[luatex] Identifying why LuaTeX 2023 is slower than LuaTeX 2022

[Hans Hagen]

On 8/30/2023 9:44 AM, Max Chernoff wrote:

> None of these patches should have any performance implications, yet
> "sys" is ~40% slower than either "2023" or "2023-initial", which makes
> me think that the LuaTeX currently in TL23 ("sys") was erroneously
> compiled with "-O0" or something.

i'm  bit puzzled now:

so the tl 2023 bin is slower but when you compile fresh the performenac 
is the same? if so ... then why waste time on it if generating a new bin 
solves the problem?

i can indeed imagine different compile flags being used, although i 
wonder if that will give a 50% performance gain .. i did observe small 
differences between gcc versions and a few options but we're talking low 
percentages here

>> Is the same
>> fontloader used for every test (which assume stability over 2017 - 2023)?
> 
> Yes.
> 
>> Is the luatex build.sh used or some tex live one?
> 
> I used the LuaTeX "build.sh" for all the "20xx" binaries, albeit with a
> few tiny patches. Very detailed build instructions here:
> 
>     https://tug.org/~mseven/luatex.html#binary-details
> 
> "2023-initial" would have been built with the standard TL build script.
> "sys" was probably built with the TL script, but it's a special ad-hoc
> release so something different might have happened.

i'm not sure if there was a reason for that (afaik texlive goes for a 
more conservative older compilation just to make sure it works on older 
os versions)

>     Benchmark 3: PATH=/tmp/texlive-testing/2023-initial/bin/x86_64-linux:/bin/ luatex -ini factorial.tex
>       Time (mean ± σ):      3.958 s ±  0.039 s    [User: 3.933 s, System: 0.024 s]
>       Range (min … max):    3.908 s …  4.031 s    10 runs
>      
>     Benchmark 4: PATH=/tmp/texlive-testing/sys/bin/x86_64-linux:/bin/ luatex -ini factorial.tex
>       Time (mean ± σ):      5.519 s ±  0.054 s    [User: 5.489 s, System: 0.029 s]
>       Range (min … max):    5.464 s …  5.589 s    10 runs
> 
> All 4 tests use the exact same TL23 texmf trees, so the TeX code, Lua
> code, and all the other binaries are identical; only the "luahbtex" and
> "luatex" binaries are different. The only difference between "2023-
> initial" and "sys" binaries should be the socket and debug/popen
> patches; they (probably) used the same compilers and the source should
> be otherwise identical.

These are as you mentioned irrelevant there.

> Yet the current TL23 LuaTeX binary ("sys") is 50% slower than the
> initially-released TL23 LuaTeX binary ("2023-initial"), so something
> weird is definitely going on. My guess would be that the current LuaTeX
> binaries were compiled with -O0 while all the other binaries use -O3, or
> something similar. Just a guess though.

Factorial does little (not spread all over token space). Some mem access 
for registers, a little amount of macro tokens that likely sit in the 
cpu cache. Plus making a macro that gets larger body every iteration (so 
that is actually the bottleneck as it involved copying tokens). As you 
start ini tokens are not scattered that much.

So, do you see the same 50 % drop with the current luatex when you 
compile without O3 ?

One thing I can imagine that there is less inlining applied for lua end 
but older compilers (doesn't tl still use gcc 7) were not that 
aggressive in that anyway. (fwiw, link time optimization at most gains a 
few % on tex but that's even more recent.)

In factorial the bottleneck might be likely in \the which (because we're 
an utf engine) has a couple of calls that might benefit from inlining, 
but imo unlikely to make the 50% drop.

Hans


-----------------------------------------------------------------
                                           Hans Hagen | PRAGMA ADE
               Ridderstraat 27 | 8061 GH Hasselt | The Netherlands
        tel: 038 477 53 69 | www.pragma-ade.nl | www.pragma-pod.nl
-----------------------------------------------------------------