On Mon, Jun 21, 2010 at 10:02 AM, Glauber Costa <[email protected]> wrote:
> On Fri, Jun 18, 2010 at 11:35:18AM -0700, Michael Davidson wrote:
>> Yes - on some Intel platforms the roundtrip time can be as much as 1800
>> clock
>> cycles - with the default test threshold of 500 cycles that always fails the
>> "roundtrip > 2 * threshold" test and we exit the loop without ever having
>> updated
>> the delta value.
>>
>> Note that the actual deta on those platforms is only a few tens of cycles.
>>
>> Most importantly, on looking more closely at the code that was added in
>> this change, it really doesn't do anything useful.(other than cause the test
>> to fail on platforms where the roundtrip time his high, even if the measured
>>
>> delta is low). It isn't as if doing a continue at that point will give you
>> another
>> shot at running the test - the loop executes a fixed number of times and
>> sets the result base on the "best" roundtrip time - continuing the loop
>> early
>> if the roundtrip time is large is pointless.
>>
> The first problem is that the patch was not applied in the way I sent.
>
> I sent:
>
> -       for (i = 0; i < NUM_ITERS; i++) {
> +       while (i < NUM_ITERS) {
>
> ...
>
> +               i++;
>
> This would guarantee that if a high roundtrip was found, we would not
> count this as a valid execution. This is very different from what it is
> in the repository right now, which is clearly wrong.

Ok, that was my fault. I will revert the changes.

> Other than thta, if there is a platform where 2 * threshold is so common,
> maybe we can change it to something bigger (maybe 5?). My main point was to 
> rule out
> scheduling issues in userspace. When they happen, they are likely to be quite 
> big.
>
> _______________________________________________
> Autotest mailing list
> [email protected]
> http://test.kernel.org/cgi-bin/mailman/listinfo/autotest
>



-- 
Lucas
_______________________________________________
Autotest mailing list
[email protected]
http://test.kernel.org/cgi-bin/mailman/listinfo/autotest

Reply via email to