Re: [Gems-users] Backoff code in LogTM

[Konstantinos Nikas <knikas@xxxxxxxxxxxxxxxxx>]

Mohammad Momin Ansari wrote
> Thanks Kostis,
>
> Would it be correct to suggest that moving the backoff code in  
> TransactionInterfaceManager.c:restartTransactionCallback() so that it  
> applies not only to XACT_LAZY_VM, and removing the backoff at the end of  
> transaction.c:tm_log_unroll() would result in roughly equivalent  
> functionality, i.e. wouldn't break anything?
> Thanks,
>
>   
I would say that the only difference is the actual number of cycles of 
the backoff, as the one in TrasactionInterfaceManager stalls the 
processor for a number of cycles while the one in the software handler 
"spins" the processor inside a (random) loop.
In terms of functionality, I don't see any real difference and how 
moving the backoff call inside the interface manager would break the 
system. On the other hand, I am sure I am not familiar with all the 
little details of LogTM, so you should wait for someone else's 
confirmation :-) .
Kind regards,

Kostis

--
Dr. Konstantinos Nikas
Computing Systems Laboratory
School of Electrical and Computer Engineering
National Technical University of Athens

Tel: +30-210-7722495
e-mail: knikas@xxxxxxxxxxxxxxxxx
http://www.cslab.ece.ntua.gr/~knikas