPUs and RAM) but the results were 2 times worse. |
Beta Was this translation helpful? Give feedback.
-
|
Hello @paweld , Alternatively, how could I potentially request another variant implementation? (I currently have 2 build-modes: Thanks again! |
Beta Was this translation helpful? Give feedback.
-
|
Hi @georges-hatem |
Beta Was this translation helpful? Give feedback.
-
|
@paweld ok thanks for the note. In my LPI, I have 5 different configurations, of which there are 3 I'd like to run:
They basically run three different hash methods for the lookup function. |
Beta Was this translation helpful? Give feedback.
-
|
Hi everyone, Run - threads: 32 | records: 1_000_000_000 | stations: ALL Run - threads: 32 | records: 1_000_000_000 | stations: 400 Run - threads: 32 | records: 5_000_000_000 | stations: ALL Run - threads: 8 | records: 1_000_000_000 | stations: 400 All results: 1brc_wyniki_20240430_090024.zip |
Beta Was this translation helpful? Give feedback.
15 replies
-
|
Hi @paweld Thanks a lot for your very valuable input, because @gcarreno sadly seems to have no time for running such tests... :( |
Beta Was this translation helpful? Give feedback.
-
|
@georges-hatem I understand that now I only have to compile the Release configuration for both projects? Just change the target file name for the OneBRC-nosharedname project, it is currently the same in both projects. @synopse I have already applied the changes to the script file. Tomorrow I will probably run new tests. |
Beta Was this translation helpful? Give feedback.
-
|
@paweld I realized that and made the changes to my fork (PR not merged). Readme is edited to describe the two projects |
Beta Was this translation helpful? Give feedback.
-
|
Yes, I downloaded your latest changes, but about 10 minutes ago and still the names were the same for both projects. Now I see that you have already changed :-) |
Beta Was this translation helpful? Give feedback.
-
|
@paweld there's a third version / LPR for me. Also it seems i still had an issue with my configurations as you mentioned before, should be adjusted now |
Beta Was this translation helpful? Give feedback.
-
|
Hi everyone, !!! I also want to explain why the Run - threads: 32 | records: 1_000_000_000 | stations: ALL Run - threads: 32 | records: 1_000_000_000 | stations: 400 Run - threads: 32 | records: 5_000_000_000 | stations: ALL Run - threads: 8 | records: 1_000_000_000 | stations: 400 I have not yet managed to run the And as usual, the results in the attachment: 1brc_wyniki_20240503_172904.zip Have a nice weekend. |
Beta Was this translation helpful? Give feedback.
15 replies
-
|
Note that this Java version is cheating about the system time: they The original Java challenge was officially running with 8 threads only (for a weird reason I still don't get). It is really weird that the numbers of this OVH cloud VM are always diverse between runs. |
Beta Was this translation helpful? Give feedback.
-
Unfortunately no, I don't even know how to do it. I ran simply:
During this test, it was 1 ms slower than the standard version. :-)
It all depends on the data center (ssd disk usage), the current load of other machines using the same physical processor, ram and ssd disk Example: yesterday it took ~8 minutes to generate a file with 5b measurements, and today only ~2.5 minutes. |
Beta Was this translation helpful? Give feedback.
-
|
Hello @paweld, thanks a lot for your valuable informations. You wrote that my program fails for 5b rows or 400 stations. I'm not surprised, because I (until now) never tested this and I assume overflows in the temperatures average management (sum = type longint, count = type word). Question to @ALL: do we have sha256-checksums for the results with
Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
I do the tests for
|
Beta Was this translation helpful? Give feedback.
-
|
Thank you very much @paweld for those checksums. |
Beta Was this translation helpful? Give feedback.
-
|
@paweld So it sounds that on AMD, the I have asked @gcarreno to check on his hardware, before the official run of Saturday. #38 (reply in thread) |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
@synopse @OttoCoddo @georges-hatem @hg747 @Laksen |
Beta Was this translation helpful? Give feedback.
5 replies
-
|
Hello Pawel (@paweld), thank you for this offer. I will put a new PR for a new version 2.10 today or tomorrow. On the machine of Gus it needs no forther optimization, but on your machine if would be great if you could do this 3 tests with the following command line parameters:
Thanks and Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Hello Everyone, @paweld thanks as always. Finally, i converged to a single version for the 1B rows / 40k stations. |
Beta Was this translation helpful? Give feedback.
-
|
Hello Pawel (@paweld), as said yesterday, I made a new PR for a new version (2.10). Please run on Wednesday this 3 tests with the following command line parameters:
Please pay attention, that this version does NOT work for 400 stations or 5 billion rows (due to integer overflows). If you want, you can run them only to see the timing, but the result will be incorrect. Thanks and Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
@hg747 And what values to use as the fourth parameter? |
Beta Was this translation helpful? Give feedback.
-
|
@paweld The new version (2.10) has only 3 command line parameters. |
Beta Was this translation helpful? Give feedback.
-
|
Hi everyone, Date: 2024-05-08 At the end are for comparison tests of two best solutions in Run - threads: 32 | records: 1_000_000_000 | stations: ALL Run - threads: 32 | records: 1_000_000_000 | stations: 400 Run - threads: 32 | records: 5_000_000_000 | stations: ALL Run - threads: 8 | records: 1_000_000_000 | stations: 400 Solution results from the so I think they are executed on all 32 threads. also downloadable detailed results: 1brc_wyniki_20240508_144233.zip |
Beta Was this translation helpful? Give feedback.
4 replies
-
|
So for 41K stations and 1,000,000 rows, the best Java solutions are twice slower than our pascal version (with full name check). |
Beta Was this translation helpful? Give feedback.
-
|
Thanks @paweld for this test, its safe to assume that @synopse beat java already based on the results: It's official, pascal is faster than JAVA. Thanks everyone. |
Beta Was this translation helpful? Give feedback.
-
|
is faster, but you have to remember that the |
Beta Was this translation helpful? Give feedback.
-
|
The Java version is IMHO cheating Also note that with 41K stations, our object pascal code is clearly much faster than Java. |
Beta Was this translation helpful? Give feedback.
-
|
Hi everyone, At the end are for comparison tests of two best solutions in Run - threads: 32 | records: 1_000_000_000 | stations: ALL Run - threads: 32 | records: 1_000_000_000 | stations: 400 Run - threads: 32 | records: 5_000_000_000 | stations: ALL Run - threads: 8 | records: 1_000_000_000 | stations: 400 Also downloadable are detailed test results: 1brc_wyniki_20240515.zip |
Beta Was this translation helpful? Give feedback.
5 replies
-
|
Solution results from the JAVA version - tests performed using: so I think they are executed on all 32 threads. |
Beta Was this translation helpful? Give feedback.
-
|
@paweld thank you very much for consistently providing these results, really helped in evaluating progress over the last month :) |
Beta Was this translation helpful? Give feedback.
-
|
@georges-hatem Here you go. I only now noticed that there are wrong descriptions for your versions on the summary, the correct descriptions are:
|
Beta Was this translation helpful? Give feedback.
-
|
and... back to be slower than JAVA on the 400s 1br:
|
||||||||||||||||||||||||||||||||||||||||||||||||
Beta Was this translation helpful? Give feedback.
-
|
For 400 stations, there is a slightly faster Pascal entry: Worth saying that this JAVA solution is clearly cheating by But for 41K stations, we are almost twice faster... ;) |
Beta Was this translation helpful? Give feedback.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
@gcarreno @synopse @OttoCoddo @bytebitespas
Date: 2024-04-11
Test environment: OVH Public Cloud b3-128 (32 vCPU AMD, 128 GB RAM)
Debian 12
Lazarus trunk / FPC trunk x64
Beta Was this translation helpful? Give feedback.
All reactions