In general, I would expect that this could be extremely inefficient if the amount of data that is shared between caches is modified by different processors, in a sense "thrashing" the cache. This, in my understanding is the idea behind false sharing in this case. The bullet point above about interconnect traffic is also concerning since each processor could potentially be "shouting", thus causing a lot of overhead.
This comment was marked helpful 0 times.
retterermoore
pwei - this is why the bullet point is labeled "workload-driven evaluation", there are a lot of potential increased costs but the question is will they outweigh the savings for your specific workload? Increasing the cache line size can have the problems you're mentioning, but it also reduces cold/capacity/trueshare misses and we saw in many of the graphs in this lecture that it reduced overall cache misses. The only real way to decide whether it's worth it is to think about how much of your program's runtime is taken up by cold/capacity/trueshare misses and how much false sharing would increase if you increased the cache size.
In general, I would expect that this could be extremely inefficient if the amount of data that is shared between caches is modified by different processors, in a sense "thrashing" the cache. This, in my understanding is the idea behind false sharing in this case. The bullet point above about interconnect traffic is also concerning since each processor could potentially be "shouting", thus causing a lot of overhead.
This comment was marked helpful 0 times.
pwei - this is why the bullet point is labeled "workload-driven evaluation", there are a lot of potential increased costs but the question is will they outweigh the savings for your specific workload? Increasing the cache line size can have the problems you're mentioning, but it also reduces cold/capacity/trueshare misses and we saw in many of the graphs in this lecture that it reduced overall cache misses. The only real way to decide whether it's worth it is to think about how much of your program's runtime is taken up by cold/capacity/trueshare misses and how much false sharing would increase if you increased the cache size.
This comment was marked helpful 0 times.