[mvapich-discuss] problem with timing and _MCST_SUPPORT_ with 1.9a ?

[Devendar Bureddy]

Hi Evren

On Fri, Oct 5, 2012 at 8:04 AM, Evren Yurtesen IB <eyurtese at abo.fi> wrote:
> Hi Devendar,
>
> Thanks for the patch, I will use it. So defining DISABLE_LOW_LEVEL_TIMERS
> (there was an ifndef in code about this) at any point is a bad idea since
> these timers are absolutely required?

That should also fine. But, just wanted to make sure we use low
overhead timers to avoid any performance impacts.

-Devendar

>
> Thanks,
> Evren
>
>
> On Thu, 4 Oct 2012, Devendar Bureddy wrote:
>
>> Hi Evren
>>
>> We have introduced multicast features only in 1.9a.  The timers are
>> required for UD multicast internal flow control.  unfortunately
>> --enable-fast do not suppress these warnings.   I have attached a
>> small patch to print these warnings only in debug. Please follow below
>> instructions to apply the patch
>>
>> $cd mvapich2-1.9a
>> $  patch -p1 < patch.diff
>> $ make && make install
>>
>> -Devendar
>>
>> On Thu, Oct 4, 2012 at 7:47 AM, Evren Yurtesen IB <eyurtese at abo.fi> wrote:
>>>
>>> Hello,
>>>
>>> I am getting timing warnings (due to cpu speed scaling) with 1.9a but NOT
>>> with previous versions (<=1.8.1). I configured them all with
>>> --enable-fast
>>> and exactly same configure options
>>>
>>> I tracked the issue to:
>>> src/mpid/ch3/channels/mrail/src/rdma/ch3_init.c
>>> about line 372
>>>
>>> #if defined(_MCST_SUPPORT_) || defined(_ENABLE_UD_)
>>>     mv2_init_timers();
>>> #if defined(_MCST_SUPPORT_)
>>>
>>> I have no idea where these timers are required... However in:
>>> src/mpid/ch3/channels/common/src/util/mv2_clock.c
>>> at about line 217 is a ifndef which can skip the call to
>>> mv2_init_timers()
>>> etc.
>>>
>>> #ifndef DISABLE_LOW_LEVEL_TIMERS
>>>
>>> I feel there is something strange here. If I am disabling timing
>>> collection
>>> with --enable-fast, then why some timers are getting initialized? and why
>>> mcast support requires them where nothing else seem to require them? is
>>> that
>>> just result of some quick patching?
>>>
>>> Is this a bug?
>>>
>>> Is it advisable to force DISABLE_LOW_LEVEL_TIMERS from within CFLAGS +=
>>> '-DDISABLE_LOW_LEVEL_TIMERS' ?
>>>
>>>
>>> Thanks,
>>> Evren
>>> _______________________________________________
>>> mvapich-discuss mailing list
>>> mvapich-discuss at cse.ohio-state.edu
>>> http://mail.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>>
>>
>>
>>
>> --
>> Devendar
>>
>



-- 
Devendar