Re: [m5-users] Inducing non-determinism into multiprocessor setups in M5

[Ali Saidi]

Actually in our tutorial as ASPLOS 2008 we showed an example of adding a variance to delay memory. Here is a diff to m5 that should do the trick. Just set latency_var = 'XXns' on the Physical Memory object and that should do the trick. It will still be deterministic for a given latency since the random number generator is seeded to the same thing (which is good so you can redo an experiment when you find the problem). Changing the latency_var parameter by a little bit is probably the best way to see a wide variety of results.

Ali

 

randomize_mem.diff
Description: Binary data

On Apr 9, 2008, at 4:46 PM, Steve Reinhardt wrote:

Have you read the paper by Alameldeen and Wood on this? (There may have been other coauthors.) They introduced variability by making the memory latency pseudo-random.

Steve

2008/4/9 Geoffrey Blake <[EMAIL PROTECTED]>:





Thanks,

Geoff


No virus found in this outgoing message.
Checked by AVG.


_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

I'm potentially looking for some ways to induce some non-determinism into multiprocessor setups for M5, since currently M5 is completely deterministic. Does anyone have any suggestions where I should start looking to be able to do this? I'd like to measure the variance for threaded programs over many runs. I'm thinking the best place to induce non-determinism is in bus.cc, or to even perhaps randomly shuffle the event queue for events that need to launch on the same tick. Version: 7.5.519 / Virus Database: 269.22.10/1367 - Release Date: 4/9/2008 7:10 AM

_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users