commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Luc Maisonobe <...@spaceroots.org>
Subject Re: [math] Attempt to circumvent some errors which seem to be platform-dependent.
Date Thu, 07 May 2015 13:58:14 GMT
Le 07/05/2015 15:42, Thomas Neidhart a écrit :
> I did not want to invalidate your effort,


Don't worry about that. I learned a lot doing this, including
something that really surprised me: pow(NaN, 0.0) *must* be 1.0
and not NaN. AFAIK this is the only case where a NaN input does not
propagate to output.


> but check if a faulty branch
> prediction / optimization might be the cause of the problem, which seems to
> be confirmed now.

Yes, it seems to be the problem.

> 
> If you have an improved version of the pow function, please go ahead with
> your modifications.

It doesn't improve anything, I run the performance benchmark and the
results are exactly the same as our previous implementation, neither
faster nor slower.

Could you merge your change back to master?

thanks again for this fix.
Luc

> 
> Thomas
> 
> 
> On Thu, May 7, 2015 at 3:28 PM, Luc Maisonobe <luc@spaceroots.org> wrote:
> 
>> Le 07/05/2015 15:04, Thomas Neidhart a écrit :
>>> After my last test, the test failures seem to have disappeared, see here:
>>> https://builds.apache.org/job/Commons%20Math%20H10/
>>
>> Thanks.
>>
>> I was working on this since two days. It also occurred to me that the
>> optimizers did rearrange the order of the if statements, probably
>> because it thought they were independent from each other.
>>
>> I did rewrite completely the pow function to ensure nesting of
>> conditions, and at the same time take care of some other specific cases.
>> My current rewrite was complete but created other failures elsewhere.
>> I will still check its performances in time and will try to solve the
>> failures only if it improves things a lot, otherwise i will simply drop it.
>>
>> best regards,
>> Luc
>>
>>>
>>> Thomas
>>>
>>> On Tue, May 5, 2015 at 4:39 PM, Luc Maisonobe <luc@spaceroots.org>
>> wrote:
>>>
>>>> Le 05/05/2015 10:51, Luc Maisonobe a écrit :
>>>>> Le 04/05/2015 23:30, Thomas Neidhart a écrit :
>>>>>> On 05/04/2015 09:31 PM, Luc Maisonobe wrote:
>>>>>>> Le 04/05/2015 19:13, sebb a écrit :
>>>>>>>> On 4 May 2015 at 16:43, luc <luc@spaceroots.org> wrote:
>>>>>>>>> Le 2015-05-04 14:48, Thomas Neidhart a écrit :
>>>>>>>>>>
>>>>>>>>>> Problem still remains, see here:
>>>>>>>>>> https://builds.apache.org/job/Commons%20Math%20H10/49/console
>>>>>>>>>>
>>>>>>>>>> The test failures only occur on the following slaves
it seems:
>>>>>>>>>>
>>>>>>>>>>  * H10
>>>>>>>>>>  * ubuntu-2
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> This looks like what happened a few months ago then.
>>>>>>>>>
>>>>>>>>> I will try to look at it. As we are only able to reproduce
this on
>>>> this
>>>>>>>>> build system, I guess this implies committing lots of
small changes
>>>> (with
>>>>>>>>> System.out.println and the like) and triggering a custom
buid from
>>>> the
>>>>>>>>> Jenkins configuration above. I can do that, but wonder
if there is
>>>> another
>>>>>>>>> way without committing the tests in the master branch.
Can we set
>> up
>>>> an
>>>>>>>>> h10-builds branch that would be used by the job above
and would be
>>>> ignored
>>>>>>>>> by the regular job ?
>>>>>>>>
>>>>>>>> Yes, just configure the Jenkins job to use the appropriate
branch.
>>>>>>>> It is currently using
>>>>>>>>
>>>>>>>> */master
>>>>>>>
>>>>>>> I explored the project and did not find any way to change the
build.
>>>>>>> Looking at the REST API explanation page, I found a reference
to
>>>>>>> JOBNAME/config.xml. However, if I try this URL I get an access
denied
>>>>>>> failure with message "luc is missing the Job/ExtendedRead permission"
>>>>>>> and if I try the createItem URL I get an access denied failure
with
>>>>>>> message "luc is missing the Job/Create permission".
>>>>>>>
>>>>>>> Could someone give me karma for changing these elements on the
>> Commons
>>>>>>> Math H10 job?
>>>>>>
>>>>>> I changed the job configuration to use the h10-builds branch and
>> started
>>>>>> a new build already.
>>>>>
>>>>> OK, thanks.
>>>>> It seemed to work up to commit 96903ec I made this morning, but not
>>>>> for the two next ones (67da172 and 14f24f2). They were not picked up
>>>>> even after wating a long time and so each time I trigger a new build
it
>>>>> is still done from 96903ec.
>>>>
>>>> I don't know why, but it started working again in the afternoon and I
>>>> could do a few tests.
>>>> Then it seemded to get stuck on a commit. I tried to force it reload
>>>> its work space by cleaning it. I was allowed to do it but now the build
>>>> does not work anymore. I probably not only cleaned old stuff but also
>>>> completely wiped the directory ... I still cannot change anything else
>>>> on the job.
>>>>
>>>> Gary, could you add me to the hudson-jobadmin group (looking at
>>>> <http://wiki.apache.org/general/Jenkins> it seems the way to get
>> access).
>>>>
>>>> I am sorry.
>>>> Luc
>>>>
>>>>>
>>>>> Luc
>>>>>
>>>>>>
>>>>>> I am not sure that I have the credentials to give you the permissions.
>>>>>>
>>>>>> Thomas
>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>>>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>> ---------------------------------------------------------------------
>>>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> ---------------------------------------------------------------------
>>>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>>>> For additional commands, e-mail: dev-help@commons.apache.org
>>>>
>>>>
>>>
>>
>>
>> ---------------------------------------------------------------------
>> To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
>> For additional commands, e-mail: dev-help@commons.apache.org
>>
>>
> 


---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org


Mime
View raw message