[Top][All Lists]

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

Re: Memory Leak

From: Sukanta Basu
Subject: Re: Memory Leak
Date: Mon, 21 Jan 2013 10:14:01 -0500

Hi All,

I would also like to answer James' question: "Is there someway to stop and save the current state at a certain iteration, then restart the simulation at that point?" 

I have a restart option in the code and I am using it frequently these days. When a simulation crashes due to the lack of memory, I have to restart it. At the beginning, the memory consumption is almost the same (statistically) as a new run. Then, it grows again until it reaches the limit. 


On Mon, Jan 21, 2013 at 10:04 AM, Sukanta Basu <address@hidden> wrote:
Dear All,

Thanks for your response and suggestions. I would like to give you a bit more background about this code and the errors I am getting. 

1. Originally, I wrote this code in Matlab ~8 years ago. I have used this code extensively (for publications) and never detected any memory leak in Matlab. I am not an "expert" in octave/matlab, but I follow all the "basics" (e.g., pre-allocation of matrices). In other words, none of the matrices are growing in size through iteration.

2. About 6 years ago, I used MatlabMPI to parallelize the code. Again, no problem with memory. The code was not very efficient because of latency of MatlabMPI. 

3. I created a serial version of the code in Octave ~2 months ago. At the beginning, I did not notice any memory issues. After posting the memory issue, I started looking into it more carefully. For example, I started a simulation last night: 
2:15 pm (Jan 20): 1.47 GB (swap: 0.192 GB)
8:15 pm (Jan 20): 1.53 GB (swap: 0.192 GB)
9:50 am (Jan 21): 2.21 GB (swap: 0.192 GB)
So, the serial code does have "small" amount of memory leak. 

4. I created a parallel version of the code in Octave with openmpi_ext package ~1 month ago. The parallelization is much faster than MatlabMPI (as expected). However, the memory leak using this code is severe. The same job (described in item #3) will cross 24 GB within the same time-period. 

From these experiments, I think the leak is coming from openmpi_ext package. I would like to note that I have experience working with large-scale MPI codes; so, it is unlikely due to any faulty implementation of MPI in my code. 

I do not have any experience with valgrind or similar codes. However, I will be happy to email anyone my code. Unfortunately, it is a rather bulky code with several subroutines. It will be difficult for me to trim down this code to reproduce the same error. 

If anyone could look into the openmpi_ext package that would be great! I am using v1.1.0 (the latest). 

Best regards,

On Mon, Jan 21, 2013 at 8:42 AM, Jordi Gutiérrez Hermoso <address@hidden> wrote:
On 19 January 2013 22:46, Sukanta Basu <address@hidden> wrote:
> I am using Octave 3.6.2 on Ubuntu 12.10. I have written a code for
> solving certain fluid dynamics problem (~5000 lines) using Octave
> and the openmpi_ext toolbox. I am noticing a strange problem with
> Octave. The memory usage slowly increases with each iteration.

If there is indeed a memory leak, it can be in several possible

It can be in the openmpi_ext package (due to trademark issues with The
Mathworks, we don't generally refer to Octave packages as
"toolboxes"), which would be a bug in this package.

It could be in in Open MPI. In that case, we should forward the bug to
the OpenM PI maintainers.

It can be intrinsic in your m-file code, if you're inadvertently
growing an array without bound.

It can be in Octave itself, which would be a bug in Octave.

At any rate, this shouldn't happen, and you have given no usable
diagnostic information to find the memory leak. If you are unfamiliar
with hunting memory leaks (using valgrind and such), then try to post
your code which exhibits the memory leak so that other people can hunt
it down. Optimally, you should trim down your code to a working sample
that exhibits the problem.

- Jordi G. H.

Sukanta Basu
Associate Professor
North Carolina State University

Sukanta Basu
Associate Professor
North Carolina State University

reply via email to

[Prev in Thread] Current Thread [Next in Thread]