Another CVode, segmentation violation question

[MishaZ]

Working with a rather complex model, after much additions to a previously working cell, I finally got it to compile with no errors. Almost. I get a segmentation violation, pointed to the end of a .ses file that is loaded automatically. After mucking around, I traced the problem to activation of CVode. Setting the default to inactive removes the error. However, manually activating it after compile replicates it. I also get a Signal 6 (whatever that means).
I have never had a problem with CVode outside the NMODL field -- and this time all the mod files work 100%.

So the question is, what should I look for in the code? The difference between a working and a broken simulation is mostly in complexity of synaptic activity...I am at a complete loss, and removing things bit by bit (back to the working version) will take years...

Not too familiar with core_dump, but that just makes the NRN exit right away. Any way to display, debug the processes that lead to the error, to zero in on a faulty code?

Thanks for any help!

[ted]

In order to diagnose a problem, you must first localize its source.

removing things bit by bit (back to the working version) will take years...

Then go back to the working version--you did save that, right?--and add complications
and refinements one at a time, in small increments, testing at every step. An incremental
cycle of revision and testing is essential when you're working on a complex project.

Any way to display, debug the processes that lead to the error, to zero in on a faulty code?

Your best tool is between your ears. Use it to trick NEURON into helping you localize
the problem.
Parse errors produce an immediate error message without executing your simulation.
If the error message isn't informative, comment out lines or blocks of code to try to find
the problem.
Runtime errors can be localized by inserting print statements that report the course of
program execution.

[Raj]

In Dutch we have a saying "Als het kalf verdronken is, dempt men de put" meaning in rough translation "After the calf has drowned the well is filled up". So here a suggestion to get rid of the well and overcome your inability to restore your future working versions, install version control.

The website
http://www.third-bit.com/swc/
gives a gentle introduction to version control using Subversion.

See also the hot tip:
https://www.neuron.yale.edu/phpBB2/viewtopic.php?t=198

Combined with a graphical component RapidSVN and a good diff program it will be a great tool to keep your code tidy and it will allow you to leap back to a previous working version at any time.


Here some some instructions I give to students and colleagues whom I try to lure into using version control (on windows):

Code: Select all

For version control with subversion (under windows) please install the following, in the given order from an administrator account:

   Apache server and libraries:
   http://apache.nedzone.nl/httpd/binaries/win32/apache_2.0.55-win32-x86-no_ssl.msi

   Subversion itself, built on Apache libraries:
   http://subversion.tigris.org/files/documents/15/29065/svn-1.3.0-setup.exe

   Widgets wxwindows:
   http://prdownloads.sourceforge.net/wxwindows/wxMSW-2.6.2-Setup.exe


   RapidSVN Subversion GUI built on Apache and wxwidgets libraries:
   http://rapidsvn.org/download/RapidSVN-0.9.0.exe


Other programs that make life easier are:

   A diff program, showing differences between two files (text or word
   documents)
   http://www.componentsoftware.com/Products/CSDiff/download.htm

   A good plain text editor:
   http://www.pspad.com/en/download.php

General intro on best programming practices, including INTRODUCTION TO SUBVERSION

   http://www.third-bit.com/swc/
   
Configuration:
    
    To benefit most from version management you have to automate a few things. 
    RapidSVN offers a nice GUI but adding the right settings will give this tool extra leverage.
    
    In the menu  under View->Preferences (Programs Tab) you can add links to programs needed for common tasks.
    In my case I have the following settings, please modify to fit your program locations etc.
    
            Standard editor:
                D:\ProgramFiles\PSPadEditor\PSPad.exe
            Program arguments (%1=selected file)
                %1
               
            Standard file explorer:     
                explorer.exe
            Program Arguments (%1=selected directory)
                %1
            
            Diff tool:
                D:\ProgramFiles\ComponentSoftware\CSDiff\CSDiff.exe
            Program Arguments (%1=file1, %2=file2):
                %2 %1
                
    PSPad offers syntax highlighting, this can be customized using Settings->Highlighter Settings.
    For NEURON syntax highlighting files can be found at:
            https://www.neuron.yale.edu/phpBB2/viewtopic.php?t=148
                

[MishaZ]

Thank you for your suggestions!

This case is a bit different from normal practices I would follow, due to the nature of the model (for ex., I had to make some rather radical changes, splitting input streams, etc) -- it would've been hard to do that incrementally, it was more of a tear-down-and-rebuld kind of thing. So what I am stuck with is a code that compiles (so no parse errors), but the runtime error is not really runtime since the run() statement never gets a chance to be called -- all it takes for the error to pop up is setting CVode[0].active(1), after compilation.

I have followed previous similar threads looking for a way to diagnose these kind of errors, but have not found any way to track NEURON's instruction executions to find out what made it give a generic segmentation error (again, core_dump does not help -- but something like WinXP's "debug" button would be awesome :) ).

Sorry if this is monotonous -- I suspect it is probably something simple, hiding in pages of code, which is why I would rather first try to debug, rather than incrementally scale back in many different modules at once...

[Raj]

Do you have an accessed section at the time you create CVode and/or call a function on it?

[MishaZ]

Yes, my "home cell" is accessed...

[ted]

runtime error is not really runtime since the run() statement never gets a chance to be called -- all it takes for the error to pop up is setting CVode[0].active(1)

. . .

I suspect it is probably something simple, hiding in pages of code, which is why I would rather first try to debug, rather than incrementally scale back in many different modules at once...

If the code is truly modular, it should be a simple matter to comment out large chunks of it
and embed print statements at critical points.

[MishaZ]

If the code is truly modular, it should be a simple matter to comment out large chunks of it and embed print statements at critical points.

Ok, I am starting to feel that I am really missing something here...but I will present my case anyway :) I admit I could've designed the code better suited for scaling back...

Taking out chunks of code presents its own problems in that I will have to go over the rest of the code and remove newly dead function, object and segment calls and references -- I do not mean to sound lazy, I just want to make sure it really is the only resort, since it is not an attractive option.

The comment-out method would only work if the code was being executed -- but my concern and difficulty is that the error occurs without execution (since the run() statement is never called -- unless activating CVode somehow executes parts of my code ). And again, everything compiles without errors. NEURON does not print (in this case) the last executed instruction after CVode[0].active(1) is set. What exactly happens in NEURON when CVode is activated, more specifically, how does it interact with the model code, outside of running the actual simulation? And is there any way to track those processes within NEURON itself? I guess this falls back to the debug quesions again...

[ted]

I just want to make sure it really is the only resort

Unless you can get somebody else to do it for you, it's the last resort. Hey, it's all your
code, and it's fresh code at that--it's not like somebody gave you a bag full of somebody
else's old code and told you to refactor it.

my concern and difficulty is that the error occurs without execution (since the
run() statement is never called

All top-level statements (statements that appear outside of a proc, func, or template) are
executed, regardless of whether run() has been called. All procs and funcs that are
called by top level statements are likewise executed. Statements inside a template will
also be executed, as appropriate, if an instance of that class is created.

everything compiles without errors

You mean the mod files? Have you checked them with modlunit? Do they use dt?

NEURON does not print (in this case) the last executed instruction after CVode[0].active(1) is set.

Exactly what happens when a run-time error occurs.

[Raj]

Another question:
In my experience some graphs introduce suboptimal stability into NEURON when used with CVode although admittedly mostly at runtime. Anyway, you might want to try to activate CVode without opening graphs. Do you have any graphs open, while changing to variable step method?