add precompile runs for xparse and xparse2 #108

rafaqz · 2022-02-03T23:15:07Z

This reduces ttfx of CSV.jl by 12 seconds in these benchmarks:

Current timing with main branches of Parsers.jl and CSV.jl:

julia> @time using CSV
  3.086746 seconds (6.63 M allocations: 375.584 MiB, 7.03% gc time, 88.55% compilation time)

julia> const input="""
       time,ping,label
       1,25.7,x
       2,31.8,y
       """
"time,ping,label\n1,25.7,x\n2,31.8,y\n"

julia> io = IOBuffer(input)
IOBuffer(data=UInt8[...], readable=true, writable=false, seekable=true, append=false, size=34, maxsize=Inf, ptr=1, mark=-1)

julia> @time file = CSV.File(io)
 17.450644 seconds (48.24 M allocations: 1.982 GiB, 6.00% gc time, 99.98% compilation time)
2-element CSV.File:
 CSV.Row: (time = 1, ping = 25.7, label = "x")
 CSV.Row: (time = 2, ping = 31.8, label = "y")

With this PR and main of CSV.jl:

julia> @time using CSV
  3.027161 seconds (6.59 M allocations: 385.136 MiB, 7.53% gc time, 90.56% compilation time)

julia> const input="""
       time,ping,label
       1,25.7,x
       2,31.8,y
       """
"time,ping,label\n1,25.7,x\n2,31.8,y\n"

julia> io = IOBuffer(input)
IOBuffer(data=UInt8[...], readable=true, writable=false, seekable=true, append=false, size=34, maxsize=Inf, ptr=1, mark=-1)

julia> @time file = CSV.File(io)
  5.894205 seconds (12.28 M allocations: 656.183 MiB, 4.75% gc time, 99.94% compilation time)
2-element CSV.File:
 CSV.Row: (time = 1, ping = 25.7, label = "x")
 CSV.Row: (time = 2, ping = 31.8, label = "y")

With the PR at CSV.jl:

julia> @time using CSV
  0.871332 seconds (2.21 M allocations: 136.379 MiB, 5.89% gc time, 78.47% compilation time)

julia> const input="""
       time,ping,label
       1,25.7,x
       2,31.8,y
       """
"time,ping,label\n1,25.7,x\n2,31.8,y\n"

julia> io = IOBuffer(input)
IOBuffer(data=UInt8[...], readable=true, writable=false, seekable=true, append=false, size=34, maxsize=Inf, ptr=1, mark=-1)

julia> @time file = CSV.File(io)
  5.757817 seconds (14.37 M allocations: 780.821 MiB, 4.84% gc time, 99.97% compilation time)
2-element CSV.File:
 CSV.Row: (time = 1, ping = 25.7, label = "x")
 CSV.Row: (time = 2, ping = 31.8, label = "y")

rafaqz · 2022-02-05T12:02:26Z

Added more precompile for xparse2 and tryparse.

This shaves multiple seconds off TTFX for packages depending on JSON.jl, like Blink.jl, Interact.jl, etc. But precompile in JSON.jl still has an effect although its mostly precompiling Parsers.jl JuliaIO/JSON.jl#337 So there could be more to add here.

Probably this PR should be quite thorough. Parsers.jl has 2200 dependents, so I'm guessing it's probably responsible for a significant fraction of TTFX for the whole julia ecosystem.

We could also rewrite some of the long unstable methods that are causing this compile time. The abstract typed fields of the objects here may contribute.

codecov · 2022-02-06T16:08:44Z

Codecov Report

Merging #108 (29f9985) into main (8bd7f84) will decrease coverage by 0.79%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##             main     #108      +/-   ##
==========================================
- Coverage   87.57%   86.77%   -0.80%     
==========================================
  Files           9        9              
  Lines        2309     2321      +12     
==========================================
- Hits         2022     2014       -8     
- Misses        287      307      +20

Impacted Files	Coverage Δ
src/precompile.jl	`11.76% <0.00%> (+11.76%)`	⬆️
src/bools.jl	`97.61% <0.00%> (-1.59%)`	⬇️
src/utils.jl	`89.28% <0.00%> (-1.54%)`	⬇️
src/floats.jl	`91.52% <0.00%> (-0.88%)`	⬇️
src/Parsers.jl	`93.87% <0.00%> (-0.38%)`	⬇️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8bd7f84...29f9985. Read the comment docs.

ufechner7 · 2022-02-07T23:48:14Z

So we need a test that calls the function precompile() to make codecov happy ...

oscardssmith · 2022-02-07T23:51:36Z

we could also just accept the code coverage regression...

rafaqz · 2022-02-08T10:13:07Z

Maybe it will look better if I add it. It is making sure all those types at least run...

Edit: added _precompile_() to the tests

src/precompile.jl

bkamins · 2022-02-08T14:51:26Z

src/precompile.jl

+    pos = 1
+    val = "a"
+    len = length(val)
+    for T in (Char, String), buf in (codeunits(val), Vector(codeunits(val)))


should also InlineStrings.jl precompilation be included here?

I wouldn't think they could be included here since they're in a separate package (InlineStrings.jl); but we could do similar precompilation over there.

The dependency goes the other way. Should we add a similar precompile block over there?

I don't actually get much improvement with precompilation in InlineStrings.jl. Precompiling for all types only takes half a second there.

What do you mean by "Precompiling for all types only takes half a second there"? Precompiling all inlinestring types? I wouldn't be too surprised by that since the heavy lifting would be handled by precompiling String in the Parsers.jl package.

Yes, all the InlineString types. Without String in Parsers.jl precompilation adds under a second to InlineStrings.jl precompilation.

It seems there is something about the numerical methods in Parsers.jl that takes a lot longer.

rafaqz · 2022-02-09T09:34:22Z

This should be ready to go. Can we change the actions settings here so that CI runs for anyone except first-time users?

src/precompile.jl

nickrobinson251 · 2022-02-09T10:41:32Z

src/precompile.jl

+            if !(T === String)
+                Parsers.xparse(T, buf, pos, len, options, T)
+            end
+            Parsers.xparse(T, buf, pos, len, options, Any)


why is this method important to precompile? is it because it is used in CSV (as _parseany)?

Yep for _parseany.

quinnj

Thanks @rafaqz!

oscardssmith · 2022-02-11T01:00:22Z

I think this might have broken CSV.

julia> using CSV
[ Info: Precompiling CSV [336ed68f-0bac-5ca0-87d4-7b16caf5d00b]
ERROR: LoadError: ccall method definition: argument 1 type doesn't correspond to a C type
Stacktrace:
 [1] top-level scope
   @ ~/.julia/packages/SentinelArrays/p1IoM/src/SentinelArrays.jl:209
 [2] include
   @ ./Base.jl:421 [inlined]
 [3] include_package_for_output(pkg::Base.PkgId, input::String, depot_path::Vector{String}, dl_load_path::Vector{String}, load_path::Vector{String}, concrete_deps::Vector{Pair{Base.PkgId, UInt64}}, source::String)
   @ Base ./loading.jl:1399
 [4] top-level scope
   @ stdin:1
in expression starting at /home/oscardssmith/.julia/packages/SentinelArrays/p1IoM/src/SentinelArrays.jl:1
in expression starting at stdin:

giordano · 2022-02-11T01:08:50Z

JuliaData/SentinelArrays.jl#69

rafaqz · 2022-02-11T09:42:11Z

The ccall isn't actually from this PR, it was there before. Maybe an issue would be better to track this.

quinnj · 2022-03-06T04:17:24Z

src/precompile.jl

+    pos = 1
+    val = "123"
+    len = length(val)
+    for T in (String, Int32, Int64, Float64, BigFloat, Dates.Date, Dates.DateTime, Bool)


@rafaqz (and as fyi @nickrobinson251 ), no idea why, but including the Dates.Date and Dates.DateTime precompiles in Parsers.jl is causing JuliaData/CSV.jl#981, i.e. precompiling CSV.jl on windows compleletly fails. Going to revert those two type precompiles for now and we can figure out what to do afterwards.

can you only disable them for Windows?

Its pretty weird that that happens. We could also rewrite a few of these methods to reduce compilation overheads, which may fix the problem.

CarloLucibello · 2022-04-07T05:59:50Z

It would be great to have this back again, it was a big quality of life improvement

quinnj · 2022-04-07T15:07:38Z

We have to figure out why it causes segfaults on windows with Julia 1.7 first during precompilation.

rafaqz · 2022-04-07T15:38:07Z

We could just rewrite those methods to reduce compilation time and see if that fixes it.

add precompile statements for xparse

cd34cd9

rafaqz mentioned this pull request Feb 3, 2022

improve time to first csv load JuliaData/CSV.jl#975

Closed

rafaqz changed the title ~~add precompile statements for xparse~~ add precompile runs for xparse Feb 4, 2022

rafaqz mentioned this pull request Feb 4, 2022

add basic precompilation JuliaIO/JSON.jl#337

Merged

rafaqz changed the title ~~add precompile runs for xparse~~ add precompile runs for xparse and xparse2 Feb 5, 2022

rafaqz force-pushed the precompile branch 2 times, most recently from 55a50de to ed2540a Compare February 5, 2022 12:40

oscardssmith approved these changes Feb 6, 2022

View reviewed changes

rafaqz force-pushed the precompile branch 2 times, most recently from 5cbe60d to 9595a55 Compare February 8, 2022 13:54

KristofferC reviewed Feb 8, 2022

View reviewed changes

src/precompile.jl Outdated Show resolved Hide resolved

KristofferC reviewed Feb 8, 2022

View reviewed changes

src/precompile.jl Outdated Show resolved Hide resolved

bkamins reviewed Feb 8, 2022

View reviewed changes

rafaqz added 2 commits February 8, 2022 15:56

add precompilation for xparse2

16997ea

use nested for loops

fb3b561

rafaqz force-pushed the precompile branch from 9595a55 to fb3b561 Compare February 8, 2022 14:56

reduce type list and use tryparse instead of xparse2

290c897

rafaqz force-pushed the precompile branch from a66a8b6 to 290c897 Compare February 8, 2022 20:14

nickrobinson251 reviewed Feb 9, 2022

View reviewed changes

remove 6 arg typed xparse precompilation

29f9985

quinnj approved these changes Feb 10, 2022

View reviewed changes

quinnj merged commit e7fe3e7 into JuliaData:main Feb 11, 2022

rafaqz deleted the precompile branch February 11, 2022 09:36

quinnj reviewed Mar 6, 2022

View reviewed changes

quinnj mentioned this pull request Mar 28, 2022

Revert precompilation #110

Merged

rafaqz mentioned this pull request Apr 21, 2022

Resolve precompilation issues #113

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add precompile runs for xparse and xparse2 #108

add precompile runs for xparse and xparse2 #108

rafaqz commented Feb 3, 2022 •

edited

rafaqz commented Feb 5, 2022 •

edited

codecov bot commented Feb 6, 2022 •

edited

ufechner7 commented Feb 7, 2022

oscardssmith commented Feb 7, 2022

rafaqz commented Feb 8, 2022 •

edited

bkamins Feb 8, 2022

quinnj Feb 8, 2022

rafaqz Feb 8, 2022

rafaqz Feb 8, 2022

quinnj Feb 8, 2022

rafaqz Feb 8, 2022 •

edited

rafaqz commented Feb 9, 2022

nickrobinson251 Feb 9, 2022

rafaqz Feb 9, 2022

quinnj left a comment

oscardssmith commented Feb 11, 2022

giordano commented Feb 11, 2022

rafaqz commented Feb 11, 2022

quinnj Mar 6, 2022

oscardssmith Mar 6, 2022

rafaqz Mar 6, 2022

CarloLucibello commented Apr 7, 2022

quinnj commented Apr 7, 2022

rafaqz commented Apr 7, 2022 •

edited

add precompile runs for xparse and xparse2 #108

add precompile runs for xparse and xparse2 #108

Conversation

rafaqz commented Feb 3, 2022 • edited

rafaqz commented Feb 5, 2022 • edited

codecov bot commented Feb 6, 2022 • edited

Codecov Report

ufechner7 commented Feb 7, 2022

oscardssmith commented Feb 7, 2022

rafaqz commented Feb 8, 2022 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rafaqz Feb 8, 2022 • edited

Choose a reason for hiding this comment

rafaqz commented Feb 9, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

quinnj left a comment

Choose a reason for hiding this comment

oscardssmith commented Feb 11, 2022

giordano commented Feb 11, 2022

rafaqz commented Feb 11, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

CarloLucibello commented Apr 7, 2022

quinnj commented Apr 7, 2022

rafaqz commented Apr 7, 2022 • edited

rafaqz commented Feb 3, 2022 •

edited

rafaqz commented Feb 5, 2022 •

edited

codecov bot commented Feb 6, 2022 •

edited

rafaqz commented Feb 8, 2022 •

edited

rafaqz Feb 8, 2022 •

edited

rafaqz commented Apr 7, 2022 •

edited