by joaofl » Tue Jul 21, 2015 10:15 pm
Dear Olofsson, thank you for your reply.
I took some time to take a better look into the topic you suggested (about the arbitrations), and the patent of yours. I got some, and I'll probably get many more of the details I need there.
But what I still miss is more details about the protocol itself. For example, an overall perspective of what I'm trying to reproduce on my simulator:
Like in the "all-to-one" epiphany example in my parallella, In a write operation from all nodes to node 0,0, what happens in the network such that the total throughput measured is X?
In more details...:
The c-mesh does 64 bits atomic write operations, in 1.5 clock cycles per hop. If you want to write some several words, how would the packets look like? What is the protocol overhead?
Is the packet as simple as [ 12 bits for destination + 32 memory address (word aligned?) + 20 bits of data ] ?
Following packets look the same? If not, does the first one establish some virtual channel? Is it blocking?
The manual says that the r-mesh performs one read operation every 8 clock cycles. Is the read operation an 8 bytes atomic transaction as well?
Does It takes 8 clock cycles because the bandwidth is smaller, or because it has some sort of time division due some limited resource?
Does the packet look like: [ 12 bits for destination + 12 origin + 32 memory address (word aligned?) + 8 offset ] ?
Does it uses wormholing? What are the flit sizes? How does the flow control works, and the time it consumes? If not, how does it do queuing?
Sorry if these look like a silly question, but I believe the more details I understand, the closer the simulator can get to simulate the real architecture network performance. With that, better metrics can be taken and understood.
One good thing about ns3 is that it uses c++, and I could potentially use the same c files from an epiphany example, and run it on virtual nodes. I would be able to insulate processing time, and analyse the impact of communication (alone) in the overall performance.
Thanks again for your attention.
Best,
João