Julia 1.11 is slower than 1.10?

ufechner7 · October 17, 2024, 11:30am

I see the same thing, only that my computer is so fast that with 1.11 I might have 0.2s delay where you report 1s delay.

And I would say that this is to be expected because the REPL is no longer part to the system image.

sijo · October 17, 2024, 11:37am

julia> versioninfo()
Julia Version 1.11.1
Commit 8f5b7ca12ad (2024-10-16 10:53 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 8 × Intel(R) Core(TM) i7-10510U CPU @ 1.80GHz
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, skylake)
Threads: 1 default, 0 interactive, 1 GC (on 8 virtual cores)

@ufechner7 show-off

Closer inspection suggests my delay is closer to 500-600ms, but that still feels sluggish.

roflmaostc · October 17, 2024, 11:41am

Not quite, both startup-file=no are fast with Julia 1.11.1 vs 1.10.4

└─(13:40:10)──> time julia-1.10.4 --startup-file="no" -e "exit()"                                                                                                             ──(Do,Okt17)─┘
julia-1.10.4 --startup-file="no" -e "exit()"  0.06s user 0.10s system 127% cpu 0.126 total
(base) ┌─(/tmp)──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────(felix@saturn:pts/8)─┐
└─(13:40:17)──> time julia-1.10.4 --startup-file="yes" -e "exit()"                                                                                                            ──(Do,Okt17)─┘
At startup Revise.jl and !OhMyREPL.jl loadedjulia-1.10.4 --startup-file="yes" -e "exit()"  0.34s user 0.83s system 343% cpu 0.342 total
(base) ┌─(/tmp)──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────(felix@saturn:pts/8)─┐
└─(13:40:19)──> time julia-1.11  --startup-file="no" -e "exit()"                                                                                                              ──(Do,Okt17)─┘
julia-1.11 --startup-file="no" -e "exit()"  0.07s user 0.06s system 130% cpu 0.105 total
(base) ┌─(/tmp)──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────(felix@saturn:pts/8)─┐
└─(13:40:31)──> time julia-1.11  --startup-file="yes" -e "exit()"                                                                                                             ──(Do,Okt17)─┘
At startup Revise.jl and !OhMyREPL.jl loadedjulia-1.11 --startup-file="yes" -e "exit()"  1.69s user 0.23s system 202% cpu 0.946 total

sijo · October 17, 2024, 11:46am

I don’t think that measures the interactive mode initialization. If I use -e "exit()" all versions are quite fast:

~ $ time julia +1.11.1 --startup-file=no -e "exit()"

________________________________________________________
Executed in  145.18 millis    fish           external
   usr time   84.10 millis    0.76 millis   83.34 millis
   sys time   63.36 millis    1.14 millis   62.22 millis

~ $ time julia +1.10 --startup-file=no -e "exit()"

________________________________________________________
Executed in  153.63 millis    fish           external
   usr time   92.41 millis    0.41 millis   91.99 millis
   sys time   65.85 millis    1.06 millis   64.79 millis

It’s only when I let Julia show the banner and REPL that it’s slow.

roflmaostc · October 17, 2024, 11:49am

Does that make a difference?

└─(13:49:12)──> time julia-1.11 --startup-file=no -e "using REPL; exit()"                                                                                                     ──(Do,Okt17)─┘
julia-1.11 --startup-file=no -e "using REPL; exit()"  1.00s user 0.10s system 398% cpu 0.274 total
(base) ┌─(~)─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────(felix@saturn:pts/8)─┐
└─(13:49:15)──> time julia --startup-file=no -e "using REPL; exit()"                                                                                                          ──(Do,Okt17)─┘
The latest version of Julia in the `release` channel is 1.11.1+0.x64.linux.gnu. You currently have `1.10.5+0.x64.linux.gnu` installed. Run:

  juliaup update

in your terminal shell to install Julia 1.11.1+0.x64.linux.gnu and update the `release` channel to that version.
julia --startup-file=no -e "using REPL; exit()"  0.07s user 0.10s system 129% cpu 0.134 total

ufechner7 · October 17, 2024, 11:59am

ufechner@framework:~$ time julia +1.11 --startup-file=no -e "using REPL; exit()"  

real	0m0,248s
user	0m0,742s
sys	0m0,076s
ufechner@framework:~$ time julia +1.10 --startup-file=no -e "using REPL; exit()"  

real	0m0,111s
user	0m0,051s
sys	0m0,070s

As expected. You can always create your own system image that contains the REPL if you don’t like this change. I am not getting nervous if I have to wait 0.25 s. Did you have used Matlab? There you probably have to wait 10s to get a prompt.

sijo · October 17, 2024, 12:05pm

Yep that accounts for most of the new latency:

~ $ time julia +1.10 --startup-file=no -e "using REPL; exit()"

________________________________________________________
Executed in  154.40 millis    fish           external
   usr time   98.02 millis   80.00 micros   97.94 millis
   sys time   62.90 millis  949.00 micros   61.95 millis


~ $ time julia +1.11.1 --startup-file=no -e "using REPL; exit()"

________________________________________________________
Executed in  401.48 millis    fish           external
   usr time  674.65 millis    8.00 micros  674.64 millis
   sys time   75.98 millis  923.00 micros   75.06 millis

Obviously not a deal breaker but I think it’s really too bad to lose that snappy feeling that Julia got in recent versions…

Palli · October 26, 2024, 6:38pm

I think my PR addresses this if backported to 1.11.2 (I thought I forgot print actually, but it’s covered by println that calls it, I believe), I just added precompile to it:

I’ve been working on eliminating allocations, and the last one is quite the rabbit hole to figure out and get rid of (I know possible, at least for single-threaded code)

I’ve already matched 1.10/fixed the 1.11 regression, and actually improved on 1.10/1.11, but getting even better to 0 allocations is elusive:

julia> @edit write(stdout, "1");

julia> s= "1"; io=stdout
Base.TTY(RawFD(15) open, 0 bytes waiting)

julia> @time write(stdout, "1");  # but what it calls below, basically, adds 1 allocation, not sure why:
1  0.000045 seconds (1 allocation: 16 bytes)


julia> @time GC.@preserve s unsafe_write(stdout, pointer(s), UInt(1));
1  0.000074 seconds (2 allocations: 32 bytes)

julia> test2() = (s = "1"; GC.@preserve s unsafe_write(stdout, pointer(s), UInt(1));)
test2 (generic function with 1 method)

julia> @time test2();
1  0.000053 seconds (2 allocations: 32 bytes)

It does actually print, not sure why segfaults right after (and thus can't count its supposed non-allocations):
julia> Base.uv_write(stdout, pointer(s), UInt(1));
1
[215864] signal (11.1): Segmentation fault

julia> @time (GC.@preserve s (Base.iolock_begin(); Base.uv_write(stdout, pointer(s), UInt(1)); Base.iolock_end()))
1
[238190] signal (11.1): Segmentation fault
in expression starting at REPL[2]:1
jl_mutex_unlock_nogc at /cache/build/builder-amdci4-4/julialang/julia-release-1-dot-10/src/julia_locks.h:85 [inlined]
...

That’s in 1.10.5. I actually get less info in 1.11.1, just:

julia> @time (GC.@preserve s (Base.iolock_begin(); Base.uv_write(stdout, pointer(s), UInt(1)); Base.iolock_end()))
1Segmentation fault (core dumped)

I’ve see take! in code (right before the other uv_write there), that I think is responsible for allocation (as opposed to the locking), and maybe _unsafe_take!(s)) is the solution like here?:

github.com

JuliaLang/julia/blob/97ecdb8595c4a1fbe68ba6f39b3244e8cdabc2c6/base/strings/io.jl#L153C5-L153C29


      
          String(_unsafe_take!(s))