The advantage of precharging immediately after a request finished is that it improves the performance of the worst case. Since when a new request comes the charging has already been finished, we only need to wait for RAS and CAS. The drawback is that it destroys the data in the row buffer, which makes the "best" case only has the same performance with the other case. However, the precharging immediately policy may still be a good choice because of an assumption. It is reasonable to assume the DRAM access is almost random, since the access would first hit cache if there is any locality.
This comment was marked helpful 1 times.
eatnow
I think there might be cases where there is much locality. For example if we have a loop over a bunch of data accesses, then the first round of cold misses would all hit memory, and they would have good locality. Also, there is the case where the working set of the loop is too big to fit in the cache. I think these cases probably happen more than we'd expect, which explains the structure of DRAM as described in the following slides.
The advantage of precharging immediately after a request finished is that it improves the performance of the worst case. Since when a new request comes the charging has already been finished, we only need to wait for RAS and CAS. The drawback is that it destroys the data in the row buffer, which makes the "best" case only has the same performance with the other case. However, the precharging immediately policy may still be a good choice because of an assumption. It is reasonable to assume the DRAM access is almost random, since the access would first hit cache if there is any locality.
This comment was marked helpful 1 times.
I think there might be cases where there is much locality. For example if we have a loop over a bunch of data accesses, then the first round of cold misses would all hit memory, and they would have good locality. Also, there is the case where the working set of the loop is too big to fit in the cache. I think these cases probably happen more than we'd expect, which explains the structure of DRAM as described in the following slides.
This comment was marked helpful 0 times.