According to the TAs at the review session, while we have buffered transmission of a packet in cut through flow control, we still lock all the nodes on the path of the packet from the source to the destination. Therefore, we cannot have interleaving of flits from different packets. A disadvantage as a result of this is that we could possibly have some links in the route idle even though there might be a packet waiting to be sent using the idle link.
This comment was marked helpful 0 times.
analysiser
I think here the cut-through flow control implies packet level flow control where the packet will be forwarded as soon as outport is free, but if the outport is not free then the whole packet has to be stored (absorbed). That's the reason for the buffer has to stay large which is similar to store-and-forward flow control.
This comment was marked helpful 0 times.
ruoyul
So if both store-and-foward and cut-through flow control buffers the entire packet in each router, is the only difference in how soon the data is forward? i.e store-and-foward send only when the entire packet arrives while cut-through start sending as soon as header arrives?
This comment was marked helpful 0 times.
spilledmilk
Question: Why would cut-through flow control be setup such that all nodes on the path of the packet from source to destination are locked? I suppose this would prevent head-of-line blocking, but it seems that the penalty on throughput in this scheme would greatly outweigh those benefits.
@ruoyul: Pretty much. One key consequence is that because cut-through flow control begins forwarding as soon as the header arrives, the validity of the packet being forwarded is not checked until the very end, whereas in store-and-forward, the validity of the packet is checked at each node.
This comment was marked helpful 0 times.
mchoquet
@ruoyul I think the idea is that because cut-through flow control can start forwarding as soon as it gets data, it probably won't have to buffer the whole packet at each node. The buffers have to be allocated for the worst case, but that's not what usually happens.
This comment was marked helpful 0 times.
nrchu
@spilledmilk, if it wasn't locked, then it's possible you would not be able to complete sending the data if the buffer isn't large enough. Imagine that you have a buffer of size of 1024 bytes and packets A and B, although the packets in this particular case are only 768~ each. The buffer fills up with packet A, then says oh I have space in my buffer let me transfer packet B. But for whatever reason packet A is blocked from going to the next node, so the buffer doesn't empty, and we only can transfer up to half of packet B. Information is going to get dropped (corrupted packet) OR we have to store a lot more metadata (recreate header for the remaining data of packet B) OR have a lot more communication regarding packet size and size of buffer etc etc.
It is a lot simpler to simply lock the path. With buffer sizes as large as the maximum packet size you do not need to worry about these edge cases and the overhead associated with solving them.
According to the TAs at the review session, while we have buffered transmission of a packet in cut through flow control, we still lock all the nodes on the path of the packet from the source to the destination. Therefore, we cannot have interleaving of flits from different packets. A disadvantage as a result of this is that we could possibly have some links in the route idle even though there might be a packet waiting to be sent using the idle link.
This comment was marked helpful 0 times.
I think here the cut-through flow control implies packet level flow control where the packet will be forwarded as soon as outport is free, but if the outport is not free then the whole packet has to be stored (absorbed). That's the reason for the buffer has to stay large which is similar to store-and-forward flow control.
This comment was marked helpful 0 times.
So if both store-and-foward and cut-through flow control buffers the entire packet in each router, is the only difference in how soon the data is forward? i.e store-and-foward send only when the entire packet arrives while cut-through start sending as soon as header arrives?
This comment was marked helpful 0 times.
Question: Why would cut-through flow control be setup such that all nodes on the path of the packet from source to destination are locked? I suppose this would prevent head-of-line blocking, but it seems that the penalty on throughput in this scheme would greatly outweigh those benefits.
@ruoyul: Pretty much. One key consequence is that because cut-through flow control begins forwarding as soon as the header arrives, the validity of the packet being forwarded is not checked until the very end, whereas in store-and-forward, the validity of the packet is checked at each node.
This comment was marked helpful 0 times.
@ruoyul I think the idea is that because cut-through flow control can start forwarding as soon as it gets data, it probably won't have to buffer the whole packet at each node. The buffers have to be allocated for the worst case, but that's not what usually happens.
This comment was marked helpful 0 times.
@spilledmilk, if it wasn't locked, then it's possible you would not be able to complete sending the data if the buffer isn't large enough. Imagine that you have a buffer of size of 1024 bytes and packets A and B, although the packets in this particular case are only 768~ each. The buffer fills up with packet A, then says oh I have space in my buffer let me transfer packet B. But for whatever reason packet A is blocked from going to the next node, so the buffer doesn't empty, and we only can transfer up to half of packet B. Information is going to get dropped (corrupted packet) OR we have to store a lot more metadata (recreate header for the remaining data of packet B) OR have a lot more communication regarding packet size and size of buffer etc etc.
It is a lot simpler to simply lock the path. With buffer sizes as large as the maximum packet size you do not need to worry about these edge cases and the overhead associated with solving them.
This comment was marked helpful 0 times.