How easy is it for a system to switch between point-to-point and broadcast communication or revert to a bit vector representation? Wouldn't switching between the different methods and having to initialize them each time you switch have a huge time penalty, especially if the switch occurs often (kind of like thrashing)?
This comment was marked helpful 0 times.
bwasti
@asinha
There won't be a time penalty because you won't need to reinitialize anything. You could potentially have bit that represents the protocol you will be using. In this case when you go to use the directory scheme you check this bit and determine if you should use a broadcast or attempt to only communicate with the processors needed.
This comment was marked helpful 0 times.
vrkrishn
I think there might be some overhead still even if you can simply switch to the broadcasting method. For example, the directory control system still needs to be updated even if the system is using broadcasting at the current step in the event that the shared count drops below the the broadcasting switch threshold.
This comment was marked helpful 0 times.
jmnash
How is it determined which of these fallback methods is implemented? Is one better than the others depending on the machine? It seems like it would depend on the applications, for example on how many nodes the application uses and how far over the max number of sharers the application will go.
However, it seems like a waste if we go one node over the max number of sharers and it instantly switches to broadcasting. Like I said above, it probably depends on the application, but to me the best fallback method seems like the third one. It seems like it would be relatively easy to implement a system where if we have max+1 sharers, each bit would represent 2 nodes, and then as soon as we have 2*max+1 sharers, each bit would represent 3 nodes, etc.
How easy is it for a system to switch between point-to-point and broadcast communication or revert to a bit vector representation? Wouldn't switching between the different methods and having to initialize them each time you switch have a huge time penalty, especially if the switch occurs often (kind of like thrashing)?
This comment was marked helpful 0 times.
@asinha There won't be a time penalty because you won't need to reinitialize anything. You could potentially have bit that represents the protocol you will be using. In this case when you go to use the directory scheme you check this bit and determine if you should use a broadcast or attempt to only communicate with the processors needed.
This comment was marked helpful 0 times.
I think there might be some overhead still even if you can simply switch to the broadcasting method. For example, the directory control system still needs to be updated even if the system is using broadcasting at the current step in the event that the shared count drops below the the broadcasting switch threshold.
This comment was marked helpful 0 times.
How is it determined which of these fallback methods is implemented? Is one better than the others depending on the machine? It seems like it would depend on the applications, for example on how many nodes the application uses and how far over the max number of sharers the application will go.
However, it seems like a waste if we go one node over the max number of sharers and it instantly switches to broadcasting. Like I said above, it probably depends on the application, but to me the best fallback method seems like the third one. It seems like it would be relatively easy to implement a system where if we have max+1 sharers, each bit would represent 2 nodes, and then as soon as we have 2*max+1 sharers, each bit would represent 3 nodes, etc.
This comment was marked helpful 0 times.