[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [bgl-discuss] File descriptor limits?




This work was done by summer interns working at IBM, we are planning to provide any improvements back to the
developers.  We just downloaded the mpiBLAST-pio version to identify and implement differences.




Robert Olson <olson@xxxxxxxxxxx>
Sent by: owner-discuss@xxxxxxxxxxxxxxx

03/14/2006 02:08 PM

To
Paul Hovland <hovland@xxxxxxxxxxx>
cc
Robert Latham <robl@xxxxxxxxxxx>, discuss@xxxxxxxxxxxxxxx
Subject
Re: [bgl-discuss] File descriptor limits?





Yes, we should do that. Does anyone know these folks? The approach in  
the paper is much more sophisticated than I'd come up with in a  
quickie attempt at doing this.

--bob

On Mar 14, 2006, at 2:00 PM, Paul Hovland wrote:

> More to the point, see this paper: http://www-users.cs.umn.edu/
> ~rangwala/final_bglBLAST.pdf
>
> pioBLAST was extended to better support BG/L.  Doesn't look like  
> this version is publicly available; perhaps we can get the code  
> from IBM.
>
> Paul
>
> Paul Hovland wrote:
>> I was about to suggest the same thing.  And, pioBLAST (~= mpiBLAST-
>> PIO) is supposed to use collective output, too.  See http://
>> mpiblast.lanl.gov/downloads/pubs/IPDPS05-pioBLAST.pdf
>>
>> Paul
>>
>> Robert Olson wrote:
>>>
>>> On Mar 14, 2006, at 12:49 PM, Robert Latham wrote:
>>>
>>>> On Tue, Mar 14, 2006 at 11:30:22AM -0600, Robert Olson wrote:
>>>>> The mpiBLAST code is opening up a bunch of files on the nodes,  
>>>>> and is
>>>>> apparently running into a 100 file descriptor limit.
>>>>
>>>> Hi Bob
>>>>
>>>> Is this the same mpiBLAST as mpiblast.lanl.gov ?  if so you may be
>>>> interested in a patch to "allow mpiBLAST to run on IBM BlueGene"
>>>> (http://mpiblast.lanl.gov/Downloads.Patches.html)
>>>
>>>
>>> Yup - I am running with that applied. I think it was mainly  
>>> configuration fixes.
>>>
>>>>
>>>> More promising is mpiBLAST-pio, which at a quick glance, looks  
>>>> like a
>>>> mpiBLAST that makes use of mpi-io to coordinate client writes.  
>>>> I've
>>>> never played with it, but that sounds like a more correct  
>>>> solution to
>>>> your problem.
>>>
>>> Aha, interesting. I'll have a look. (though this is the database-
>>> read side, not the client-write side, but I'm sure anything I/O  
>>> related will help).
>>>
>>> --bob
>>>
>>> -  
>>> --------------------------------------------------------------------
>>> To add or remove yourself from this mailing list, use the 'notifyme'
>>> command on any BGL machine. To remove: notifyme -n, to add:  
>>> notifyme -y.
>>>
>>
>> -  
>> --------------------------------------------------------------------
>> To add or remove yourself from this mailing list, use the 'notifyme'
>> command on any BGL machine. To remove: notifyme -n, to add:  
>> notifyme -y.
>>
>

- --------------------------------------------------------------------
To add or remove yourself from this mailing list, use the 'notifyme'
command on any BGL machine. To remove: notifyme -n, to add: notifyme -y.