Replies: 31 comments 88 replies
-
|
Ah this was a newer method, not available in the release. |
Beta Was this translation helpful? Give feedback.
-
|
I have made a pull request to fix this - and enhance the performance by 5% too. ;) |
Beta Was this translation helpful? Give feedback.
-
|
Your results on a cold run: time ./bin/abouchez /tmp/measurements-1_000_000_000.txt -t 30
{...}
real 0m9.134s
user 1m17.539s
sys 0m14.066s |
Beta Was this translation helpful? Give feedback.
-
|
You seem to have a "slow" SSD, reading at around 1GB/s. ;) |
Beta Was this translation helpful? Give feedback.
-
|
Please try to run all the commands I wrote in the README file, so guess which parameters are the best for your hardware. |
Beta Was this translation helpful? Give feedback.
-
|
That was a cold start. |
Beta Was this translation helpful? Give feedback.
-
|
Running the full automation: |
Beta Was this translation helpful? Give feedback.
-
|
This is with only 30 threads and with a busy system. Thanks Arnaud!! |
Beta Was this translation helpful? Give feedback.
-
|
Nice! I am a bit worried about the Please try with 16 and 24 threads too. |
Beta Was this translation helpful? Give feedback.
-
|
With 24 threads on |
Beta Was this translation helpful? Give feedback.
-
|
2.4 seconds and 16 threads is consistent what I expected, since I reached 4 seconds on my Intel PC with 8-10 threads over 6 P-cores (and some E-cores). |
Beta Was this translation helpful? Give feedback.
-
|
If the watchdog bugs me, I'll dig into that link. Thanks !! Btw, I have the threads as a configuration field on my |
Beta Was this translation helpful? Give feedback.
-
|
We would rather put |
Beta Was this translation helpful? Give feedback.
-
|
Awesome sauce, will do that !! |
Beta Was this translation helpful? Give feedback.
-
|
Did you try with the Just curious, if thread affinity helps here. |
Beta Was this translation helpful? Give feedback.
-
|
Did not, gimme a sec 😄 |
Beta Was this translation helpful? Give feedback.
-
|
Didn't improve with affinity. |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), In the spirit of cross platform, at least, would it be too much to ask if you can do the If it's too much I guess we can skip that. Would really appreciate it if you could make the effort. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
There are already asm version of To learn about cross-platform and cross-compiler, look at the mORMot source code, it is a much better place. |
Beta Was this translation helpful? Give feedback.
-
|
Awesome !! |
Beta Was this translation helpful? Give feedback.
-
|
I tried to clean the comments about the asm, and updated the README thanks to your feedback. |
Beta Was this translation helpful? Give feedback.
-
|
I'm absolutely LOVING your Thank you so very much!!! |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Does this matter? If so, where should I start investigating? Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
This is an Lazarus parsing bug only. So, as I wrote above, I think this is a false positive. (to be honest, I never use lazbuild, I do prefer a manual .sh file directly calling fpc with the proper switches) |
Beta Was this translation helpful? Give feedback.
-
|
I'll take your word. BTW, Székely sends his congratulations on such a magnificent result!! |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Latest PR. 16 Threads 24 threads 32 Threads Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Amazing! |
Beta Was this translation helpful? Give feedback.
-
|
More than welcome !! |
Beta Was this translation helpful? Give feedback.
-
|
Did you make a try with I doubt it could be better, but anyway.... |
Beta Was this translation helpful? Give feedback.
-
|
No I did not. Here is is: $ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 2.522 s ± 0.014 s [User: 29.072 s, System: 1.763 s]
Range (min … max): 2.502 s … 2.543 s 10 runs
===========
$ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 2.522 s ± 0.014 s [User: 29.072 s, System: 1.763 s]
Range (min … max): 2.502 s … 2.543 s 10 runs
===========
$ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.729 s ± 0.010 s [User: 33.461 s, System: 2.023 s]
Range (min … max): 1.710 s … 1.743 s 10 runs
=========== |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Was trying to run your entry with the 400 stations file( 8 threads ), thanks to @paweld, and I got this: $ ./bin/abouchez /tmp/measurements-400-1_000_000_000.txt -t 8
Fatal exception EAccessViolation raised with message
Access violation
Program will now abortCheers, |
Beta Was this translation helpful? Give feedback.
-
|
Should be fixed now with my latest pull request. |
Beta Was this translation helpful? Give feedback.
-
|
Awesome sauce !! |
Beta Was this translation helpful? Give feedback.
-
|
Latest patch results. 3rd run with $ time ./bin/abouchez /tmp/measurements-400-1_000_000_000.txt -t 8
{...}
real 0m2.746s
user 0m16.317s
sys 0m1.391s4th or 5th run ( I forget ) for the 400/8 with affinity: time ./bin/abouchez /tmp/measurements-400-1_000_000_000.txt -a -t 8
{...}
real 0m2.698s
user 0m16.381s
sys 0m1.315sand $ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.732 s ± 0.031 s [User: 33.319 s, System: 1.956 s]
Range (min … max): 1.700 s … 1.805 s 10 runs
=========== |
Beta Was this translation helpful? Give feedback.
-
|
So does thread affinity |
Beta Was this translation helpful? Give feedback.
-
|
On the official run? |
Beta Was this translation helpful? Give feedback.
-
|
$ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.768 s ± 0.012 s [User: 33.286 s, System: 2.067 s]
Range (min … max): 1.743 s … 1.782 s 10 runs
===========
$ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.771 s ± 0.010 s [User: 33.415 s, System: 2.056 s]
Range (min … max): 1.760 s … 1.786 s 10 runs
===========I'm really bad at the calculation of "statistically significant", so what do you say? |
Beta Was this translation helpful? Give feedback.
-
|
My guess is that we are in the error margin, and usual feedback is that thread affinity could do more harm than good, with a modern OS. So, we better not set thread affinity. |
Beta Was this translation helpful? Give feedback.
-
|
Aye aye Sir!! No affinity it is !! |
Beta Was this translation helpful? Give feedback.
-
|
I have added #56 With updated README file. Thanks a lot! |
Beta Was this translation helpful? Give feedback.
-
|
About the new And see the results. |
Beta Was this translation helpful? Give feedback.
-
|
I'm at the very, very end of my awake cycle, and maybe pushing it a bit too far, so I'll do these tests after I've had some sleep. These will be easy since I only have to copy/paste the same line and then just edit the I'll keep you posted once I've run them!! |
Beta Was this translation helpful? Give feedback.
-
|
No hurry - we have plenty of time. Nice dreams! 💤 |
Beta Was this translation helpful? Give feedback.
-
|
Merci bien !! 🫂 ❤️ |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Hope you don't mind me doing a bit of formatting clean-up on your If you do, please chastise me a ton and I'll, maybe, promise not to do it again 😉 🫂 Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
I'm back on |
Beta Was this translation helpful? Give feedback.
-
|
Has requested: Which one should I leave for the official run? |
Beta Was this translation helpful? Give feedback.
-
|
It is a bit weird that the default chunk makes 1.733 s mean, but Chunk 4 only 1.766 s. So perhaps we could use |
Beta Was this translation helpful? Give feedback.
-
|
So it seems that the new SSE2 code was included in this run. Could you please try to disable the SSE2 code: to verify that the SSE2 code is actually faster than our new unrolled loop in pure pascal? Thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Can you please send a PR? Yes, I'm lazy 😺 I'm also dealing with this ATM: https://github.com/gcarreno/1brc-ObjectPascal/tree/main/spelunking/roundex Many thanks !!! And I'll edit the automation to use |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Your latest Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
So the pure pascal naive code seems to be actually faster than optimized SSE2 ! Thanks for the feedback. |
Beta Was this translation helpful? Give feedback.
-
|
As usual, you're more than welcome !!! |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), I think I messed up!! And I must give you my most profound apologies!!! I think I forgot to compile you last PR or something on those lines, cuz now that I did and performed the runs, you're back at 1.9s and not 1.7s. I think this means that the Again, really sorry to have messed this up!! The weekly Saturday results are up. Will update them when you make a new PR reverting what you did lastly. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
I'll do that. |
Beta Was this translation helpful? Give feedback.
-
|
Ok, just ran your last PR: $ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.635 s ± 0.010 s [User: 32.665 s, System: 1.864 s]
Range (min … max): 1.620 s … 1.651 s 10 runsThis is even quicker that the ~1.7s we had before !!! |
Beta Was this translation helpful? Give feedback.
-
|
@gcarreno |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), I'm really sorry, but no. Got a bit distracted by the new toys and I'm still a bit behind. Will have to be a Monday run, just this once. I blame the shinies !! 😝 Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Completely forgot that I only listed the new toys on the Discords... |
Beta Was this translation helpful? Give feedback.
-
|
It is weird that the last run did return numbers higher that what you found here. Could you try with |
Beta Was this translation helpful? Give feedback.
-
|
Yeah, sure: $ hyperfine -w 1 -r 10 -N -n abouchez './bin/abouchez /tmp/measurements-1_000_000_000.txt -t 32 -c 1'
Benchmark 1: abouchez
Time (mean ± σ): 1.764 s ± 0.010 s [User: 33.068 s, System: 2.169 s]
Range (min … max): 1.754 s … 1.784 s 10 runs |
Beta Was this translation helpful? Give feedback.
-
|
The same code did make 1.635 last week (see above). Weird. |
Beta Was this translation helpful? Give feedback.
-
|
Another question: why some results are using lazarus-3.99, fpc-3.3.1 and some lazarus-3.0, fpc-3.2.2 ? My guess is that fpc-3.3.1 should generate better code, so it seems a bit unfair. |
Beta Was this translation helpful? Give feedback.
-
|
This was a recent thing due to 2 entries that do not compile on fpc-3.2.2. I've even attempted to compile all with fpc-3.3.1, but was not able to. And due to my new gear I've been a bit distracted from My intention is to have everyone ported to 3.3.1, eventually. Gimme some time and I'll migrate everyone there. Even you. And on that question, would you want me to also use the |
Beta Was this translation helpful? Give feedback.
-
|
No need of using the mormot2/main branch. |
Beta Was this translation helpful? Give feedback.
-
|
Awesome sauce!! I've done the necessary changes to the automation tools in order to have control on choosing which I think I need to update some constants with the correct strings on the results table generator, but that's about it, I think. Again, all will be sorted before next Saturday. |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Latest $ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.766 s ± 0.031 s [User: 33.130 s, System: 2.374 s]
Range (min … max): 1.737 s … 1.849 s 10 runs |
Beta Was this translation helpful? Give feedback.
-
|
Here are the scripts |
Beta Was this translation helpful? Give feedback.
-
|
Regarding any |
Beta Was this translation helpful? Give feedback.
-
|
@synopse This is indeed weird. When I initially wrote this shared-jumper approach, I compared it with your results, and on my machine (i7 Serries 11, and Windows), it was nearly 2X faster (doing 3.5 s compared to yours 5.9 s). But on the @gcarreno machine, it was even a bit slower than yours. |
Beta Was this translation helpful? Give feedback.
-
|
@OttoCoddo In fact, if we compare with my machine: So, L1/L2 cache seems to be too small for our process, on both CPUs. So the L3 cache seems to be the bottleneck. And L3 cache is bigger on the AMD, and all our data is likely to fit in it . Without sharing, the per-thread temperature data size is 16 bytes * 43413 stations * 32 threads = 21 MB. My initial guess what that maintaining a hash table for each thread could be a good idea. And it seems it was on the Ryzen CPU, but not on Intels. But since the same code is reported with diverse results on @gcarreno machine (e.g. 1.635s vs 1.764s - see above), there may be some instability on this computer (e.g. background tasks, thermal throttling, configuration). Then the fastest entries < 2 seconds are within the error margin. So the results are not very relevant. Both entries are just too fast. ;) |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for the input. The difference is huge if we count the program as a universal project and not made only for that model of CPU. Yes, we can optimise for it as a competition, but our discussion about the reasons is even more interesting. About the test result, I agree—maybe a thermal throttling? Maybe too much IO from previous tests (5 for yours and 7 for mine) makes the machine tired. Yes, both are very fast—faster than almost anything I saw in other languages. Maybe we should try a 10BRC :D |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse) I just got a feed back from Craig Chapman, on his 32 cores, 64 threads and 128GiB of
That blows the JavaBros completely out of the water !!! Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), I jumped the gun, sorry!! This was a run with incorrect params and a premature exit. Again, sorry!! Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
^c is the maker of performance. |
Beta Was this translation helpful? Give feedback.
-
|
Very interesting video. |
Beta Was this translation helpful? Give feedback.
-
|
Yes, it was. Beside how it works, the process they went through to find out how is also fascinating. |
Beta Was this translation helpful? Give feedback.
-
|
Besides the information, it is totally baffling, that he can write like that. From the wrong side. |
Beta Was this translation helpful? Give feedback.
-
|
He is just right-handed and has an inverted t-shirt, I guess. ;) |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Last $ ./run_all.sh abouchez
******** Run ********
===== Arnaud Bouchez ======
-- SSD --
Benchmark 1: abouchez
Time (mean ± σ): 1.644 s ± 0.009 s [User: 30.553 s, System: 1.932 s]
Range (min … max): 1.631 s … 1.656 s 10 runs |
Beta Was this translation helpful? Give feedback.
-
|
Great! So we will wait for the next round then. ;) |
Beta Was this translation helpful? Give feedback.
-
|
Indeed!! |
Beta Was this translation helpful? Give feedback.
-
|
I have committed yesterday some new versions, done during the weekend. On my computer they are noticeably faster. Please try with and without To be accurate, here are the actual numbers on my local machine, with the latest current version of our code: Which are pretty good numbers, especially thanks to our If we force the full station name lookup (with the The "pure pascal" version of the code is not so far away: Interesting reading: @paweld @gcarreno Last but not least... |
Beta Was this translation helpful? Give feedback.
-
|
Hi, |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), I still need to keep my promise to Hartmut(@hg747) to get the perfect chunk size for the official run. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Hello Arnaud @synopse , If you have the time (and aside the great links you've posted in your readme), any recommended readings for such bit-manipulations and Delphi asm for beginners? |
Beta Was this translation helpful? Give feedback.
-
|
I would like to be able to run the measurements on the same input than Java or others - so we may need to have consistent line feed, i.e. only LF. From my own tests, it sounds like if at least my entry is faster than the Java or C entries I got. With less complicated code. Processing 3 lines at once may be not so efficient in practice: AVX/AVX2/AVX512 numbers seem to give almost no benefit when all CPU threads are consumed - the cores seem saturated with regular code, because the bottleneck is the CPU caches. This is also what make my entry peculiar: because I try to take into special account the CPU caches. No need to AVX512 if you need to wait for the data to be in your CPU L1 cache. ;) About asm there is https://stackoverflow.com/questions/7087086/good-resources-for-using-assembly-in-delphi but it is a bit outdated, and 32-bit only. In practice, the way I do it is to write a fast pascal version first. Then I use a disassembler (e.g. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks very much for this info, and the time for compiling them! I'll be researching each of those points 🫡 |
Beta Was this translation helpful? Give feedback.
-
|
But note that asm helps only for the final percents - unless you have something heavily parallelizable where SIMD could shine. For most process, like our CSV parsing, algorithms matter more than assembly. |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), I need to run these: On a quiet system. But I promise to give you proper results before next Saturday. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
@gcarreno Could you try to run for the 41K stations official run: And pickup the fastest for the official run. On AMD, it seems that And for the 400 stations with full station check (to be compared with Java): And pickup the fastest for the official run. |
Beta Was this translation helpful? Give feedback.
-
|
@gcarreno Any news? 🚶♂️ |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Yes!! I'm putting all Just concluded personal tests with Georges(@georges-hatem). Will continue with you and @hg747 next. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Hello Arnaud, |
Beta Was this translation helpful? Give feedback.
-
|
We have several arrays:
This was tuned to leverage the L1/L2/L3 CPU caches, and gives the best results from our tests. Using a prime number needs a modulo operation, and even if FPC is able to replace this modulo per a constant with a multiplication per the reciprocal, it is still slower than an arithmetic |
Beta Was this translation helpful? Give feedback.
-
|
noted, thanks for the clarification. the link to Lemire's blog was an interesting read |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), On a not so quiet system.
$ hyperfine -w 1 -r 5 -N -n 'abouchez-full-a' --export-json 'results/abouchez-full-a.json' './bin/abouchez /tmp/measurements-1_000_000_000.txt -a'
Benchmark 1: abouchez-full-a
Time (mean ± σ): 1.358 s ± 0.026 s [User: 18.044 s, System: 2.442 s]
Range (min … max): 1.328 s … 1.399 s 5 runs |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), Kewl, kewl!! Gotcha. I'm just sorry I didn't read this before launching this week's run. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
I am lost. Where can we find this |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), I'm not including it on the repo because it has some private data. Can you please tell me where's the best place to do so? Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
ab at synopse dot info |
Beta Was this translation helpful? Give feedback.
-
|
Hey Arnaud(@synopse), File sent. Cheers, |
Beta Was this translation helpful? Give feedback.
-
|
Hello Arnaud (@synopse), |
Beta Was this translation helpful? Give feedback.
-
|
There are some comments on each operation in the asm part. |
Beta Was this translation helpful? Give feedback.
-
Hey Arnaud(@synopse),
I'm getting an error:
I installed the
tar.gzfrom Release v2.2.How can I solve this?
Cheers,
Gus
Beta Was this translation helpful? Give feedback.
All reactions