Re: [Gems-users] Deadlock avoidance in Gems


Date: Tue, 03 Apr 2007 14:45:00 -0400
From: Niket Agarwal <niketa@xxxxxxxxxxxxx>
Subject: Re: [Gems-users] Deadlock avoidance in Gems
Thanks Hongxia for the reply.

I looked into the code and figured out that definitely there are virtual networks present. What I gathered is: - At the Coherence protocol level, different Message Buffers are assigned different virtual network numbers. Thus the messages which get into these buffers, will pass through only those particular virtual network.

- At the network level, each switch picks from a particular Message buffer (and hence knows the vnet) and outputs the message into the same vnet of an outlink. Thus, as you had pointed out, a particular message does not change it's virtual network during its whole life.

Please could you confirm that I am understanding it correctly.

I have one issue though. Suppose, a particular memory request is made by a processor and its a cache miss and simics is stalled now. The request messages flow around into the network and when they are served, simics should be woken up. Where is this being done ?

Thanks for the help.
Niket.
[← Prev in Thread] Current Thread [Next in Thread→]