search

liyonghelpme / bwapi

Automatically exported from code.google.com/p/bwapi
0 stars 0 forks source link

(Starcraft) Memory Leak for unknown Dialog graphic object #322

Closed GoogleCodeExporter closed 9 years ago

GoogleCodeExporter commented 9 years ago 
What steps will reproduce the problem?
1. Trying to run a Evolutionary Computation on my bot (running in client mode 
with CMD). 
2. Using the BWAPI.ini file to restart game (auto_menu = SINGLE_PLAYER and 
auto_restart = ON)
3.After around 4500 games I get an error from StarCraft saying: This 
application has encountered a critical error: Not enough storage is available 
to process this command.

What is the expected output? What do you see instead?
I exspected that StarCraft would continue running like my Bot. In CMD it says 
"reconnecting", so this part is still running. Bot StarCraft crashes with an 
error (see enclosed files).
I'm not sure if this is BWAPI or StarCraft that has a problem. My computer has 
over 4GB HD storage left, so it is not my HD that is out of storage. I'm not 
sure about if there is a memory leak somewhere.

What version of the product are you using? On what operating system?
BWAPI 3.3Beta
bwapi-mono-bridge 3.3a
Windows 7 x64
Starcraft 1.16.1

Please provide any additional information below.

Original issue reported on code.google.com by twsandb...@gmail.com on 27 Nov 2010 at 8:12

Attachments:

GoogleCodeExporter commented 9 years ago 
Pretty sure this is because many objects are created but never destroyed.
I think this is mostly Units and Bullets.

Should probably store every created object in a vector that never gets 
overwritten and delete them onGameEnd. (like a cumulative object array)

I'm going to create a map that illustrates the issue.

Original comment by AHeinerm on 27 Nov 2010 at 3:46

GoogleCodeExporter commented 9 years ago 
I might be wrong, but I didn't see where those objects were destroyed.

Original comment by AHeinerm on 27 Nov 2010 at 4:06

GoogleCodeExporter commented 9 years ago 
Yeah it must be something else, I've created and killed thousands of units but 
there isn't much of a change.

Also when it mentions it has run out of memory, it's not referring to HD 
storage, but RAM.

Original comment by AHeinerm on 27 Nov 2010 at 5:52

GoogleCodeExporter commented 9 years ago 
Yeah it make perfect sence that it is a memory problem, because when I ran it 
again I could see that StarCraft continued using more and more RAM and ended up 
with using just over 400 MB, started with just over 100 MB. I just couldn't see 
in the error message anything about memory other than storage, but I guess it 
wouldn't have made the memory dump if it has been an error on my HD?

Hope you find a solution to what is not deleted correctly.

Original comment by twsandb...@gmail.com on 28 Nov 2010 at 8:19

Attachments:

GoogleCodeExporter commented 9 years ago 
I forgot to say that I'm using the build in Reset method 
BWAPI.bwapi.Broodwar.restartGame() in the onEnd event and just before the 
restart I call Sleep(1000). I'm not sure if it is nessesary to sleep it, but I 
saw you did that for the map  GoliathsZerglings.scm on the forum:  
http://www.broodwarai.com/forums/index.php?s=c80bc889964ef0bb501c3486c36c7063&sh
owtopic=815.
Bot my bot crashed, when using the map, so I just thought that is was easier to 
just use the build-in restart game method, than find out the bug that made it 
crash on the map.

Original comment by twsandb...@gmail.com on 28 Nov 2010 at 8:32

GoogleCodeExporter commented 9 years ago 
No, a Sleep in the bot is not necessary, but is in the map so that the bot has 
time to recognize that it has won/lost and perform any updates or calculations 
using the remaining units, rather than having them immediately destroyed and 
not having access to them anymore.

Original comment by AHeinerm on 28 Nov 2010 at 4:00

GoogleCodeExporter commented 9 years ago 
I'm leaning towards some kind of client/server issue. My Broodwar isn't 
exceeding 56 MB over the course of several games (using the restart() method as 
well).

Original comment by AHeinerm on 28 Nov 2010 at 4:48

GoogleCodeExporter commented 9 years ago 
Tried again and ran for much longer time than last time and crashed with the 
use of 800+ MB RAM. Do you think the problem is in the BWAPI client/server or 
between the Mono bridge and BWAPI?

Original comment by twsandb...@gmail.com on 28 Nov 2010 at 6:43

GoogleCodeExporter commented 9 years ago 
This time I got this error file.

Original comment by twsandb...@gmail.com on 28 Nov 2010 at 6:45

Attachments:

GoogleCodeExporter commented 9 years ago 
It's the same error, Storm allocating memory for the standard.SNP module. It's 
accumulating memory in Starcraft, so the problem should be in BWAPI and/or any 
loaded modules.

Original comment by AHeinerm on 29 Nov 2010 at 2:38

GoogleCodeExporter commented 9 years ago 
I found and fix some memory leaks in r3299 and r3300. I don't know if this 
fixed the bug completely or if there is still something else causing a memory 
leak.

Original comment by lowerlo...@gmail.com on 7 Dec 2010 at 11:25

GoogleCodeExporter commented 9 years ago 
Thats great but I can unfortunately not test it before there is a new version 
of BWAPI and the bwapi-mono-bridge has been updated to match the version.
I don't know how to do all the SWIG thing that is needed to talk with BWAPI.

Original comment by twsandb...@gmail.com on 8 Dec 2010 at 9:19

GoogleCodeExporter commented 9 years ago 
Developers:
mono-bridge: http://code.google.com/p/bwapi-mono-bridge/
Currently at 3.3, waiting for update to 3.4 to verify if this issue should 
remain.

Original comment by AHeinerm on 6 Jan 2011 at 12:39

GoogleCodeExporter commented 9 years ago 
If possible, can you simplify your bot, log the game count, and create a port 
of the simplified bot to C++ so that we can identify if the problem is in BWAPI 
or bwapi-mono-bridge (or both)?

Original comment by AHeinerm on 6 Jan 2011 at 12:42

GoogleCodeExporter commented 9 years ago 
I have tried to run it using Mono command prompt instead of the standard 
Windows command prompt. I haven't had any brake downs yet, after using Mono 
CMD. I'm not sure if it is just a coincidence.

Original comment by twsandb...@gmail.com on 7 Jan 2011 at 8:57

GoogleCodeExporter commented 9 years ago 
Unfortunately it was only a coincidence. I have now also had a crash using Mono 
CMD

Original comment by twsandb...@gmail.com on 9 Jan 2011 at 10:46

GoogleCodeExporter commented 9 years ago 
Hi, there. I am having the same problem in Windows XP. 
I am training a Reinforcement learning for unit micro management (3vs3, 6vs6 
scenarios)
Game count is different, 60,000 6v6, 100,000 for 3v3, 30,000 for 6v6. 
The computer that I am wokring on has 2GB of Ram. Though I am using the version 
before the final release until now.

This does not happen often in my own laptop, it has windows 7 and 3gb of ram 
(probably becaause of more RAM)

My error points to another file though:

File: Starcrat\SWAR\lang\net_record.cpp.
Line:690

What I mainly use in my code are std vectors to store frame information, action 
rewards and Q values for my algorithm. And I reset them after each round. 

I will try to see if I can find anything suspicous.

Original comment by shantia....@gmail.com on 11 Jan 2011 at 8:45

GoogleCodeExporter commented 9 years ago 
@shantia.amirhossein Do you also use bwapi-mono-bridge? Knowing the type of 
module and any bwapi-related dependencies are most important. What version of 
BWAPI are you referring to?

The error reports are useless for this issue. It also seems to be 
OS-independant as well. This is definitely a Starcraft, BWAPI, or client/bridge 
issue.

Original comment by AHeinerm on 12 Jan 2011 at 3:40

GoogleCodeExporter commented 9 years ago 
No, I directly use the c++ version of BWAPI. I was using version 3.3, now I am 
testing with 3.4. I will report the results. I found a leak in my program 
already so I am testing again.

Original comment by shantia....@gmail.com on 17 Jan 2011 at 1:49

GoogleCodeExporter commented 9 years ago 
ok I ran a small scenario with 2vs2 zealots and ran it for 30 seconds. It took 
about 1 minute for the memory leak debugger to stop writing on output.

I am using
#define _CRTDBG_MAP_ALLOC
#include <stdlib.h>
#include <crtdbg.h>

and putting _CrtDumpMemoryLeaks(); in the onEnd() function.

I am doing it in debug mode but it does not tell me any program lines. 
I still can see some variable of my code here like:
{507782} normal block at 0x077982B8, 32 bytes long.
 Data: <unit_grid_target> 75 6E 69 74 5F 67 72 69 64 5F 74 61 72 67 65 74 

but I also see strange stuff like:
{507585} normal block at 0x04185310, 20 bytes long.
 Data: < R   S  pS      > B0 52 18 04 D0 53 18 04 70 53 18 04 03 00 00 00 

or some more meaningful stuff like:
{507218} normal block at 0x077824E0, 32 bytes long.
 Data: <UNIT_NOT_VISIBLE> 55 4E 49 54 5F 4E 4F 54 5F 56 49 53 49 42 4C 45 
and
{507122} normal block at 0x07781970, 32 bytes long.
 Data: <SUNKEN_COLONY_TE> 53 55 4E 4B 45 4E 5F 43 4F 4C 4F 4E 59 5F 54 45 

I will try get rid of those lines from my code, but what are the other ones? Is 
it because that the AImodule is a part of the Dll and maybe the function should 
be called at somewhere else?

I am not very experienced with it so all suggestions are welcome.

Original comment by shantia....@gmail.com on 17 Jan 2011 at 5:45

GoogleCodeExporter commented 9 years ago 
Correction:
unit_grid_target is not a variable, is a part of a string that I print using 
sprintf_s function to initials my neural network (flann)

Original comment by shantia....@gmail.com on 17 Jan 2011 at 6:14

GoogleCodeExporter commented 9 years ago 
The memory leak bug is still there. I have updated to BWAPI mono-bridge 3.4.
I ran a evolutionary algorithm and 600 games were played before StarCraft 
crashed. (StarCraft has stopped working...)

StarCraft Memory use were: 60.816 K, my client AI 16.896 K.

My AI client was still running after StarCraft crashed. But there is no error 
messages.

What version of the product are you using? On what operating system?
BWAPI 3.4Beta
bwapi-mono-bridge 3.4a
Windows 7 x64
Starcraft 1.16.1

Original comment by twsandb...@gmail.com on 20 Jan 2011 at 10:47

GoogleCodeExporter commented 9 years ago 
Thanks for the responses.
I added the leak dumping in r3418 for BWAPI DEBUG mode, as well as fixed a 
memory leak that probably isn't related to the issue.

I am also getting those "leaks" mentioned in Comment 20, most of them are 
strings, but there are occasional data segments, and of course these ones won't 
output the file either.

Maybe we should breakpoint when the memory usage starts to become insane and 
iterate all memory blocks to see where exactly something is happening. Storm 
allocations (inside Starcraft, and some used in BWAPI) will have a unique 
header containing the file and line number, and if we can find out which header 
appears the most then we've identified our leak. If there is no header then at 
least we know the leak is not in Starcraft or any of its related modules.

I am going to try testing again to see if I can reproduce it and perform the 
above.

Original comment by AHeinerm on 21 Jan 2011 at 2:29

GoogleCodeExporter commented 9 years ago 
Just want to mention that I can reproduce the issue just fine.

Original comment by AHeinerm on 21 Jan 2011 at 3:27

GoogleCodeExporter commented 9 years ago 
Figured it out. THIS IS A STARCRAFT ISSUE, so if you didn't use bwapi and just 
sat in single player and went to "Restart Game" a few thousand times, you would 
encounter this same error. The reason why this hasn't surfaced is because 
nobody has played a few thousand games without exiting the game once (and 
reported it).

I will verify these findings in a test that will log all SMem handles that 
havn't been freed.

Once I've verified this and any other possible Starcraft leaks, a simple patch 
in BWAPI will fix them.

Original comment by AHeinerm on 21 Jan 2011 at 4:54

GoogleCodeExporter commented 9 years ago 
Oh yeah, the object I was referring to is located here:
0x004197B0 DlgGrp_Constructor

And more specifically (the repeated handles I noticed in the debugger), here: 
0x004EE2D0 LoadRaceUI

Original comment by AHeinerm on 21 Jan 2011 at 4:56

GoogleCodeExporter commented 9 years ago 
Using the hack I added, the memory leak is confirmed inside of Starcraft.

Starcraft\SWAR\lang\game.cpp:183 - dlgs\protoss.grp
.?AU_DLGGRP@@:-2 - dlgs\protoss.grp
Starcraft\SWAR\lang\game.cpp:183 - dlgs\protoss.grp
.?AU_DLGGRP@@:-2 - dlgs\protoss.grp
Starcraft\SWAR\lang\game.cpp:183 - dlgs\protoss.grp
.?AU_DLGGRP@@:-2 - dlgs\protoss.grp
...

These lines will repeat for the number of games.

Original comment by AHeinerm on 21 Jan 2011 at 6:01

GoogleCodeExporter commented 9 years ago 
thx for the findings :)
We are desperately waiting for the patch :) If there is a way to somehow exit 
the StarCraft and rerun it with chaos luncher via an external program 
automatically, we can survive till the patch comes out.
Cheers,

Original comment by shantia....@gmail.com on 21 Jan 2011 at 8:15

GoogleCodeExporter commented 9 years ago 
If it really is the memory leak that causes StarCraft to crash after a random 
number of restarts. How can it be that there is much difference in how many 
games restart that can be made before the game crashes.

This time I could run around 5500 games (last time around 600 games) before 
StarCraft crashed. Ram consumption was 184,044 K at the crash this time.

Original comment by twsandb...@gmail.com on 21 Jan 2011 at 2:36

GoogleCodeExporter commented 9 years ago 
Forgot to say that there was around 1 GB unused RAM left, when the game crashed 
(78 % physical memory used).

Original comment by twsandb...@gmail.com on 21 Jan 2011 at 2:39

GoogleCodeExporter commented 9 years ago 
For me it is not very different. Depending on the number of units I use in 
scenario, it scales up and down. But my SC crashes usually around 1GB of memory 
use. In a 6v6 scenario it is around 60,000 restarts (though I do the restarts 
by calling the function inside my .cpp file and I am not using Auto_Restart 
thingi).

Original comment by shantia....@gmail.com on 21 Jan 2011 at 2:41

GoogleCodeExporter commented 9 years ago 
I also use the build in game restart from code in a 1vs1 SSC map, with 9 
goliaths against 21 hydralisks. How many days do you spent on running 60,000 
games. I think that I can only run about 10,000 games a day.

Original comment by twsandb...@gmail.com on 21 Jan 2011 at 3:10

GoogleCodeExporter commented 9 years ago 
Well, a 6v6 scenario is quiet fast, the most thing that takes time is the 
update of my neural network (I use FANN). But around 8 hours is enough for 
60,000 games in a 6v6 scenario for marines or zealots.

Original comment by shantia....@gmail.com on 21 Jan 2011 at 3:34

GoogleCodeExporter commented 9 years ago 
You can restart the game almost immediately after it starts to test the leak. A 
full game does not need to be run. As for crashing with memory remaining, it 
happens to me as well.

Original comment by AHeinerm on 21 Jan 2011 at 4:07

GoogleCodeExporter commented 9 years ago 
Fixed in r3422.

I think it crashes sooner because it doesn't use upper bounds of memory 
(highest is 0x7FFFFFFF).

Original comment by AHeinerm on 21 Jan 2011 at 6:53

GoogleCodeExporter commented 9 years ago 
Also it seems to follow a very high block alignment rule. The allocation is 
probably limited by Storm.

Original comment by AHeinerm on 21 Jan 2011 at 6:54

GoogleCodeExporter commented 9 years ago

Original comment by AHeinerm on 21 Jan 2011 at 6:55

GoogleCodeExporter commented 9 years ago 
[deleted comment]
GoogleCodeExporter commented 9 years ago 
Thanks and Cheers

Original comment by shantia....@gmail.com on 21 Jan 2011 at 8:47

GoogleCodeExporter commented 9 years ago 
I still use gui, turning it off does not change the results very much for
me, but I am not using the mono bridge, I used that once 1 year ago and it
was sooooo slow and incomplete back then. I also run in window mode. It is
quite fast indeed, with your units at least you should be able to do 20,000
runs in 8 hours.

Original comment by shantia....@gmail.com on 21 Jan 2011 at 8:53