Relating to Problem 2A. from sample exam questions,
Question: how would the topology of an interconnect affect the performance of above computation?
This comment was marked helpful 0 times.
ycp
Basically the mesh interconnect system basically exactly fits this problem. Each processor would take the cell that matches with its position in the interconnect and then each processor will only have to communicate with other processors that are 1 link away from them.
This comment was marked helpful 0 times.
vrkrishn
Analyzing the different interconnect begins with the analysis that the computation on one data point on a processor will most likely require communication to processors that are surrounding on all cardinal directions.
In an interconnect system like the ring interconnect, the structure does not match the communication pattern of the program. We will have to wait 16 complete ticks for data to leave one processor, arrive at the next processor, and go back to the start.
On a a grid interconnect, this is limited down to 4 for any processor that is one cardinal direction away (assuming proc location influences data point distribution).
It all comes down to matching a communication pattern with an interconnect
Relating to Problem 2A. from sample exam questions,
Question: how would the topology of an interconnect affect the performance of above computation?
This comment was marked helpful 0 times.
Basically the mesh interconnect system basically exactly fits this problem. Each processor would take the cell that matches with its position in the interconnect and then each processor will only have to communicate with other processors that are 1 link away from them.
This comment was marked helpful 0 times.
Analyzing the different interconnect begins with the analysis that the computation on one data point on a processor will most likely require communication to processors that are surrounding on all cardinal directions.
In an interconnect system like the ring interconnect, the structure does not match the communication pattern of the program. We will have to wait 16 complete ticks for data to leave one processor, arrive at the next processor, and go back to the start.
On a a grid interconnect, this is limited down to 4 for any processor that is one cardinal direction away (assuming proc location influences data point distribution).
It all comes down to matching a communication pattern with an interconnect
This comment was marked helpful 0 times.