Re: [PATCH] gdb.base/siginfo-thread.exp: Increase timeout for 'gcore' command

[Wei-min Pan <weimin.pan@oracle.com>]

Pedro Alves wrote:
> On 03/16/2017 06:27 PM, Wei-min Pan wrote:
>   
>> Pedro Alves wrote:
>>     
>>> On 03/16/2017 04:00 PM, Wei-min Pan wrote:
>>>  
>>>       
>>>> Yao Qi wrote:
>>>>     
>>>>         
>>>  
>>>       
>>>>> Did you see timeout fails in all gcore related tests?  gdb_gcore_cmd is
>>>>> used in many places in gdb testsuite.  Did you investigate why it is so
>>>>> slow to generate coredump in gdb?
>>>>>
>>>>>         
>>>>>           
>>>> No, only this test failed with timeout and did so consistently. The
>>>> generated core file was fine.
>>>> We suspect the slow disk performance was the culprit.
>>>>     
>>>>         
>>> I agree with Yao, and I'm not convinced.  The generated core file is just
>>> "8.6M" on my x86_64 Fedora 23 and the test runs in under 1s here.
>>>
>>> $ time make check TESTS="*/siginfo-thread.exp"
>>> ...
>>> real    0m0.781s
>>> user    0m0.554s
>>> sys     0m0.152s
>>>
>>>
>>> What's the size of the core you get?  If you run the test manually,
>>> do we notice any kind of slowness?
>>>   
>>>       
>> The core size is a little over 9.0M but it took much longer to run this
>> individual test:
>>
>> % time make check TESTS="*/siginfo-thread.exp"
>> ...
>> real    0m11.743s
>> user    0m3.892s
>> sys     0m7.572s
>>
>> And I didn't notice any slowness if the test was run by hand.
>>     
>
> You mean that by hand it went faster than that?
> So what is GDB doing differently when run via make check
> that makes it slower than running by hand?
>   

Yes, but not by much faster:

% cat in
run
gcore tmp.gcore
quit

% time my_gdb siginfo-thread -x in
...
real    0m13.327s
user    0m3.504s
sys     0m7.572s

Thanks.
>   
>>> If you have a general slowness issue in your testing host, then
>>> this should be affecting all gcore tests the same.  We have some
>>> tests that generate big cores on purpose even.
>>>   
>>>       
>> Like I said, only this test consistently failed and the core file
>> generated was not that big.
>>     
>
> Which suggests bumping the timeout is not the right thing to do.
>
> Thanks,
> Pedro Alves
>
>