"If head flit blocks, rest of packet stops"
In case of a long packet, where head flit can be multiple hops ahead of the source, how does the source know that head flit was blocked and, hence, stop transmission?
Is there some kind of ack which is sent by switch, once a flit exits from switch?
Firephinx
For long messages, latency is almost entirely independent of network distance because the latency would depend on the throughput of the network and how long the message is. This is because each flit would be moving through the network one after the other so after the initial latency of the head flit due to the network distance, the rest of the latency depends on the number of flits in the packet and the throughput of the network to move the flits from the source to the destination.
muchanon
Wormhole control provides significantly better latency than store-and-forward because in wormhole flow control, the flits do not need to recombine until they all reach the destination, while with store-and-forward, each part of the packet must take the exact same path and recombine at each router on that path.
lya
What if the links are not equally long? How do the flits do synchronization with each other?
nishadg
If one just waits for an opening in the next, then naturally the latency will be bounded by the longest link.
"If head flit blocks, rest of packet stops" In case of a long packet, where head flit can be multiple hops ahead of the source, how does the source know that head flit was blocked and, hence, stop transmission? Is there some kind of ack which is sent by switch, once a flit exits from switch?
For long messages, latency is almost entirely independent of network distance because the latency would depend on the throughput of the network and how long the message is. This is because each flit would be moving through the network one after the other so after the initial latency of the head flit due to the network distance, the rest of the latency depends on the number of flits in the packet and the throughput of the network to move the flits from the source to the destination.
Wormhole control provides significantly better latency than store-and-forward because in wormhole flow control, the flits do not need to recombine until they all reach the destination, while with store-and-forward, each part of the packet must take the exact same path and recombine at each router on that path.
What if the links are not equally long? How do the flits do synchronization with each other?
If one just waits for an opening in the next, then naturally the latency will be bounded by the longest link.