Excise REPL and its dependencies from sysimg #51399
Conversation
vchuravy
added
REPL
|
Positive startup impact for But just |
function time_repl_startup(julia_cmd="./julia")
@async begin sleep(1); println("please press ^d") end
buf = IOBuffer()
t0 = time()
run(pipeline(`$julia_cmd --startup-file=no -ie 'atreplinit(_->println(time()))'`, buf))
s = String(take!(buf))
parse(Float64, strip(match(r"\n\n\d\.\d\d\d\d+e\d\n", s).match))-t0
end |
Another way suggested by Jameson iirc: just |
|
For me, "snappiness" refers to elapsed wall-clock time so reported user time might be misleading in the presence of CPU utilization other than 100% (either less or more). |
|
So looks like we spend time in codegen. |
|
So the extra-overhead is coming from |
vchuravy
force-pushed
the
vc/excise_repl
branch
from
d509367
to
d829a18
Compare
vchuravy
force-pushed
the
vc/excise_repl
branch
from
d829a18
to
f9ef95e
Compare
|
Would be great if folks try this out on different machines and see if they think the latency is now okay. |
|
I tried this PR out. Julia's REPL itself does launch perceptably slower (perhaps 100-150 ms slower, but don't know how to measure it), but once the banner appears, the REPL is responsive immediately with no latency. |
vchuravy
force-pushed
the
vc/excise_repl
branch
from
f9ef95e
to
75014c5
Compare
|
So... it turns out my recommendation above underreports time till the "julia>" prompt appears because the Results: Scriptfrom subprocess import Popen, PIPE, STDOUT
from time import time, sleep
import statistics
def repl_startup_time(exe):
sleep(.3)
t0 = time()
p = Popen(exe + " --startup-file=no", stdout = PIPE, stderr = STDOUT, shell = True)
str = []
end = "julia>"
while True:
out = p.stdout.read(1)
if out == b'':
raise RuntimeError("Unexpected end")
else:
str.append(out.decode("utf-8"))
if len(str) > len(end) and ''.join(str[-len(end):]) == end:
break
# print(''.join(str))
t1 = time()
p.kill()
sleep(.3)
return t1 - t0
def no_repl_startup_time(exe):
sleep(.3)
t0 = time()
p = Popen(exe + " --startup-file=no -e 'println(sqrt(2))'", stdout = PIPE, stderr = STDOUT, shell = True)
p.wait()
# print(p.stdout.read())
t1 = time()
sleep(.3)
return t1 - t0
def process(fun, exe, n):
data = [fun(exe) for _ in range(n)]
mn, med, mean = min(data), statistics.median(data), statistics.mean(data)
return "{:.0f}/{:.0f}/{:.0f}".format(1000*mn, 1000*med, 1000*mean)
print("system info:")
p = Popen("julia --startup-file=no -e 'using InteractiveUtils; versioninfo()'", shell = True)
p.wait()
print("master:")
p = Popen("julia/julia --startup-file=no -e 'Base.banner(short=true)'", shell = True)
p.wait()
print("pr:")
p = Popen("julia2/julia --startup-file=no -e 'Base.banner(short=true)'", shell = True)
p.wait()
process(repl_startup_time, "julia/julia", 3)
process(repl_startup_time, "julia2/julia", 3)
process(no_repl_startup_time, "julia/julia", 3)
process(no_repl_startup_time, "julia2/julia", 3)
print(" min/med/mean (ms)")
print("master repl: " + process(repl_startup_time, "julia/julia", 30))
print("pr repl: " + process(repl_startup_time, "julia2/julia", 30))
print("master no repl: " + process(no_repl_startup_time, "julia/julia", 30))
print("pr no repl: " + process(no_repl_startup_time, "julia2/julia", 30))According to my measurements, this PR is great for startup times. It would be nice to include this sort of end to end startup time benchmarking in BaseBenchmarks.jl. |
I don't quite trust this... It feels slower, but maybe its more that I have a debug build on? |
|
I also don't trust it. Testing this out interactively in VS code and MacOS Terminal revealed some issues VS Code MacOS Terminal We should make sure CI catches these issues and also fix them. |
|
One way to measure latency with REPL is to take a screen recording... |
|
What is your environment? This looks like you are missing |
|
@LilithHafner I made the debug message more useful. |
|
I usually record a gif and count the frames until I see the julia repl prompt. |
https://discourse.julialang.org/t/10-15-minute-ttfp-with-plots-jl-please-help/92636/53?u=lilith
Original source: JuliaLang/Pkg.jl#2669 (comment) Also if you could do `JULIA_DEBUG=loading`.
This fixes it, but my startup file is currently fine on master, so I think this is a regression of some sort. I'll try to distill my startup file to a MWE. |
|
It's particularly weird that we can load Pkg, but not REPL. Pkg depends on REPL. |
|
The culprit in my startup.jl file seems to be |
|
A test case that covers this for me is running at the command line. Does that test case fail on your machine too? Regardless, I'm not sure how to do this within Julia's testing framework. |
|
Okay loading Revise breaks the |
vchuravy
force-pushed
the
vc/excise_repl
branch
from
c7c81f4
to
e8fc317
Compare
vchuravy
changed the title
giordano
changed the title
| # @which is undefined | ||
| @test_broken v == ("false\nREPL: InteractiveUtilstrue\n", true) |
There was a problem hiding this comment.
This has introduced a distracting/confusing error show into CI
This reverts commit 2defa57.
Needs #51189 for proper precompilation of Pkg.
Moves the precompilation of the REPL into it's own pkgimg.
Similar to #51350, but without moving any code out of Base.