The heading of this slide confuses me. How is this a problem of fixed problem sized analysis specifically? Can't super-linear speedups occur of time and memory fixed speedup analysis as well?
EggyLv999
This particular form of problem (cache-related) wouldn't occur in a fixed memory speedup analysis.
cloudhary
This speedup is throwing me off a tad bit. Much like how we did in a previous assignment, shouldn't the speedup be measured specifically for just the computation and not the entire function?
BBB
If the function is compute-bound, then maybe you can get away with neglecting memory access time. In general though ignoring memory access time when computing speedup would grossly misrepresent speedup, especially for memory-bound functions.
In the real-world, you have wait for both the computation and the memory accesses. This is why memory latency and bandwidth are so crucial to multi-core performance. If you can't feed enough data into the compute units, they can't compute anything. If you have an application that spends
2 seconds getting its inputs from memory and 10 seconds computing the answer, and a second one that spends 7 seconds getting the same inputs from memory and only 5 seconds computing the answer, you can't say that the second one is twice as fast.
The heading of this slide confuses me. How is this a problem of fixed problem sized analysis specifically? Can't super-linear speedups occur of time and memory fixed speedup analysis as well?
This particular form of problem (cache-related) wouldn't occur in a fixed memory speedup analysis.
This speedup is throwing me off a tad bit. Much like how we did in a previous assignment, shouldn't the speedup be measured specifically for just the computation and not the entire function?
If the function is compute-bound, then maybe you can get away with neglecting memory access time. In general though ignoring memory access time when computing speedup would grossly misrepresent speedup, especially for memory-bound functions.
In the real-world, you have wait for both the computation and the memory accesses. This is why memory latency and bandwidth are so crucial to multi-core performance. If you can't feed enough data into the compute units, they can't compute anything. If you have an application that spends 2 seconds getting its inputs from memory and 10 seconds computing the answer, and a second one that spends 7 seconds getting the same inputs from memory and only 5 seconds computing the answer, you can't say that the second one is twice as fast.