... | ... | @@ -51,9 +51,4 @@ In addition to cache allocation, cache/memory bandwidth monitoring feature was i |
|
|
We setup various [experiments](https://gitlab.com/sosy-lab/software/pqos-wrapper/blob/master/docs/experiments.md) to check effect of cache availability and sharing on parallel benchmarks. Following scenarios were created by modifying the behaviour of `pqos_wrapper`:
|
|
|
- [full-cache-sharing](https://gitlab.com/sosy-lab/software/pqos-wrapper/tree/full-cache-sharing): In this experiment we allocated full cache ways to each benchmark, thus replicating the default cache configuration in a system where each core has a completely shared access for the LLC.
|
|
|
- [min-cache-sharing](https://gitlab.com/sosy-lab/software/pqos-wrapper/tree/min-cache-sharing): In this experiment we replicated a noisy neighbour scenario where a process is using most of the cache by allocating minimum possible cache ways to each benchmark.
|
|
|
- [optimal-cache-sharing-and-isolation](https://gitlab.com/sosy-lab/software/pqos-wrapper/tree/master): In this experiment equal and isolated cache ways were allocated to each of the parallel benchmark which was used to test how cache isolation affects performance of benchmarks.
|
|
|
|
|
|
### Observations
|
|
|
- `full-cache-sharing` experiment provides the same performance as without `pqos_wrapper` which was expected as it replicates the default cache configuration.
|
|
|
- `optimal-cache-sharing-and-isolation` experiment provides a little bit worse performance than `full-cache-sharing`. This is due to the fact that the cache is partitioned and whenever one core does not fully utilize its cache, the others cannot make use of it, so cache usage in total goes down a little bit. But this also shows that there is an influence between parallel benchmarks and using `pqos_wrapper` makes the benchmarks slightly less nondeterministic.
|
|
|
- `min-cache-sharing` experiment's performance is 5 times slower than everything else, this large difference shows how cache availability has a large influence on performance of a benchmark and thus it makes cache allocation a very important aspect for benchmarking. |
|
|
\ No newline at end of file |
|
|
- [optimal-cache-sharing-and-isolation](https://gitlab.com/sosy-lab/software/pqos-wrapper/tree/master): In this experiment equal and isolated cache ways were allocated to each of the parallel benchmark which was used to test how cache isolation affects performance of benchmarks. |
|
|
\ No newline at end of file |