Re: [Gems-users] Deadlock avoidance in GEMS

[Niket Agarwal <niketa@xxxxxxxxxxxxx>]

Thanks Hongxia and Greg for the replies.

Hongxia, I still want to confirm one thing.

An example L1 controller has this:-
> machine(L1Cache, "MSI Directory L1 Cache CMP") {
>
>   // NODE L1 CACHE
>   // From this node's L1 cache TO the network
>   // a local L1 -> this L2 bank, currently ordered with directory 
> forwarded requests
>   MessageBuffer requestFromL1Cache, network="To", virtual_network="1", 
> ordered="false";
>   // a local L1 -> this L2 bank
>   MessageBuffer responseFromL1Cache, network="To", 
> virtual_network="3", ordered="false";
>   MessageBuffer unblockFromL1Cache, network="To", virtual_network="4", 
> ordered="false";
>   MessageBuffer messageToPrivateL2, network="To", virtual_network="0", 
> ordered="true";
>
>   MessageBuffer messageFromPrivateL2, network="From", 
> virtual_network="0", ordered="true";
>   MessageBuffer requestToL1Cache, network="From", virtual_network="2", 
> ordered="false";
>   // a L2 bank -> this L1
>   MessageBuffer responseToL1Cache, network="From", 
> virtual_network="3", ordered="false";
>  
I guess here, a virtual net is assigned to message buffers like 
requestFromL1Cache assigned virtual_network="1".
Now the PerfectSwitch::wakeup() is written such that it picks a message 
from a particular "vnet" of an incoming link and then finds the outlink 
and forwards the message into the same vnet of the outlink.
I hope, this is how protocol level deadlocks are avoided. Please confirm 
this.
One more question I have. How is it determined, which message buffers 
should be ordered and which ones unordered?
Thanks for your answers,
Niket.