This side provides an example of an operation (sending a message on a two hop network) that can be broken down into multiple steps, each of which requires a different resource to perform it (the sending processing, link 1, the bridge, link 2, etc.). The actual steps shown on this slide are not particularly important. For example, I do not expect you to memorize the steps.
But why do you think I included this example? How does it relate to the laundry example we started off with? And how does it relate to the previous slide where the sending of messages was not pipelined?
ankit1990
@kavonf: Taking a shot at the question. Please correct me if I am wrong:-
I think the whole idea behind the slide is to the recognize that bandwidth of a communication/pipeline(or any other operation) is constrained by the slowest link.
For instance, while we are carrying out saxpy operations over a large number of elements, the two entities are:-
1. Memory bus which is transferring the data to the processor.
2. Bandwidth(rate) at which the processor can compute.
Here, the limiting factor is (1) which determines the overall "bandwidth" of the operation.
Similarly, in the laundry example, we noted that the bandwidth was constrained by the "dryer" and not "washer" or "folder".
The previous slide is a simplified form of the general model with just one communication link.
kayvonf
@ankit1990: Great work!
ankit1990
@kayvonf Thank you :D
andymochi
This may be a stupid question...
Is this model an accurate summary of latency in internet communication?
I would think that network delay >> occupancy because the network delay would include routing a packets through an ISP?
From a hardware design perspective, this problem is pretty interesting because it's probably possible to find a <B_small, B_large> such that (occupancy == network delay). It reminds me of the same cost optimization problem we saw on slide 22.
This side provides an example of an operation (sending a message on a two hop network) that can be broken down into multiple steps, each of which requires a different resource to perform it (the sending processing, link 1, the bridge, link 2, etc.). The actual steps shown on this slide are not particularly important. For example, I do not expect you to memorize the steps.
But why do you think I included this example? How does it relate to the laundry example we started off with? And how does it relate to the previous slide where the sending of messages was not pipelined?
@kavonf: Taking a shot at the question. Please correct me if I am wrong:-
I think the whole idea behind the slide is to the recognize that bandwidth of a communication/pipeline(or any other operation) is constrained by the slowest link.
For instance, while we are carrying out saxpy operations over a large number of elements, the two entities are:- 1. Memory bus which is transferring the data to the processor. 2. Bandwidth(rate) at which the processor can compute.
Here, the limiting factor is (1) which determines the overall "bandwidth" of the operation.
Similarly, in the laundry example, we noted that the bandwidth was constrained by the "dryer" and not "washer" or "folder".
The previous slide is a simplified form of the general model with just one communication link.
@ankit1990: Great work!
@kayvonf Thank you :D
This may be a stupid question...
Is this model an accurate summary of latency in internet communication?
I would think that network delay >> occupancy because the network delay would include routing a packets through an ISP?
From a hardware design perspective, this problem is pretty interesting because it's probably possible to find a <B_small, B_large> such that (occupancy == network delay). It reminds me of the same cost optimization problem we saw on slide 22.