Re: Why 3 NoCs to avoid deadlock?

[Jonathan Balkind]

Hi Marcelo,

As you say, it is protocol-level deadlock that necessitates the three NoCs (with strict priority). The caches themselves also have to be designed accordingly to avoid protocol-level deadlock.

Thanks,

Jon

On Mon, 26 Apr 2021 at 09:37, Marcelo Ruaro <mcelo...@gmail.com> wrote:

Hello,

First, congratulation for this impressive many-core framework. I was reading the papers and documentations and a simple question arrived.

My question comes from sentences like this from docs and papers:

• "While the cache coherence protocol is designed to be logically deadlock free, it also depends on the physical layer and routing to also be deadlock free."
  
• "Packets are routed using dimension-ordered wormhole routing to ensure a deadlock-free network"
• "Coherent messages between L1.5 caches and L2 caches communicate through three NoCs, carefully designed to ensure deadlock-free operation

My question: If the cache is deadlock free and one NoC is deadlock free too, what justifies the adoption of three NoC instead one to avoid deadlock? I am imagining that without the three NoCs  the cache coherence protocol will enter in deadlock, this is right?

Thank you in advance,

Best,

Marcelo

--
You received this message because you are subscribed to the Google Groups "OpenPiton Discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpiton+...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/openpiton/8337cded-1ad3-4692-bdd0-dc3cdbfa268dn%40googlegroups.com.