Bad file access on the rise

[Doug Ledford]

On 06/09/2013 09:53 AM, Roberto Ragusa wrote:
> On 06/08/2013 04:13 PM, Doug Ledford wrote:
>>
>> Yes, but none of these results show the .12s time that your first
>> noatime test run showed in your original post.  If you are now saying
>> that atime is faster than noatime by about .005 to .010s, then these
>> results seem to show that.  But your original post was from .019 to .12,
>> or a difference of .10+s.  That was cache load time, not just the
>> syscall difference.
> 
> Hmm, someone is misreading the results.
> I've reread multiple times, and I see a difference of 12s, not .12s.
> 
> ---> real	0m12.645s
> ---> user	0m0.003s
> ---> sys	0m0.159s
> 
> And 12 seconds (elapsed, with 0.159s system) means 12s/5000=2.4ms
> which could only be explained with the auditing system doing fsync
> calls on its log files.
> 

You're right, I was going from memory and I did not remember it being 12
seconds, I thought it was .12 seconds.  This most certainly is not just
the effects of loading cache.  A 12 second cache load would only happen
if you are loading from stone tablet...