So one thing to take away from this lecture is to make sure to correctly organize data for the processors so that our cache misses are lowered?
This comment was marked helpful 0 times.
RICEric22
I would rephrase that as data that is well organized can reduce your cache misses, which is useful only if the program itself exhibits good locality. If the process of re-ordering / reorganizing the data is very time consuming in comparison to the actual computation that occurs afterwards, it may not be worth it.
This comment was marked helpful 0 times.
jhhardin
To add a bit to what @RICEric22 said, another thing to keep in mind is load balancing. Locality is good, but remember the Mandelbrot example from homework 1. In that case, we actually made it significantly faster by assigning threads data that are further apart than assigning by contiguous blocks. While this potentially hurt locality, load balancing was greatly improved.
Overall, locality is something to keep in mind, but there are many factors that should be weighed in when thinking about distributing data.
So one thing to take away from this lecture is to make sure to correctly organize data for the processors so that our cache misses are lowered?
This comment was marked helpful 0 times.
I would rephrase that as data that is well organized can reduce your cache misses, which is useful only if the program itself exhibits good locality. If the process of re-ordering / reorganizing the data is very time consuming in comparison to the actual computation that occurs afterwards, it may not be worth it.
This comment was marked helpful 0 times.
To add a bit to what @RICEric22 said, another thing to keep in mind is load balancing. Locality is good, but remember the Mandelbrot example from homework 1. In that case, we actually made it significantly faster by assigning threads data that are further apart than assigning by contiguous blocks. While this potentially hurt locality, load balancing was greatly improved.
Overall, locality is something to keep in mind, but there are many factors that should be weighed in when thinking about distributing data.
This comment was marked helpful 0 times.