Timon Evenblij
2018-07-30 14:37:18 UTC
Hi all,
I recently ran some parsec benchmarks on ARM in full-system mode, to
evaluate the performance hit of a couple of scenarios using a cache with
higher latencies. However, I found that in some cases, the performance
increased instead of decreased (up to 20% lower execution time!), while
running the exact same benchmark.I would think this is too indeterminism in
full system mode, but
https://www.mail-archive.com/gem5-***@gem5.org/msg12088.html suggests
indeterminism should be limited to system calls only. Since parsec is
commonly used as a benchmarking application, I would think there should be
no indeterminism, or at least very limited? Running multiple runs of parsec
and averaging the results would be too time consuming in a simulator.
If anyone can comment to clear things up, that would probably help a lot of
people, including me!
Best regards,
Timon
I recently ran some parsec benchmarks on ARM in full-system mode, to
evaluate the performance hit of a couple of scenarios using a cache with
higher latencies. However, I found that in some cases, the performance
increased instead of decreased (up to 20% lower execution time!), while
running the exact same benchmark.I would think this is too indeterminism in
full system mode, but
https://www.mail-archive.com/gem5-***@gem5.org/msg12088.html suggests
indeterminism should be limited to system calls only. Since parsec is
commonly used as a benchmarking application, I would think there should be
no indeterminism, or at least very limited? Running multiple runs of parsec
and averaging the results would be too time consuming in a simulator.
If anyone can comment to clear things up, that would probably help a lot of
people, including me!
Best regards,
Timon