[Vm-dev] Question about memory allocation in VM

Mariano Martinez Peck marianopeck at gmail.com
Tue May 4 17:38:13 UTC 2010


Ok....a completely new world to explore for me. I was just thinking about
this but maybe I will have to "do it" in a couple of months. I will be
slowly reading and analyzing each of your comments and I will ask any
further question.

Thanks to everybody.

Mariano

On Tue, May 4, 2010 at 6:22 PM, John M McIntosh <
johnmci at smalltalkconsulting.com> wrote:

>
> http://isqueak.org/sqAllocateMemory
> http://isqueak.org/sqGrowMemoryBy
> http://isqueak.org/sqShrinkMemoryBy
> http://isqueak.org/sqMemoryExtraBytesLeft
>
>
> That said, there are some issues which can defeat your vision of only
> working with part of the oops space
>
> (a) I can't remember if this was removed, but at startup time we would
> iterate over the oops space confirming the setting of the root bit to
> fix a problem related to images created in mm what 1992?  This should not
> be a concern for closure based images since cleanup
> code at image save time fixed the issue and was introduced 10 or so years
> ago.
>
> (b) If you invoke allInstances it ends up iterating over the entire oops
> space, you'll find there are a dozen or so uses of that pattern in the image
> that are triggered via startup:
>
> (c) If your oops space start address isn't the same as it was when you
> saved the image, then the vm will iterate over all the oops readjusting the
> oops address to the start address offset.  Adjusting the platforms Virtual
> memory mapping system to place the oops space at 512MB helps avoid that.
>
> (d) The untuned GC will at some point invoke a full GC if you don't adjust
> it to bias against full GC activity.
>
> Given all that then for example for WikiServer for the iPhone where the
> operating system does not pre-load virtual memory mapped pages and where
> I've removed the desire to do allInstances and tuned the GC not to do a full
> GC at the drop of a hat. Then that results in a image that spans 1,932 4K
> pages (7,913,472), of which 806 pages (3,301,376) are swapped in from Flash,
> and 68 pages are modified to get the first wiki page visible.  This means
> 4,612,096 bytes are left in Flash, which on a iPhone 3G shaves 2 or 3
> seconds off the startup time.
>
>
> On 2010-05-04, at 7:56 AM, Andreas Raab wrote:
>
> > Look at memory mapping (mmap on unix, MapViewOfFile on Windows). What
> you're describing can be done simply by mapping and unmapping memory regions
> from an image file. I believe the iPhone does just that.
> >
> > Cheers,
> >  - Andreas
> >
> > On 5/4/2010 6:12 AM, Mariano Martinez Peck wrote:
> >> Sorry in advance if this was asked before...I have something in my mind
> >> that says it was :(
> >>
> >> I want to know how the memory is allocated between the VM and the OS.
> >> For example, I know (actually, I think I know) that in Mac OS VM when
> >> you run an image, the VM allocates (malloc() I guess) the same size at
> >> least, of the image. So...I couple of questions:
> >>
> >> 1) Is the same behavior for all OS ? If not, which happens with the
> rest?
> >>
> >> 2) Where is such behavior defined? I don't think it is in VMMaker but in
> >> the C support code of each VM. But where (.c and function anme) , any
> >> quotes?
> >>
> >> 3) In am working (actually yet thinking) a mechanism that will probably
> >> use much less memory than the image. The idea is to swap to disk unused
> >> objects. Then....the amount of memory I will use should be MUCH smaller
> >> than the size of the image. So my question is, which changes should I do
> >> to the VM so that it uses only THAT amount of memory and not the size of
> >> the image.
> >>
> >> As I said, suppose I have an image size (in disk) on 15MB. Suppose I run
> >> that image, and then, as I swapped a lot of objects to disk, the actual
> >> usage of memory is 3 MBs. So what I want is that OS only uses 3 MB and
> >> not 15. Can I do that ?
> >>
> >> Thanks a lot for answering to this newbie :)
> >>
> >> Mariano
>
> --
> ===========================================================================
> John M. McIntosh <johnmci at smalltalkconsulting.com>   Twitter:
>  squeaker68882
> Corporate Smalltalk Consulting Ltd.  http://www.smalltalkconsulting.com
> ===========================================================================
>
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20100504/7b18c290/attachment.htm


More information about the Vm-dev mailing list