[LITMUS^RT] litmus-dev Digest, Vol 60, Issue 10

[Björn Brandenburg]

> On 23 Feb 2017, at 03:17, Shuai Zhao <zs673 at york.ac.uk> wrote:
> 
> Here I attached the testing results for the original pfp-scheduler. We use a Quad-Core AMD 8350 Processor 2GHz with a three level cache architecture (512kb, 2048kb, 2mb). 

Sorry, I have no experience with this architecture. Are you sure there are no significant NMIs on this platform?

> 
> We have not done anything to prevent the cache interference... However, even if so, it should give a relatively stable max value for each identical test with a sufficient number of samples (i.e. executing the longest code path in pfp-scheduler), am I correct? 
> 
> I am running a "meaningless" test with a increasing number of processor, from 1 to 16, and expect a relatively stable max overheads for each test (just to have a view of how the system can interfer the scheduler execution). On each processor there are 5 tasks that keep releasing and executing.

As more cores become active, generally speaking, one would expect interference and thus unpredictability to increase if cache interference is not mitigated for. 

Looking at the data that you sent, I see two things: the total number of samples seems rather small (try running it for much longer), and the 99.9th percentile overheads seem quite stable, which means that the software is likely behaving as expected and that you are looking at hardware issues (that’s my best guess at least). 

- Björn