On Tue, Jun 16, 2009 at 2:38 PM, Shoaib Akram <[email protected]> wrote:

> 1) On a multiprocessor system, when a upgradeReq hits the global bus, the
> owner sends a UpgradeResp and all sharers invalidate their copy. The
> invalidation by sharers and UpgradeResp happens as a single transaction?


Yes.


> 2) I observe some WriteReq packets on the global bus. on a Write miss ina
> last-level cache, does not the cache make a ReqReq packet and send it over
> the global bus. Are the WriteReq packets possibly from the I/O subsystem?


Yes.  One way to verify this is that the WriteReq packets should be for
smaller sizes (1-8 bytes) and not for full cache lines (64 bytes).


> 3) In the presence of UpgradeReq and UpgradeResp, why are the commands
> WriteInvalidateReq and WriteInvalidateResp needed in packt.hh?


They're intended to be used for doing write-throughs to coherent blocks.
The cache model doesn't currently implement that though so they're basically
unused in the current code base.

Steve


>
>
> Thanks.
>
> _______________________________________________
> 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