I don't know of anyone who's done it, but it sounds cool.  My vote is
for option #3.  Option #1 seems really invasive and would make it
tricky to deal with wrong-path code or anything else that might be
non-deterministic.  Option #2 doesn't scale past N=2.  Option #3 might
not be too bad if you interleave things on a page-by-page basis; e.g.,
for an N-thread workload, physical page p of thread i maps to physical
page N*p+i of the new system.  This would also give you reasonable
conflict behavior in a physically indexed cache (another issue that
might backfire on you in options 1 or 2... particularly in option 1 if
all the threads have overlapping fake physical addresses).

Steve

On Mon, Jun 9, 2008 at 6:24 PM, Rick Strong <[EMAIL PROTECTED]> wrote:
> Hi all,
>
> I am going to add a system that would allow opening two single threaded
> checkpoints in SE mode and run them on a detailed SMT simulation. A few
> questions to make sure no one has implemented this already:
>
> 1) Does anyone already have this support and is willing to share?
>
> Deciding between three options:
>
> 1) Add to the checkpoint all memory reads for the simulation interval of
> interest and then redirect all M5 memory reads to use this file to
> satisfy requests instead of its simulated physical memory and to make
> sure that memory reads for different threads go to their respective files.
>
> 2) Each thread could be checkpointed two times with a memory allocator
> that allocates all memory in either the beginning or end of physical
> memory. That way when two threads are started, clever selection of the
> two checkpoint files will make sure that physical memory address do not
> collide. The negative of this situation is that the number of
> checkpoints necessary will double, but this isn't too bad.
>
> 3) The third option is extending option 2 to just use a single
> allocator, but make a script which dynamically creates new checkpoint
> files from two single threaded checkpoints, and re-assigns
> virtual-to-physical memory mappings for the phymem file and TLB section.
>
> If anyone has a good alternative or convincing argument why they would
> choose one alternative, pipe it on up and let me know.
>
> Thanks in advance,
> -Rick
> _______________________________________________
> m5-users mailing list
> [email protected]
> http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
>
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

Reply via email to