Meng-Ju et al.,

Following your suggestion I worked worth Meng and was able to, with the 
author's help, able to get it working.

In the meantime I have developed doubts on your suggested direction.

Problem restatement: in the early 1990s authors showed that by having 
some information on the location of a task's footprint in cache the 
scheduler could make better decisions reducing the cache reload 
transient.

What I wonder if by using threads I am undermining the results. 
Speculatively, I could have four identical cores, three polluters and 
one core that (pollutes then loads; loads then reloads; etc). A problem 
is that this isn't a full context switch so I suspect that the 
footprint shall not be very discrete and thus harder to show.

It seems feasible that this could be accomplished by running the system 
in FS, however, that could incur quite large penalties to run.

I (and my advisers) have investigated a sundry of ways to make this 
work, such as: checkpointing and trying various restarts; and, the 
method you suggested.

The checkpointing seems most able to show what we want but from my 
reading of the code it isn't going to work without some modification.

The least pleasant alternative is to abuse the code and actually do 
context switches then have the simulator be the Omni-scheduler.

Thoughts?




In some respects I need an Omnipotent/ Omniscient scheduler.


   


Meng-Ju Wu wrote:


>Hi William,
>
>In Meng's patches, you can change the code to schedule/deschedule the
>threads as you want. It can give you a very precise control on thread
>scheduling. If you want to pollute the L2,  you can just spawn two
>threads on two different cores and suspend one. Then you can
>deschedule the first thread and active the second one.
>
>Meng-Ju
>_______________________________________________
>m5-users mailing list
>[email protected]
>http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
>

-----------------------------------
Will Beazley|Sys. Software Analyst
409.880.7847|[email protected]
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

Reply via email to