Re: Calculate average latencies on the fly

[John Kacur <jkacur@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 993 bytes --]


Please inline the patches

There are 1000000 microseconds / second, the default interval is 1000us, 
but let's say we did it more frequently at 100us

1000000 us /  100 intervals = 10000 samples / second * 60 sec/min * 60 
min/hr * 24 hrs/day * 365 days / year * 1000 (very pessimistic diff) =
an accumulated diff of 3.1536×10^14

With a max double of 9x10^15/3.15*10^14 we could run for 30 years before 
we move into normal numbers.

So, the problem with your patch, is not just the problem which it proposes 
to fix, is one I see as mostly theoretical, but there is the potential for 
accumulating small inaccuracies.

Of course, there is always the chance that I made some kind of error in my 
analysis here, please feel free to point it out!

What might be more interesting is how long would it take for the cycles to 
overflow?

Thanks

John

On Mon, 14 Nov 2016, Piotr Gregor wrote:

> There was an error in previous patch. Please find attached corrected patch.
> 
> cheers,
> Piotr
>