Cool, that is reproducible on Linux, digging it further.

yours,
anton.

On Mon, Nov 23, 2009 at 3:21 PM, Anton Muhin <[email protected]> wrote:
> Thanks a lot, I can now reproduce it on my box.
>
> My immediate hypothesis would be it's due to virtual address space
> fragmentation: notice that v8 dies after performing 10 mark sweeps
> which fail to reclaim enough of garbage (last ones, going from 125.1
> -> 125.0, the last one 125.0 -> 125.0)---that should be enough to
> exhaust process memory space.  But it's only a hypothesis.  I'd check
> if it's reproducible under Linux and update.
>
> yours,
> anton.
>
> On Sat, Nov 21, 2009 at 12:38 AM, dracflamloc
> <[email protected]> wrote:
>>
>> Sorry for so many replies in a row, but I've isolated the code that
>> crashes v8. You can test this with the normal d8.exe and it crashes at
>> the same point my custom exe does.
>>
>> http://pastebin.com/m2f610e8
>>
>> On Nov 20, 4:29 pm, dracflamloc <[email protected]> wrote:
>>> Small update: I rebuilt the lib using the visual studio sln instead of
>>> scons, and now the splay test runs fine using my executable. However
>>> my js still crashes. Can someone take a look at the cpp code and let
>>> me know if theres something I'm missing? Thanks
>>>
>>> On Nov 20, 3:53 pm, dracflamloc <[email protected]> wrote:
>>>
>>>
>>>
>>> > What do you mean by allocate a 'thing' in between?
>>>
>>> >http://pastebin.com/m10af9629
>>>
>>> > Here is the cpp code for my exe.
>>>
>>> > On Nov 20, 3:13 pm, Anton Muhin <[email protected]> wrote:
>>>
>>> > > Could you post the code?  If we could reproduce that under Windows, it
>>> > > might be easier to understand what goes on.
>>>
>>> > > How do you build v8, btw?
>>>
>>> > > I can run Erik's variant of splay just fine:
>>>
>>> > > d:\chromium\src\v8>shell.exe public-splay.txt --trace-g
>>> > > Scavenge 0.7 -> 0.6 MB, 2 ms.
>>> > > Scavenge 1.0 -> 1.0 MB, 2 ms.
>>> > > Scavenge 1.3 -> 1.3 MB, 2 ms.
>>> > > Scavenge 2.1 -> 2.0 MB, 5 ms.
>>> > > Scavenge 2.8 -> 2.8 MB, 5 ms.
>>> > > Mark-sweep 4.3 -> 4.2 MB, 20 ms.
>>> > > Scavenge 5.7 -> 5.6 MB, 7 ms.
>>> > > Mark-sweep 8.6 -> 8.5 MB, 38 ms.
>>> > > Mark-sweep 11.5 -> 11.3 MB, 45 ms.
>>> > > Scavenge 17.3 -> 17.1 MB, 31 ms.
>>> > > Mark-sweep 23.1 -> 22.8 MB, 99 ms.
>>> > > Mark-sweep 28.8 -> 28.5 MB, 119 ms.
>>> > > Scavenge 34.5 -> 34.2 MB, 34 ms.
>>> > > Mark-sweep 39.3 -> 39.1 MB, 156 ms.
>>> > > Scavenge 45.1 -> 44.8 MB, 35 ms.
>>> > > Mark-sweep 50.8 -> 50.5 MB, 191 ms.
>>> > > Scavenge 56.5 -> 56.3 MB, 34 ms.
>>> > > Scavenge 62.3 -> 62.0 MB, 33 ms.
>>> > > Mark-sweep 68.0 -> 67.7 MB, 249 ms.
>>> > > Scavenge 73.7 -> 73.4 MB, 34 ms.
>>> > > Scavenge 79.4 -> 79.2 MB, 35 ms.
>>> > > Scavenge 85.2 -> 84.9 MB, 35 ms.
>>> > > Mark-sweep 90.9 -> 90.6 MB, 323 ms.
>>> > > Scavenge 96.6 -> 96.4 MB, 36 ms.
>>> > > Scavenge 102.4 -> 102.1 MB, 36 ms.
>>> > > Scavenge 108.1 -> 107.8 MB, 36 ms.
>>> > > Scavenge 113.8 -> 113.5 MB, 37 ms.
>>> > > Mark-sweep 119.5 -> 119.3 MB, 371 ms.
>>> > > Scavenge 125.3 -> 125.0 MB, 37 ms.
>>> > > Scavenge 131.0 -> 130.7 MB, 37 ms.
>>> > > Scavenge 136.8 -> 134.5 MB, 28 ms.
>>> > > Scavenge 140.5 -> 134.5 MB, 16 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > Time (splay): 1954 us.
>>>
>>> > > (I am at trunk's 3339 now)
>>>
>>> > > Are you sure you do not allocate a thing in between?  (it could lead
>>> > > to fragmentation of address space and eventual OOM).
>>>
>>> > > yours,
>>> > > anton.
>>>
>>> > > On Fri, Nov 20, 2009 at 10:29 PM, dracflamloc
>>>
>>> > > <[email protected]> wrote:
>>>
>>> > > > Thanks for the help, btw =)
>>>
>>> > > > I ran your splay test and it crashes as well. here is the trace:
>>> > > > Scavenge 34.5 -> 34.2 MB, 43 ms.
>>> > > > Mark-sweep 39.2 -> 38.9 MB, 199 ms.
>>> > > > Scavenge 45.0 -> 44.7 MB, 41 ms.
>>> > > > Mark-sweep 50.7 -> 50.4 MB, 243 ms.
>>> > > > Scavenge 56.4 -> 56.1 MB, 41 ms.
>>> > > > Scavenge 62.1 -> 61.9 MB, 42 ms.
>>> > > > Mark-sweep 67.9 -> 67.6 MB, 316 ms.
>>> > > > Scavenge 73.6 -> 73.3 MB, 43 ms.
>>> > > > Scavenge 79.3 -> 79.1 MB, 42 ms.
>>> > > > Scavenge 85.1 -> 84.8 MB, 42 ms.
>>> > > > Mark-sweep 90.8 -> 90.5 MB, 409 ms.
>>> > > > Scavenge 96.5 -> 96.2 MB, 44 ms.
>>> > > > Scavenge 102.2 -> 102.0 MB, 43 ms.
>>> > > > Scavenge 108.0 -> 107.7 MB, 44 ms.
>>> > > > Scavenge 113.7 -> 113.4 MB, 46 ms.
>>> > > > Mark-sweep 119.4 -> 119.1 MB, 519 ms.
>>> > > > Scavenge 125.2 -> 124.9 MB, 44 ms.
>>> > > > Scavenge 130.9 -> 130.6 MB, 45 ms.
>>> > > > Scavenge 136.6 -> 134.5 MB, 35 ms.
>>> > > > Scavenge 140.5 -> 134.5 MB, 18 ms.
>>> > > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > > Scavenge 142.5 -> 134.5 MB, 6 ms.
>>> > > > Scavenge 142.5 -> 134.5 MB, 7 ms.
>>> > > > crash...
>>>
>>> > > > I could post the code, however I'm not just using the vanilla v8
>>> > > > engine. I've wrapped pdcurses functions and created my own. So my code
>>> > > > won't work in plain v8. However since even the splay test is crashing,
>>> > > > we already have a common code we can work from to try and figure this
>>> > > > out.
>>>
>>> > > > On Nov 20, 1:58 pm, Erik Corry <[email protected]> wrote:
>>> > > >> I attached the file.  I am running with the command line
>>>
>>> > > >> ./v8 public-splay.txt --trace-gc
>>>
>>> > > >> I am using the Linux build.
>>>
>>> > > >> 2009/11/20 dracflamloc <[email protected]>:
>>>
>>> > > >> > Perhaps I'm missing something here. Could you post the flags you're
>>> > > >> > passing in, if any?
>>> > > >> > I assume the 140 you're getting is from the gc trace output?
>>> > > >> > Are there any special build parameters you're using?
>>>
>>> > > >> > I don't get why its happening, but it is. Is there anything I 
>>> > > >> > could do
>>> > > >> > to investigate this better?
>>>
>>> > > >> If the JS file is something you can share then sending it to us would
>>> > > >> help us find out what is going on.
>>>
>>> > > >> > On Nov 20, 1:12 pm, Erik Corry <[email protected]> wrote:
>>> > > >> >> It seems to work for me.  I quadrupled the size of the splay 
>>> > > >> >> benchmark
>>> > > >> >> and I can run that without problems.  It tops out at over 
>>> > > >> >> 250Mbytes in
>>> > > >> >> the 64 bit version, around 140Mbytes in the 32 bit version (those 
>>> > > >> >> fat
>>> > > >> >> pointers take up space!).
>>>
>>> > > >> >> 2009/11/20 dracflamloc <[email protected]>:
>>>
>>> > > >> >> > trace shows it gets to about 125MB and then can no longer free 
>>> > > >> >> > memory
>>> > > >> >> > so it throws OOM.
>>> > > >> >> > I've attempted to use --max-old-space-size but always get the 
>>> > > >> >> > same
>>> > > >> >> > result
>>>
>>> > > >> >> > I have tried this with 3339 and still it caps out at 125 in the 
>>> > > >> >> > trace.
>>>
>>> > > >> >> > The thing is, I know where the memory is going (Lots of arrays),
>>> > > >> >> > however I don't know why its capping out at such a small RAM 
>>> > > >> >> > usage.
>>>
>>> > > >> >> > On Nov 20, 4:20 am, Erik Corry <[email protected]> wrote:
>>> > > >> >> >> 2009/11/20 dracflamloc <[email protected]>:
>>>
>>> > > >> >> >> > #
>>> > > >> >> >> > # Fatal error in (null)
>>> > > >> >> >> > # Allocation failed - process out of memory
>>> > > >> >> >> > #
>>>
>>> > > >> >> >> > Hello, I have v8 running in a VC++ 2005 console application. 
>>> > > >> >> >> > I'm
>>> > > >> >> >> > running a script thats pretty simple, but is memory 
>>> > > >> >> >> > intensive. However
>>> > > >> >> >> > the script always crashes after a few seconds while its 
>>> > > >> >> >> > creating
>>> > > >> >> >> > arrays/objects with an out of memory error.
>>>
>>> > > >> >> >> > This computer has 3GB of RAM. How can I enable a script to 
>>> > > >> >> >> > access all
>>> > > >> >> >> > available memory before throwing the OOM error, if it is 
>>> > > >> >> >> > indeed
>>> > > >> >> >> > cutting off at 150MB?
>>>
>>> > > >> >> >> We had a bug until recently where you could get out of memory 
>>> > > >> >> >> errors
>>> > > >> >> >> without actually having run out of memory.  make sure you have 
>>> > > >> >> >> a
>>> > > >> >> >> version at least as recent as 3285 on the bleeding edge or 
>>> > > >> >> >> 3333 on the
>>> > > >> >> >> trunk.
>>>
>>> > > >> >> >> You can run the VM with --trace-gc to see how much memory it 
>>> > > >> >> >> is actually using.
>>>
>>> > > >> >> >> There is an option --max-old-space-size for increasing the 
>>> > > >> >> >> amount of
>>> > > >> >> >> memory available to V8.
>>>
>>> > > >> >> >> I think the developer tools have some functions for tracing 
>>> > > >> >> >> where
>>> > > >> >> >> memory is going, but I am not personally familiar with them.
>>>
>>> > > >> >> >> --
>>> > > >> >> >> Erik Corry
>>>
>>> > > >>  public-splay.txt
>>> > > >> 14KViewDownload
>> >>
>>
>

--~--~---------~--~----~------------~-------~--~----~
v8-users mailing list
[email protected]
http://groups.google.com/group/v8-users
-~----------~----~----~----~------~----~------~--~---

Reply via email to