On 7 December 2011 11:24, Reuti <[email protected]> wrote:
> Am 07.12.2011 um 09:33 schrieb William Hay:
>
>> On 6 December 2011 16:44, Reuti <[email protected]> wrote:
>>> Am 06.12.2011 um 17:32 schrieb William Hay:
>>>
>>>> On 6 December 2011 16:03, Reuti <[email protected]> wrote:
>>>>> Am 06.12.2011 um 17:01 schrieb William Hay:
>>>>>
>>>>>> On 6 December 2011 13:10, Reuti <[email protected]> wrote:
>>>>>>> Am 06.12.2011 um 12:16 schrieb William Hay:
>>>>>>>
>>>>>>>> On 6 December 2011 09:48, Reuti <[email protected]> wrote:
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> Am 06.12.2011 um 10:04 schrieb William Hay:
>>>>>>>>>
>>>>>>>>>> 128 > 2147483648
>>>>>>>>>>
>>>>>>>>>> A user has submitted a job requesting 128 slots:
>>>>>>>>>> qstat -j producing the following output:
>>>>>>>>>> parallel environment:  qlc-[1ABCDEFGHIJTWKLMNOPX] range: 128
>>>>>>>
>>>>>>> How are you submitting this? You can use wildcards, but not a regular 
>>>>>>> expression.
>> Not according to the qsub manpage which implies only a literal pe_name
>> ie wildcards are verbotten.
>
> Well, `man qsub` refers to `man sge_types` and there under "parallel_env" 
> wildcards are mentioned.
>
>
>>>>>> <snip>
>>>
>>> Whoa, then I wonder whether this is an undocumented feature, as it's not 
>>> possible for me to enter such a PE name on the command line in 6.2u5.
>> I believe it is as I can't enter it at the qsub command line either.
>> I fix it up in the jsv.  I believe you can qalter the PE to include
>> square brackets though which works.
>
> It works in a different way than `qsub`, i.e. a different error message, but 
> it's in the end not working too.
>
> $ qalter -pe "smp[abc]" 12 3590
> denied: parallel environment "smp[abc]" does not exist
>
> Sure, it should be rejected not searched for.
>
>
>>   The man page says the -pe option
>> accepts a parallel_environment and the commentary implies that this is
>> a pe_name.
>
> In `man sge_types` it says "pe_name := object_name" which being:
>
>   object_name
>       An  object name is a sequence of up to 512 ASCII string characters 
> except "\n", "\t", "\r", " ", "/", ":", "´", "\", "[",
>       "]", "{", "}", "|", "(", ")", "@", "%", "," or the " character itself.
>
>
>>  However while qsub enforces this (excluding square
>> brackets) the scheduler interprets it as a wc_pe_name
>> which permits them.
>
> But you entered them with a trick, at least it all should work consistently.
>
> - a JSV allows to enter a "wc_pe_name", which is not allowed for `qsub`or 
> `qalter`
> - `qalter` takes in the specified PE literally the [ or ], it should be 
> rejected as with `qsub` or both should handle it in the same way
> - documentation should be more consistent. A "wc_pe_name" is defined in `man 
> sge_types` but nowhere allowed to be entered (besides the JSV trick)
> -  "wc_pe_name" allows a wildcard * anywhere (defined under "pattern"), but 
> `qsub` or `qalter` allow it only at the end of the string.
>
>
>>  The ? and * characters are valid in either
>> pe_name and wc_pe_name but with different meanings (literal in the
>> first, metacharacters in the second).
>
> Where is it used literally? In both cases it's used to match zero or infinity 
> characters. But for a PE only at end of the string.
Well by implication when pe_name is equated to object type neither of
which define a special meaning to any characters.  I read the -pe
switch as taking a pe_name since it passes a parameter of that name to
the JSV.


>
>
>> However that isn't the problem.  The wildcard PE works for every other
>> job.  The oddity is that this job doesn't get a reservation when other
>> jobs with lower priority and less restrictive resource requirements
>> (and similar wildcard PEs) do and that qalter claims these PEs have
>> "only" 2 billion slots which is apparently not enough for a 128 slot
>> job.  While I suppose it is legal for use of an undocumented feature
>> to cause demons to fly out of my nose I doubt that is what is
>> happening here.
>
> Can you try to create a copy of the job with `qresub` and change for the copy 
> the resource requests like a time limit. Any change?
It's not my job as such I'll need to get the user's permission to
impersonate them.  I tried resub as a manager and it complains that
I'm not a member of the relevant project.
>
> The error happened already in the past, but the discussion threads closed 
> before they spotted the real cause:
>
> http://arc.liv.ac.uk/pipermail/gridengine-users/2009-February/thread.html#23435
>
> (search for 2147483648).
Well there is a slight distinction.  That bug report appears to be
about pe's with only -2147483648 slots available while mine is
complaining about a positive number of slots available which makes
even less sense.
>
> -- Reuti
>

_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to