[mpich-discuss] mpich2 hangs on Ubuntu beowulf cluster(with NFS)

[Pavan Balaji]

There are some tools available that will try to find deadlocks in MPI 
codes, such as this for example: 
http://www.cs.utah.edu/formal_verification/ISP-release/

But in some cases (I believe, if the number of pending requests or wild 
card requests is too many), it might run very slowly.  Still worth a try.

  -- Pavan

On 01/04/2012 07:20 PM, Nicolas Rosner wrote:
>> the code we use is an old one
>
> The fact that the code was written long ago can be the main or only
> cause of some problems, but in practice, it rarely is. (It probably
> was somehow broken in the first place, maybe in some subtle way that
> went unnoticed under the older MPI.)
>
> Had the code been correct by, say, the MPI 1.1 standard, in most cases
> it would still be correct by the latest standard.
> (Backward-compatibility is an important goal, carefully preserved by
> those in charge of improving the standard over time.)
>
> When legacy code fails, the real fault usually turns out to be more
> than merely its age -- perhaps it was written under one implementation
> of MPI and never tested with any other one, for instance?
>
>
>> But this is an interesting topic so I will
>> probably read about deadlock and
>> I may come up with a solution
>
> Yeah, it is interesting, and understanding what it is and how it works
> is important. I think it's a wise decision to read up on the matter;
> it should pay off.
>
> Then again, while you are still learning the basics, finding solutions
> to existing deadlock problems in legacy code may be not so simple.
> Figuring out some small, few-line examples and classic problems before
> attempting to fix real-world trouble with lots of distractions around
> it (at least before doing so on your own, without help) can spare you
> some frustration and speed up the learning curve, I think.
>
> Good luck! N.
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji