By: Travis (travis.downs.delete@this.gmail.com), December 21, 2017 8:23 am
Room: Moderated Discussions
Adrian (a.delete@this.acm.org) on December 20, 2017 11:09 pm wrote:
> In the "asm" results, the number of high values in the table
> and their positions are very different from your results.
>
>
> When I have run the test for the first time, I have thought that maybe the differences are meaningful.
>
>
> Nevertheless, after repeating the test a few times I have seen that the results are not reproducible,
> i.e. the results depend on the current state of the cache, i.e. on the history of program executions.
>
> At each execution of the "asm" test, the results vary wildly, both
> in the number of high values and in their positions in the table.
By "high values" you mean the really high outliers that show in purple, and greater than 10 cycles, right?
Yes, these appear randomly. The are the "slow mode" I was talking about which is about 2x as slow as the fast mode (which is also slow).
It seems to go in and out of this mode randomly. The results would indicate that they still occur only 10% of the time or less, but this is an underestimate because this matrix is taking the fastest time out of 10 runs, and often both fast and slow modes are present in those 10 runs and so the fast time is reported.
> The other 3 tests give consistently low values with very little differences from your results.
Makes sense: the CPU side of Kaby Lake is micro-architecturally identical to Skylake.
> In the "asm" results, the number of high values in the table
> and their positions are very different from your results.
>
>
> When I have run the test for the first time, I have thought that maybe the differences are meaningful.
>
>
> Nevertheless, after repeating the test a few times I have seen that the results are not reproducible,
> i.e. the results depend on the current state of the cache, i.e. on the history of program executions.
>
> At each execution of the "asm" test, the results vary wildly, both
> in the number of high values and in their positions in the table.
By "high values" you mean the really high outliers that show in purple, and greater than 10 cycles, right?
Yes, these appear randomly. The are the "slow mode" I was talking about which is about 2x as slow as the fast mode (which is also slow).
It seems to go in and out of this mode randomly. The results would indicate that they still occur only 10% of the time or less, but this is an underestimate because this matrix is taking the fastest time out of 10 runs, and often both fast and slow modes are present in those 10 runs and so the fast time is reported.
> The other 3 tests give consistently low values with very little differences from your results.
Makes sense: the CPU side of Kaby Lake is micro-architecturally identical to Skylake.