This chapter explains the use of threads by MMTK and MMTK's parallelization support. This is an advanced topic, and not essential for the majority MMTK applications. You need to read this chapter only if you use multiprocessor computers, or if you want to implement multi-threaded programs that use MMTK.
Threads are different execution paths through a program that are executed in parallel, at least in principle; real parallel execution is possible only on multiprocessor systems. MMTK makes use of threads in two ways, which are conceptually unrelated: parallelization of energy evaluation on shared-memory multiprocessor computers, and support for multithreaded applications. Thread support is not available on all machines; you can check if yous system supports threads by starting a Python interpreter and typing import threading. If this produces an error message, then your system does not support threads, otherwise it is available in Python and also in MMTK. If you do not have thread support in Python although you know that your operating system supports threads, you might have compiled your Python interpreter without thread support; in that case, MMTK does not have thread support either.
Another approach to parallelization is message passing: several processors work on a program and communicate via a fast network to share results. A standard library, called MPI (Message Passing Interface), has been developped for sharing data by message passing, and implementations are available for all parallel computers currently on the market. MMTK contains elementary support for parallelization by message passing: only the energy evaluation has been paralellized, using a data-replication strategy, which is simple but not the most efficient for large systems. MPI support is disabled by default. Enabling it involves modifying the file Src/Setup.template prior to compilation of MMTK. Furthermore, an MPI-enabled installation of ScientificPython is required, and the mpipython executable must be used instead of the standard Python interpreter.
Threads and message passing can be used together to use a cluster of shared-memory machines most efficiently. However, this requires that the thread and MPI implementations being used work together; sometimes there are conflicts, for example due to the use of the same signal in both libraries. Refer to your system documentation for details.
The use of threads for parallelization on shared-memory systems is very simple: Just set the environment variable MMTK_ENERGY_THREADS to the desired value. If this variable is not defined, the default value is 1, i.e. energy evaluations are performed serially. For choosing an appropriate value for this environment variable, the following points should be considered:
The number of energy evaluation threads should not be larger than the number of processors that are fully dedicated to the MMTK application. A larger number of threads does not lead to wrong results, but it can increase the total execution time.
MMTK assumes that all processors are equally fast. If you use a heteregenous multiprocessor machine, in which the processors have different speeds, you might find that the total execution time is larger than without threads.
The use of threads incurs some computational overhead. For very small systems, it might be faster not to use threads.
Not all energy terms necessarily support threads. Of the force field terms that part of MMTK, only the multipole algorithms for electrostatic interactions does not support threads, but additional force fields defined outside MMTK might also be affected. MMTK automatically evaluates such energy terms with a single thread, such that there is no risk of getting wrong results. However, you might not get the performance you expect.
If second derivatives of the potential energy are requested, energy evaluation is handled by a single thread. An efficient implementation of multi-threaded energy evaluation would require a separate copy of the second-derivative matrix per thread. This approach needs too much memory for big systems to be feasible. Since second derivatives are almost exclusively used for normal mode calculations, which need only a single energy evaluation, multi-thread support is not particularly important anyway.
Parallelization via message passing is somewhat more complicated. In the current MMTK parallelization model, all processors execute the same program and replicate all tasks, with the important exception of energy evaluation. Energy terms are divided evenly between the processors, and at the end the energy and gradient values are shared by all machines. This is the only step involving network communication. Like thread-based parallelization, message-passing parallelization does not support the evaluation of second derivatives.
A special problem with message-passing systems is input and output. The MMTK application must ensure that output files are written by only one processor, and that all processors correctly access input files, especially in the case of each processor having its own disk space. See the example MPI/md.py for illustration.
Multithreaded applications are applications that use multiple threads in order to simplify the implementation of certain algorithms, i.e. not necessarily with the goal of profiting from multiple processors. If you plan to write a multithreaded application that uses MMTK, you should first make sure you understand threading support in Python. In particular, you should keep in mind that the global interpreter lock prevents the effective use of multiple processors by Python code; only one thread at a time can execute interpreted Python code. C code called from Python can permit other threads to execute simultaneously; MMTK does this for energy evaluation, molecular dynamics integration, energy minimization, and normal mode calculation.
A general problem in multithreaded applications is access to resources that are shared among the threads. In MMTK applications, the most important shared resource is the description of the chemical systems, i.e. universe objects and their contents. Chaos would result if two threads tried to modify the state of a universe simultaneously, or even if one thread uses information that is simultaneously being modified by another thread. Synchronization is therefore a critical part of multithreaded application. MMTK provides two synchronization aids, both of which described in the documentation of the class MMTK.Universe.Universe: the configuration change lock (methods acquireConfigurationChangeLock and releaseConfigurationChangeLock), and the universe state lock (methods acquireReadStateChangeLock, releaseReadStateChangeLock, acquireWriteStateChangeLock, and releaseWriteStateChangeLock). Only a few common universe operations manipulate the universe state lock in order to avoid conflicts with other threads; these methods are marked as thread-safe in the description. All other operations should only be used inside a code section that is protected by the appropriate manipulation of the state lock. The configuration change lock is less critical; it is used only by the molecular dynamics and energy minimization algorithms in MMTK.