The probability of a capacity miss can be reduced by enlarging the cache simply because capacity misses occur due to lack of space in the cache. With a conflict miss, too much of the working set maps to the same cache set, resulting in evictions of cache lines even when there is space elsewhere in the cache (in different sets). Increasing cache associativity provides more lines per set, alleviating the conflicts within a set. These two solutions make some sense to me. However, I am not sure I understand how increasing line size decreases cold misses. I was under the impression that cold misses were unavoidable since they just result from the cache being empty. If this is the case, then how can cold misses be decreased with a larger cache line size?
This comment was marked helpful 0 times.
ron
"spatial locality present" is an important caveat there. On a clean cache with 32-byte cache lines, iterating across a 64-byte array will cost you two cold misses. If you make the lines 64-bytes long instead, you'll only get one cold miss. If the data set you're working with isn't contiguous in memory, on the other hand, you can't make the same conclusion.
The probability of a capacity miss can be reduced by enlarging the cache simply because capacity misses occur due to lack of space in the cache. With a conflict miss, too much of the working set maps to the same cache set, resulting in evictions of cache lines even when there is space elsewhere in the cache (in different sets). Increasing cache associativity provides more lines per set, alleviating the conflicts within a set. These two solutions make some sense to me. However, I am not sure I understand how increasing line size decreases cold misses. I was under the impression that cold misses were unavoidable since they just result from the cache being empty. If this is the case, then how can cold misses be decreased with a larger cache line size?
This comment was marked helpful 0 times.
"spatial locality present" is an important caveat there. On a clean cache with 32-byte cache lines, iterating across a 64-byte array will cost you two cold misses. If you make the lines 64-bytes long instead, you'll only get one cold miss. If the data set you're working with isn't contiguous in memory, on the other hand, you can't make the same conclusion.
This comment was marked helpful 0 times.