Feeding a Large-scale Physics Application to Python

David M. Beazley
Department of Computer Science
University of Utah
Salt Lake City, Utah 84112
beazley@cs.utah.edu

Peter S. Lomdahl
Theoretical Division
Los Alamos National Laboratory
Los Alamos, New Mexico 87545
pxl@lanl.gov

(Presented at the 6th International Python Conference, San Jose, California. October 14-17, 1997).

Abstract

We describe our experiences using Python with the SPaSM molecular dynamics code at Los Alamos National Laboratory. Originally developed as a large monolithic application for massively parallel processing systems, we have used Python to transform our application into a flexible, highly modular, and extremely powerful system for performing simulation, data analysis, and visualization. In addition, we describe how Python has solved a number of important problems related to the development, debugging, deployment, and maintenance of scientific software.

Background

For the past 5 years, we have been developing a large-scale physics code for performing molecular-dynamics simulations of materials. This code, SPaSM (Scalable Parallel Short-range Molecular-dynamics), was originally developed for the Connection Machine 5 massively parallel supercomputing system and later moved to a number of other machines including the Cray T3D, multiprocessor Sun and SGI systems, and Unix workstations [1,2]. Our goal has been to investigate material properties such as fracture, crack propagation, dislocation generation, friction, and ductile-brittle transitions [3]. The SPaSM code helps us investigate these problems by performing atomistic simulations--that is, we simulate the dynamics of every atom in a material and hope to make sense of what happens. Of course, the underlying physics is not so important for this discussion.

While the development of SPaSM is an ongoing effort, we have been hampered by a number of serious problems. First, typical simulations generate tens to hundreds of gigabytes of data that must be analyzed. This task is not easily performed on a user's workstation, nor is it economically feasible to buy everyone their own personal desktop supercomputer. A second problem is that of interactivity and control. We are constantly making changes to investigate new physical models, different materials, and so forth. This would usually require changes to the underlying C code--a process that was tedious and not very user friendly. We wanted a more flexible mechanism. Finally, there were many difficulties associated with the development and maintenance of our software. While we are only a small group, it was not uncommon for different users to have their own private copies of the software that had been modified in some manner. This, in turn, led to a maintenance nightmare that made it almost impossible to update the software or apply bug-fixes in a consistent manner.

To address these problems, we started investigating the use of scripting languages. In 1995, we wrote a special purpose parallel-scripting language, but replaced it with Python a year later (although the interface generation tool for that scripting language lives on as SWIG). In this paper, we hope to describe some of our experiences with Python and how it has helped us solve practical scientific computing problems. In particular, we describe the organization of our system, module building process, interesting tools that Python has helped us develop, and why we think this approach is particularly well-suited for scientific computing research.

Why Python?

Although our code originally used a custom scripting language, we decided to switch to Python for a number of reasons :

Another factor, not to be overlooked, is the increasing acceptance of Python elsewhere in the computational science community. Efforts at Lawrence Livermore National Laboratory and elsewhere were attractive--in particular, we saw Python as being a potential vehicle for sharing modules and utilizing third-party tools [4,9,10].

System Organization

Our goal was to build a highly modular system that could perform simulation, data analysis, and visualization--often performing all of these tasks simultaneously. Unfortunately, there is a tendency to do this by building a tightly integrated monolithic package (perhaps using a well-structured C++ class hierarchy for example). In our view, this is too formal and restrictive. We wanted to support modules that might only be loosely related to each other. For example, there is no need for a graphics library to depend on the same structures as a simulation code or to even be written in the same language for that matter. Likewise, we wanted to exploit third-party modules where no assumptions could be made about their internal structure.

The major components of our system take the form of C libraries. When Python is used, the libraries are compiled into shared libraries and dynamically loaded into Python as extension modules. The functionality of each library is exposed as a collection of Python "commands." Because of this, the C and Python programming environments are closely related. In many cases, it is possible to implement the same code in both C or Python. For example :

/* A simple function written in C */
#include "SPaSM.h"
void run(int nsteps, double Dt, int freq) {
    int i;
    char filename[64];
    for (i = 0; i < nsteps; i++) {
         integrate_adv_coord(Dt);
         boundary_periodic();
         redistribute();
         force_eam();
         integrate_adv_velocity(Dt);
         if ((i % freq) == 0) {
            sprintf(filename,"Dat%d",i);
            output_particles(filename);
         }
    }
}
Now, in Python :
# A function written in Python
from SPaSM import *
def run(nsteps, Dt, freq):
    for i in xrange(0, nsteps):
         integrate_adv_coord(Dt)
         boundary_periodic()
         redistribute()
         force_eam()
         integrate_adv_velocity(Dt)
         if (i % freq) == 0) :
            output_particles('Dat'+str(i))
While C libraries provide much of the underlying functionality, the real power of the system comes in the form of modules and scripts written entirely in Python. Users write scripts to set up and control simulations. Several major components such as the visualization and data-analysis system make heavy use of Python. We also utilize a variety of modules in the Python library and have dynamically loadable versions of Tkinter and the Python Imaging Library [11].

Embedding Python and Hiding System Dependencies

One of the biggest implementation problems we have encountered is the fact that the SPaSM code is a parallel application that relies heavily upon the proper implementation of low-level system services such as I/O and process management. Currently, it is possible to run the code in two different configurations--one that uses message passing via the MPI library and another using Solaris threads. Using Python with both versions of code requires a degree of care--in particular, we have found it to be necessary to provide Python with enhanced I/O support to run properly in parallel. This work has been described elsewhere [5,6].

Handling two operational modes introduces a number of problems related to code maintenance and installation. Traditionally, we would recompile the entire system and all of its modules for each configuration (placing the files in an architecture dependent subdirectory). Users would have to decide which system they wanted to use and compile all of their modules by linking against the appropriate libraries and setting the right compile-time options. Changing configurations would typically require a complete recompile.

With dynamic loading and shared libraries however, we have been able to devise a different approach to this problem. Rather than recompiling everything for each configuration, we use an implementation independent layer of system wrappers. These wrappers provide a generic implementation of message passing, parallel I/O, and thread management. All of the core modules are then compiled using these generic wrappers. This makes the modules independent of the underlying operational mode---to use MPI or threads, we simply need to supply a different implementation of the system wrapper libraries. This is easily accomplished by building two different versions of Python that are linked against the appropriate system libraries. To run in a particular mode, we now just run the appropriate version of Python (i.e. 'python' or 'pythonmpi'). The neat part about this approach is that all of the modules work with both operational modes without any recompilation or reconfiguration. If a user is using threads, but wants to switch to MPI, they simply run a different version of Python--no recompilation of modules is necessary.

A full discussion of writing system-wrappers can be found elsewhere. In particular, a discussion of writing parallel I/O wrappers for Python can be found in [5]. An earlier discussion of the technique we have used for writing message passing and I/O wrappers can also be found in [7].

Module Building with SWIG

To build modules, we have been using SWIG [8]. Each module is described by a SWIG interface file containing the ANSI C declarations of functions, structures, and variables in that module. For example :
// SWIG interface file
%module SPaSM
%{
#include "SPaSM.h"
%}
void integrate_adv_coord(double Dt);
void boundary_periodic();
void redistribute();
void force_eam();
void integrate_adv_velocity(double Dt);
int  output_particles(char *filename);
SWIG provides a logical mapping of the underlying C implementation into Python. During compilation, interface files are automatically converted into wrapper code and compiled into Python modules. This process is entirely transparent--changes made to the interface are automatically propagated to Python whenever a module is recompiled. Given the constantly evolving nature of research applications, this makes it easy to extend and maintain the system.

Separation of Implementation and Interface

An important aspect of SWIG is that it is requires no modifications to existing C code which allows us to maintain a strict separation between the implementation of C modules and their Python interface. We believe that this results in code that is more organized and generally reusable. There is no particular reason why a C module should depend on Python (one might want to use it as a stand-alone package or in a different application). Despite using Python extensively, the SPaSM code can still be compiled with no Python interface (of course, you lose all of the benefits gained by having Python). We feel that most physics codes tend to have a rather long life-span. Maintaining a separation of implementation and interface helps insure that our physics code will be usable in the future--even if there are drastic changes in the interface along the way.

Providing Access to Data Structures

For the purposes of debugging and data exploration, we have used SWIG to provide wrappers around C data structures. For example, a collection of C structures such as the following
typedef struct {
     double x,y,z;
} Vector;

typedef struct {
     int    type;
     Vector r;
     Vector v;
     Vector f;
} Particle;
can be turned into Python wrapper classes. In addition, SWIG can extend structures with member functions as follows :
// SWIG interface for vector and particle structures
%include datatypes.h

// Extend data structures with a few useful methods
%addmethods Vector {
     char *__str__() {
         static char a[1024];
         sprintf(a,"[ %0.10f, %0.10f, %0.10f ]", self->x, self->y, self->z);
         return a;
     }
}
%addmethods Particle {
     Particle *__getitem__(int index) {
          return self+index;
     }
     char *__str__() {
          // print out a particle
          ...
     }
}
When the Python interface is built, C structures now appear like Python objects. By providing Python-specific methods (such as __getitem__) we can even provide array access. For example, the following Python code would print out all of the coordinates of stored particles :
# Print out all coordinates to a file
f = open("part.data","w")
p = SPaSM_first_particle()          # Get first particle pointer
for i in xrange(0,SPaSM_count_particles()):
    f.write("%f, %f, %f\n" % (p[i].r.x, p[i].r.y, p[i].r.z))
f.close()
While only a simple example, having direct access to underlying data has proven to be quite valuable since we view the internal representation of data, check values, and perform diagnostics.

Improving the Reliability of Modules

When instrumenting our original physics application to use scripting, we found that many parts of the code were not written in an entirely "reliable" manner. In particular, the code had never been operated in an event-driven manner. Functions often made assumptions about initializations and rarely checked the validity of input parameters. To address these issues, most sections of code were gradually changed to provide some kind of validation of input values and error recovery. The SWIG compiler has also been extended to provide some of these capabilities as well.

One of Python's most powerful features is its exception handling mechanism. Exceptions are easily raised and handled in Python scripts as follows :

# A Python function that throws an exception
def allocate(nbytes):
	ptr = SPaSM_malloc(nbytes)
        if ptr == "NULL": raise MemoryError,"Out of memory!"

# A Python function that catches an exception
def foo():
	try:
		allocate(NBYTES)
	except:
		return    # Bailing out

We have borrowed this idea and implemented a similar exception handling mechanism for our C code. This is accomplished using functions in the <setjmp.h> library and defining a few C macros for "Try", "Except", "Throw", etc... Using these macros, many of our library functions now look like the following :
/* A C function that throws an exception */
void *SPaSM_malloc(size_t nbytes) {
    void *ptr = (void *) malloc(nbytes);
    if (!ptr) Throw("SPaSM_malloc : Out of memory!");
    return ptr;
}
Like Python, we allow C functions to catch exceptions and provide their own recovery as follows :
/* A C function catching an exception */
int foo() {
    void *p;
    Try {
        p = SPaSM_malloc(NBYTES);
    } Except {
        printf("Unable to allocate memory. Returning!\n");
        return -1;
    }
}
In the case of a stand-alone C application, exceptions can be caught and handled internally. Should an uncaught exception occur, the code prints an error message and terminates. However, when Python is used, we can generate Python exceptions using a SWIG user-defined exception handler such as the following :
%module SPaSM
// A SWIG user defined exception handler
%except(python) {
     Try {
          $function
     } Except {
          PyErr_SetString(PyExc_RuntimeError,SPaSM_error_msg());
     }
}
// C declarations 
...
The handler code gets placed into all of the Python "wrapper" functions and is responsible for translating C exceptions into Python exceptions. Doing this makes our physics code operate in a more seamless manner and gives it a precisely defined error recovery procedure (i.e. internal errors always result Python exceptions). For example :
>>> SPaSM_malloc(1000000000)
RuntimeError: SPaSM_malloc(1000000000). Out of memory!
(Line 52 in memory.c)
>>>
We have found error recovery to be critical. Without it, simulations may continue to run, only to generate wrong answers or a mysterious system crash.

More Than Scripting

When we originally started using scripting languages, we thought they would mainly be a convenient mechanism for gluing C libraries together and controlling them in an interactive manner. However, we have come to realize that scripting languages are much more powerful than this. Now, we find ourselves implementing significant functionality entirely in Python--bypassing C altogether. The most surprising (well, not really that surprising) fact is that Python makes it possible to build very powerful tools with only a small amount of extra programming. In this section, we present a brief overview of some of the tools we have developed--most of which have been implemented largely in Python.

Object-Oriented Visualization and Data Analysis

Our original goal was to add a powerful data analysis and visualization component to our application. To this end, we developed a lightweight high-performance graphics library implemented in C. This library supports both 2D and 3D plotting and produces output in the form of GIF images. To make plots, a user writes simple C functions such as the following :
void plot_spheres(Plot3D *p3, DataFunction func, double min, double max,
                  double radius) {
    Particle *p = Particles;
    int       i, npart, color;
    npart = SPaSM_count_particles();
    for (i = 0; i < npart; i++, p++) {
        /* Compute color value */
        value = (*func)(p);
        color = (value-min)/(max-min)*255;
        /* Plot it */
        Plot3D_sphere(p3,p->r.x,p->r.y,p->r.z,radius,color);
    }
}
When executed, this function produces a raw image such as the following showing stacking-faults generated by a passing shock wave in an fcc crystal of 10 million atoms :

While simple, we often want our images to contain more information including titles, axis labels, colorbars, time-stamps, and bounding boxes. To do this, we have built an object-oriented visualization framework in Python. Python classes provide methods for graph annotation and common graph operations. If a user wants to make a new kind of plot, they simply inherit from an appropriate base class and provide a function to plot the desired data. For example, a "Sphere" plot using the above C function would look like this :

class Spheres(Image3D):
      def __init__(self, func, min, max, radius=0.5):
             Image3D.__init__(self, ... 
             self.func = func
             self.min  = min
             self.max  = max
             self.radius = radius
      def draw(self):
             self.newplot()
             plot_spheres(self.p3,self.func,self.min,self.max,self.radius)
To use the new image, one simply creates an object of that type and manipulates it. For example :
>>> s = Spheres(PE,-8,-3.5)   # Create a new image
>>> s.rotd(45)                # Rotate it down
>>> s.zoom(200)               # Zoom in
>>> s.show()                  # generate Image and display it
>>> ...

Thus, with a simple C function and a simple Python class, we get a lot of functionality for free, including methods for image manipulation, graph annotation, and display. In the current system, there are about a dozen different types of images for 2D and 3D plotting. It is also possible to perform filtering, apply clipping planes, and other advanced operations. Most of this is supported by about 2000 lines of Python code and a relatively small C library containing performance critical operations.

Web-Based Simulation Monitoring

One feature of many physics codes is that they tend to run for a long time. In our case, a simulation may run for tens to hundreds of hours. During the course of a simulation, it is desirable to check on its status and see how it is progressing. Given the strong Internet support already bundled with Python, we decided to write a simple physics web-server that could be used for this purpose. Unlike a more traditional server, our server is used by "registering" various objects that one wants to look at during the course of a simulation. The simulation code then periodically polls a network socket to see if anyone has requested anything. If so, the simulation will stop for a moment, generate the requested information, send it to the user, and then continue on with the calculation. A simplified example of using the server is as follows :
# Simplified script using a web-server
from vis import *
from web import *

web = SPaSMWeb()
web.add(WebMain(""))
web.add(WebFileText("Msg"+`run_no`))

# Create an image object

set_graphics_mode(HTTP)
ke = Spheres(KE,0,20)
ke.title = "Kinetic Energy"
web.add(WebImage("ke.gif",ke))

def run(nsteps):
	for i in xrange(0,nsteps):
		integrate(1)           # Integrate 1 timestep
                web.poll()             # Anyone looking for me?

# Run it
run(10000)
While simple, this example gives the basic idea. We create a server object and register a few "links." When a user connects with the server, they will be presented with some status information and a list of available links. When the links are accessed, the server feeds real-time data back to the user. In the case of images, they are generated immediately and sent back in GIF format. During this process, no temporary files are created, nor is the server transmitting previously stored information (i.e. everything is created on-the-fly).

While we are still refining the implementation, this approach has turned out to be useful--not only can we periodically check up on a running simulation, this can can be done from any machine that has a Web-browser, including PCs and Macs running over a modem line (allowing a bored physicist to check on long-running jobs from home or while on travel). Of course, the most amazing fact of all is that this was implemented by one person in an afternoon and only involved about 150 lines of Python code (with generous help from the Python library).

Development Support

Finally, we have found Python to be quite useful for supporting the future development of our code. Some of this comes in the form of sophisticated debugging--Python can be used to analyze internal data structures and track down problems. Since Python can read and modify most of the core C data structures, it is possible to prototype new functions or to perform one-time operations without ever loading up the C compiler.

We have also used Python in conjunction with code management. One of the newer problems we have faced is the task of finding the definitions of C functions (for example, we might want to know how a particular command has been implemented in C). To support this, we have written tools for browsing source directories, displaying definitions, and spawning editors. All of this is implemented in Python and can be performed directly from the Physics application. Python has made these kinds of tools very easy to write--often only requiring a few hours of effort.

Python and the Development of Scientific Software

The adoption of Python has had a profound effect on the overall structure of our application. With time, the code became more modular, more reliable, and better organized. Furthermore, these gains have been achieved without significant losses in performance, increased coding complexity, or substantially increased development cost. It should also be stressed that automated tools such as SWIG have played a critical role in this effort by hiding the Python-C interface and allowing us to concentrate on the real problems at hand (i.e. physics).

One point, that we would like to strongly emphasize, is the dynamic nature of many scientific applications. Our application is not a huge monolithic package that never changes and which no one is supposed to modify. Rather, the code is constantly changing to explore new problems, try new numerical methods, or to provide better performance. The highly modular nature of Python works great in this environment. We provide a common repository of modules that are shared by all of the users of the system. Different users are typically responsible for different modules and all users may be working on different modules at the same time. Whenever changes to a module are made, they are immediately propagated to everyone else. In the case of bug-fixes, this means that patched versions of code are available immediately. Likewise, new features show up automatically and can be easily tested by everyone. And of course, if something breaks--others tend to notice almost immediately (which we feel is a good thing).

Limitations of Python

While we consider our Python "experiment" to be highly successful, we have experienced a number of problems along the way. The first is merely a technical issue--it would be nice if the mechanism for building static extensions to Python were made easier. While we are currently using dynamic loading, many machines, especially supercomputing systems, do not support it. On these machines it is necessary to staticly link everything. Libraries help simplify the task, but combining different statically linked modules together into a single Python interpreter still remains magical and somewhat problematic (especially since we are creating an application that changes a lot--not merely trying to patch the Python core with a new extension module).

A second problem is that of education--we have found that some users have a very difficult time making sense of the system (even we're somewhat amazed that it really works). Interestingly enough, this does not seem to be directly related to the use of C code or Python for that matter, but to issues related to the configuration, compilation, debugging, and installation of modules. While most users have written programs before, they have never dealt with shared libraries, automated code generators, high-level languages, and third-party packages such as Python. In other cases, there seems to be a "fear of autoconf"--if a program requires a configure script, it must be too complicated to understand (which is not the case here). The combination of these factors has led some users to believe that the system is far more complicated and fragile than it really is. We're not sure how to combat this problem other than to try and educate the users (i.e. it's essentially the same code as before, but with a really cool hack).

Conclusions and Future Work

Python is cool and we plan to keep using it. However much work remains. One of the more exciting aspects of Python is its solid support for modular programming and the culture of cooperation in the Python community. There are already many Python-related scientific computing projects underway. If these efforts were coordinated in some manner, we feel that the results could be spectacular.

Acknowledgments

We would like acknowledge our collaborators Brad Holian, Tim Germann, Shujia Zhou, and Wanshu Huang at Los Alamos National Laboratory. We would also like to acknowledge Paul Dubois and Brian Yang at Lawrence Livermore National Laboratory for many interesting conversations concerning Python and its use in physics applications. We would also like to acknowledge the Scientific Computing and Imaging Group at the University of Utah for their continued support. Finally, we'd like to thank the entire Python development community for making a truly awesome tool for solving real problems. Development of the SPaSM code has been performed under the auspices of the United States Department of Energy.

References

[1] D.M.Beazley and P.S. Lomdahl, "Message Passing Multi-Cell Molecular Dynamics on the Connection Machine 5," Parallel Computing 20 (1994), p. 173-195.

[2] P.S.Lomdahl, P.Tamayo, N.Gronbech-Jensen, and D.M.Beazley, "50 Gflops Molecular Dynamics on the CM-5," Proceedings of Supercomputing 93, IEEE Computer Society (1993), p.520-527.

[3] S.J.Zhou, D.M. Beazley, P.S. Lomdahl, B.L. Holian, Physical Review Letters. 78, 479 (1997).

[4] P. Dubois, K. Hinsen, and J. Hugunin, Computers in Physics (10), (1996) p. 262-267.

[5] D. M. Beazley and P.S. Lomdahl, "Extensible Message Passing Application Development and Debugging with Python", Proceedings of IPPS'97, Geneva Switzerland, IEEE Computer Society (1997).

[6] T.-Y.B. Yang, D.M. Beazley, P. F. Dubois, G. Furnish, "Steering Object-oriented Computations with Python", Python Workshop 5, Washington D.C., Nov. 4-5, 1996.

[7] D.M. Beazley and P.S. Lomdahl, "High Performance Molecular Dynamics Modeling with SPaSM : Performance and Portability Issues," Proceedings of the Workshop on Debugging and Tuning for Parallel Computer Systems, Chatham, MA, 1994 (IEEE Computer Society Press, Los Alamitos, CA, 1996), pp. 337-351.

[8] D.M. Beazley, "Using SWIG to Control, Prototype, and Debug C Programs with Python", Proceedings of the 4th International Python Conference, Lawrence Livermore National Laboratory, June 3-6, 1996.

[9] K. Hinsen, The Molecular Modeling Toolkit, http://starship.skyport.net/crew/hinsen/mmtk.html

[10] Jim Hugunin, Numeric Python, http://www.sls.lcs.mit.edu/jjh/numpy

[11] Fredrik Lundh, Python Imaging Library, http://www.python.org/sigs/image-sig/Imaging.html