[LITMUS^RT] Why write overheads to disk during experiments?

[Björn Brandenburg]

On Jan 24, 2012, at 1:21 AM, Mac Mollison wrote:
> The alternative I'm thinking about is saving overheads (i.e.
> feather-trace data) to a large buffer in memory, and then dumping them
> to disk after the experiment has concluded.

In kernel space, it is difficult to get a sufficiently large buffer. In user space, that certainly would be an option.

> The main advantage would be not having interrupts due to disk I/O
> during experiments.

Sure, that's a reasonable workaround. Long-term, a better solution would be to integrate with / rebase on top of PREEMPT-RT, so that interrupts are less of a problem in general.

> Another very minor advantage would be not having gaps in the data. (I
> say "very minor" because the gaps aren't actually a problem.)

No, I don't think so. Gaps can also be created simply because ftcat is starved by real-time tasks. If starved for sufficiently long time, gaps will appear.

> Glenn mentioned the idea of just ftcating to a file in a ram-based
> filesystem, which is another way of accomplishing the same thing with
> minimal modifications to existing infrastructure.

Sounds like a good solution to me, if you have enough RAM that you can dedicate to this. You probably need to move the trace files to stable storage after each task set completes.

- Björn