Previous | Next --- Slide 31 of 47
Back to Lecture Thumbnails
kayvonf

Can someone summarize the artifactual communication that it depicted on this slide?

zeppelin

The inherent communication is all the red dots on the screen, as the algorithm needs that information in order to complete the computation. On the top and bottom rows, we see little to no artifactual communication because the cache lines we load are row major and contain only red dots as a result. The left and right columns, on the other hand, exhibit more artifactual communication because the cache line we receive when we process the red dots contains data that we do not need for computation. As a result, the amount of artifactual communication will increase linearly with the line size.

parallelfifths

What happens at the edges of the data? That is, I understand that artifactual communication just refers to the fact that we must pull a cache line of a certain size and only a fraction of that cache line is necessary for the computation, but what exactly happens when the size of the cache line exceeds the memory addresses that are devoted to the desired data?

sanchuah

In this scenario, the data is ordered in 2D matrix. Because the cache line is row based and the length of cache line is larger than it needs in the left and right columns, it will load unnecessary data and cause false sharing.