The bandwidth goes up as cache line increases and it's bad in a sense that we're more likely to be bandwidth-bound.
Reducing cache misses does not always result in reducing data BW.
This comment was marked helpful 0 times.
shabnam
Slides clearly show us that our cache line size may help latency but may affect bandwidth negatively.
This comment was marked helpful 0 times.
uhkiv
Let me see if I get this: so, in the previous slide, we showed that as cache line size increases, miss rates decrease, with the exception of radix sort. Over here, we see that as cache line size increases, bandwidth skyrockets in all cases except for LU. We can say that LU (and some other cases, to a certain extent) has good spatial locality since increasing the cache line size does not increase the bandwidth, which means that once it loads a cache line in, it will take a long time to miss again.
Now, with other cases, without Radix sort, we saw that miss rates were decreasing. However, the reason why bandwidth increases is that each time we miss, we need to load a big chunk of data from memory (since the cache line is bigger).
With radix sort, we saw that miss rates were increasing, so obviously bandwidth will increase.
If we are bandwidth-bound, then we might get more performance by decreasing the cache size and taking more misses (utilizing more latency). If we are latency-bound, then we might get more performance by increasing the cache size and taking less misses (utilizing more bandwidth).
This comment was marked helpful 0 times.
mofarrel
@taegyunk
Its not necessarily true that this will make us more easily bandwidth bound. If we are reading a large data set in, we will definitely end up using the entire cache line (even if it is large). This would mean that the cache line size does not actually matter. This is more of a concern if we are running programs that are reading small amounts of data from many sporadic locations.
This comment was marked helpful 0 times.
eatnow
What is LU and why is its traffic usage trends so different from the other applications? Also, what exactly is sent over the data bus and address bus?
The bandwidth goes up as cache line increases and it's bad in a sense that we're more likely to be bandwidth-bound.
Reducing cache misses does not always result in reducing data BW.
This comment was marked helpful 0 times.
Slides clearly show us that our cache line size may help latency but may affect bandwidth negatively.
This comment was marked helpful 0 times.
Let me see if I get this: so, in the previous slide, we showed that as cache line size increases, miss rates decrease, with the exception of radix sort. Over here, we see that as cache line size increases, bandwidth skyrockets in all cases except for LU. We can say that LU (and some other cases, to a certain extent) has good spatial locality since increasing the cache line size does not increase the bandwidth, which means that once it loads a cache line in, it will take a long time to miss again.
Now, with other cases, without Radix sort, we saw that miss rates were decreasing. However, the reason why bandwidth increases is that each time we miss, we need to load a big chunk of data from memory (since the cache line is bigger).
With radix sort, we saw that miss rates were increasing, so obviously bandwidth will increase.
If we are bandwidth-bound, then we might get more performance by decreasing the cache size and taking more misses (utilizing more latency). If we are latency-bound, then we might get more performance by increasing the cache size and taking less misses (utilizing more bandwidth).
This comment was marked helpful 0 times.
@taegyunk Its not necessarily true that this will make us more easily bandwidth bound. If we are reading a large data set in, we will definitely end up using the entire cache line (even if it is large). This would mean that the cache line size does not actually matter. This is more of a concern if we are running programs that are reading small amounts of data from many sporadic locations.
This comment was marked helpful 0 times.
What is LU and why is its traffic usage trends so different from the other applications? Also, what exactly is sent over the data bus and address bus?
This comment was marked helpful 0 times.