Daniel Seither wrote:
Am 20.05.2010 19:00, schrieb Sven Eckelmann:
> Sven Eckelmann wrote:
> Ok, have a small suggestion. Please check that mdev - ... is positive.
> Otherwise we would have something like (mdev - avg * avg == -0.000.....,
> sqrt(...) == NaN):
> PING 02:00:00:00:00:02 (02:00:00:00:00:02) 19(47) bytes of data
> 19 bytes from 02:00:00:00:00:02 icmp_seq=1 ttl=50 time=0.10 ms
> --- 02:00:00:00:00:02 ping statistics ---
> 1 packets transmitted, 1 received, 0% packet loss
> rtt min/avg/max/mdev = 0.101/0.101/0.101/nan ms
There seems to be a problem with the floats... The value of mdev - avg^2
cannot become negative if computed with infinite precision. In this
case, it seems to be slightly less than zero because of rounding errors.
In iputils' ping, all calculations are done using integers which
prevents this kind of problems.
Correct. Because of that I recommended to check before calling sqrt.
I think there are two solutions for the problem you discovered:
1) check whether the difference is <= 0 (easy)
2) rewrite the time measurement to use integers (more code to change,
but calculations will be exact)
Please check whether replacing the line calling sqrt by the following
code fixes the problem for you (first solution):
mdev = mdev - avg * avg;
if (mdev > 0.0)
mdev = sqrt(mdev);
mdev = 0.0;
Already done that - that's why I recommended it :)
Or should we aim for the second solution?
I think the first solution is good enough for now. Please send the patch again
with that fix and it is ok for now. If you want you can rewrite the code later
to use only integers.