Re: [Gems-users] Necessity of NP state

["Mike Marty" <mikem@xxxxxxxxxxx>]

For a M replacement, the cache block is deallocated and the data is moved into a TBE via i_allocateTBE.  The PUTX message sends dirty data.

 

When transition(M_W, WB_Ack, I) occurs, the writeback operation completes because the TBE is deallocated.  Since nothing is available to hold a cache state (no cache block, no TBE entry), the result is actually NP instead of I.  See setState()--> even though the transition indicates the resulting state is I, there is nothing to hold the new state.

 

--Mike

 

----- Original Message -----

From: Juyoung Jung

To: gems-users@xxxxxxxxxxx

Sent: Friday, June 22, 2007 2:31 PM

Subject: Re: [Gems-users] Necessity of NP state

Mike,


Your explanation is so helpful for me. Thanks a lot, Mike.
By the way, I have another quick question of cache replacement operation.
Looking into MESI_SCMP_tiledL2bank_directory- L1cache.sm file, I saw that cache state will change to Invalid state when any cache line in shared(S) or exclusive(M) state encounters L1_Replacement event. However, cache replacement operation should entail getting missed line, and thus I guess the updated line should eventually be marked as non invalid state such as S or M. Please look at the below code from L1 cache controller sm file.

For example, for a case of line replacement in M state, state transition will be "Exclusive state --> transient state waiting for ack --> Invalid state" according to
transition(M, L1_Replacement, M_I) {
     i_allocateTBE;
     g_issuePUTX;   // send data, but hold in case forwarded request (but clean data not sent)
     ff_deallocateL1CacheBlock;
   }

transition(M_I, WB_Ack, I) {
     s_deallocateTBE;
     o_popIncomingResponseQueue;
   }

transition(S, L1_Replacement, I) {
     gs_issuePUTS;
     ff_deallocateL1CacheBlock;
   }


When the modified cache line to replace receives acknowledgement for write-back completion, it terminates its replacement operation by marking cache line as Invalid.
I think further operation may be needed to get new block after receiving WB_Ack; Shouldn't lower-level memory fill the missed line with the required data? (This situation is also shown from cache replacement in S state.)
Does Ruby decouple cache line replacement from processor reference such Load and Store? If I have wrong thought, please disabuse me.
Thank you.




Regards,

Young

---

_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/" to your search.