new (?) performance profiling tool #6297

New Issue

flux · 2024-02-15T21:15:44Z

flux commented

2024-02-15 21:15:44 +00:00

i swear someone sent me a flame graph before, but i've apparently never used this tool myself:

https://content.minetest.net/packages/jwmhjwmh/jitprofiler/

i'm currently using it locally to try to analyze the performance of spawnit and other things.

i swear someone sent me a flame graph before, but i've apparently never used this tool myself: https://content.minetest.net/packages/jwmhjwmh/jitprofiler/ i'm currently using it locally to try to analyze the performance of spawnit and other things. ![image](/attachments/a61b4cc3-f8ca-4ee1-92f9-d8be435f1811) ![image](/attachments/97b8bc9f-9642-4ede-8e10-c44a3f5f9882)

image.png

80 KiB

image.png

90 KiB

flux added the

1. kind/enhancement

label 2024-02-15 21:15:44 +00:00

whosit commented

2024-02-15 21:38:59 +00:00

Oh, I liked to use it years before for analyzing perf record dumps, but somehow it got broken when I tried it again recently :/
I think format changed? I should try it again maybe.

But if it works for luajit, then it's good too - those visuals are nice :)

IIRC those are svg files and they have labels with full function names, etc. Maybe even interactive?

Oh, I liked to use it years before for analyzing `perf record` dumps, but somehow it got broken when I tried it again recently :/ I think format changed? I should try it again maybe. But if it works for luajit, then it's good too - those visuals are nice :) IIRC those are svg files and they have labels with full function names, etc. Maybe even interactive?

AliasAlreadyTaken commented

2024-02-16 03:04:44 +00:00

I tried it as well on testserver, but when I used it on main, it somehow killed the server. There is a warning on the repo that mesecon might have unforeseen consequences, maybe that's what happened?

flux commented

2024-02-16 17:29:36 +00:00

There is a warning on the repo that mesecon might have unforeseen consequences, maybe that's what happened?

hm. it might be that the jit profiler crashes when there's a pcall? that unfortunately makes this totally useless, as i wanted to run it on main, but we can't disable luacontrollers or there'd be a riot.

> There is a warning on the repo that mesecon might have unforeseen consequences, maybe that's what happened? hm. it might be that the jit profiler crashes when there's a `pcall`? that unfortunately makes this totally useless, as i wanted to run it on main, but we can't disable luacontrollers or there'd be a riot.

flux commented

2024-02-16 17:42:00 +00:00

hm. it might be that the jit profiler crashes when there's a pcall?

that doesn't seem to be it, the following runs just fine locally:

local profile = require("jit.profile")

local outfile = io.open("profile.out", "w")
local function record(thread, samples, vmstate)
    outfile:write(profile.dumpstack(thread, "pF;", -100), vmstate, " ", samples, "\n")
end
profile.start("vfi1", record)

local function do_stuff()
    local i = 0
    for j = 1, 10000000 do
        i = i + j
    end
    return i
end

for _ = 1, 20 do
    local ok, ret = pcall(do_stuff)
    if ok then
        print(ret)
    else
        print(error)
    end
end
profile.stop()
outfile:close()

print("done")

> hm. it might be that the jit profiler crashes when there's a `pcall`? that doesn't seem to be it, the following runs just fine locally: ```lua local profile = require("jit.profile") local outfile = io.open("profile.out", "w") local function record(thread, samples, vmstate) outfile:write(profile.dumpstack(thread, "pF;", -100), vmstate, " ", samples, "\n") end profile.start("vfi1", record) local function do_stuff() local i = 0 for j = 1, 10000000 do i = i + j end return i end for _ = 1, 20 do local ok, ret = pcall(do_stuff) if ok then print(ret) else print(error) end end profile.stop() outfile:close() print("done") ```

whosit commented

2024-02-16 18:30:44 +00:00

do we have a stack trace? error? any clue about how it crashed?

AliasAlreadyTaken commented

2024-02-17 05:39:47 +00:00

I don't remember exactly, but I believe the server just stopped responding and I had to restart it. There was no stacktrace nor any logging, even the output file was malformed.

We could try to recreate the experiment on the testserver, as far as I remember it happened there as well, but the crash took longer to surface

I don't remember exactly, but I believe the server just stopped responding and I had to restart it. There was no stacktrace nor any logging, even the output file was malformed. We could try to recreate the experiment on the testserver, as far as I remember it happened there as well, but the crash took longer to surface

flux commented

2024-02-18 05:49:12 +00:00

I don't remember exactly, but I believe the server just stopped responding and I had to restart it.

when i was running this tool locally, my server once stopped responding when i executed the "stop profiling" command. it certainly can cause hangs, i've got no ideas on why though.

> I don't remember exactly, but I believe the server just stopped responding and I had to restart it. when i was running this tool locally, my server once stopped responding when i executed the "stop profiling" command. it certainly can cause hangs, i've got no ideas on why though.

flux commented

2024-02-18 05:49:58 +00:00

this could probably be closed as "maybe" until the hang issue is better understood

AliasAlreadyTaken commented

2024-02-18 09:05:10 +00:00

Let's not close but investigate in this issue why the hangs happen. It's a useful tool and the whole of MT could profit if we find the cause.

flux commented

2024-02-18 19:37:39 +00:00

simple luac scripts are apparently not enough to trigger the bug (if it's even the luacs)

if event.type == "program" then
 interrupt(1)
elseif event.type == "interrupt" then
 port.b = not port.b
 interrupt(1)
end

simple luac scripts are apparently not enough to trigger the bug (if it's even the luacs) ```lua if event.type == "program" then interrupt(1) elseif event.type == "interrupt" then port.b = not port.b interrupt(1) end ```

flux commented

2024-02-18 19:54:30 +00:00

also, while i don't think the mod is the issue (probably it's either luajit itself or some way that minetest is using it), the developer of the mod is dead.

👀 1

Sign in to join this conversation.