With the advent of multicore machines as standard practice from laptops to supercomputers, the issue of a hybrid MPI-thread (for example, MPI-OpenMP) "version" of PETSc is of interest. In this model, one MPI process per CPU (node) and several threads on that CPU (node) work together inside that MPI process.

The core PETSc team has come to the consensus that pure MPI using neighborhood collectives and the judicious using of MPI shared memory (for data structures that you may not wish to have duplicated on each MPI process due to memory constraints) will provide the best performance for HPC simulation needs on current generation systems, next generation systems and exascale systems. It is also a much simpler programming model then MPI + threads (leading to simpler code).

Note that the PETSc team has no problems with proposals to replace the pure MPI programming model with a different programming model but only with an alternative that is demonstrably __better__, not with something more complicated that has not been demonstrated to be better nor that has any technical reason to be believed to be any better. Ever since the IBM SP2 in 1996 we've been told that "pure MPI won't scale to the next generation of machine", this has yet to be true and there is no reason to believe that it will be true.

Though the current and planned programming model for PETSc is pure MPI we do provide some limited support for use with the hybrid MPI-thread model. These are discussed below. Many people throw around the term "hybrid MPI-thread" as if it is a trivial change in the programming model. It is not -- major ramifications must be understood if such a hybrid approach is to be used successfully. Hybrid approaches can be developed in many ways that affect usability and performance.

The simple model of PETSc with threads is to contain all the thread operations inside the Mat and Vec classes, leaving the user's programming model identical to what it is today. This model can be done in two ways by having Vec and Mat class implementations that use

  1. OpenMP compiler directives to parallelize some of the methods
  2. POSIX threads (pthread) calls to parallelize some of the methods.

We tried this approach (with support for both OpenMP and pthreads) and found the code was never faster than pure MPI and cumbersome to use hence we have removed it.

An alternative simple model of PETSc with threads is to have individual threads (OpenMP or others) to each manage their own (sequential) PETSc objects (and each thread can interact only with its own objects). This is useful when one has many small systems (or sets of ODEs) that must be integrated in an "embarrassingly parallel" fashion.

To use this feature one must ./configure PETSc with the option --with-threadsafety --with-log=0 [--with-openmp or --download-concurrencykit]. src/ksp/ksp/tutorials/ex61f.F90 demonstrates how this may be used with OpenMP. The code uses a small number of #pragma omp critical in non-time-critical locations in the code and thus only works with OpenMP and not with pthreads.

A more complicated model of PETSc with threads, which would allow users to write threaded code that made PETSc calls, is not supported because PETSc is not currently thread-safe. Because the issues involved with toolkits and thread safety are complex, this short answer is almost meaningless. Thus, this page attempts to explain how threads and thread safety relate to PETSc. Note that we are discussing here only "software threads" as opposed to "hardware threads."

Threads are used in 2 main ways in HPC:

  1. Loop-level compiler control. The C/C++/FORTRAN compiler manages the dispatching of threads at the beginning of a loop, where each thread is assigned non-overlapping selections from the loop. OpenMP, for example, defines standards (compiler directives) for indicating to compilers how they should "thread parallelize" loops.
  2. User control. The programmer manages the dispatching of threads directly by assigning threads to tasks (e.g., a subroutine call). For example, consider POSIX threads (pthreads) or the user thread management in OpenMP.

Threads are merely streams of control and do not have any global data associated with them. Any global variables (e.g., common blocks in FORTRAN) are "shared" by all the threads; that is, any thread can access and change that data. In addition, any space allocated (e.g., in C with malloc or C++ with new) to which a thread has a reference can be read/changed by that thread. The only private data a thread has are the local variables in the subroutines that it has called (i.e., the stack for that thread) or local variables that one explicitly indicates are to be not shared in compiler directives.

In its simplest form, thread safety means that any memory (global or allocated) to which more than one thread has access, has some mechanism to ensure that the memory remains consistent when the various threads act upon it. This can be managed by simply associating a lock with each "memory" and making sure that each thread locks the memory before accessing it and unlocks when it has completed accessing the memory. In an object oriented library, rather than associating locks with individual data items, one can think about associating locks with objects; so that only a single thread can operate on an object at a time.

PETSc is not "generically" "thread-safe" because:

  1. All the PETSc objects created during a simulation do not have locks associated with them. Again, the reason is performance; having always to lock an object before using it, and then unlocking it at the end would have a large impact on performance. Even with very inexpensive locks, there will still likely be a few "hot-spots" that kill performance. For example, if four threads share a matrix and are each calling MatSetValues(), this will be a bottleneck with those threads constantly fighting over the data structure inside the matrix object.

Some concerns about a thread model for parallelism. A thread model for parallelism of numerical methods appears to be powerful for problems that can store their data in very simple (well controlled) data structures. For example, if field data is stored in a two-dimensional array, then each thread can be assigned a nonoverlapping slice of data to operate on. OpenMP makes managing much of this sort of thing reasonably straightforward.

When data must be stored in a more complicated opaque data structure (for example an unstructured grid or sparse matrix), it is more difficult to partition the data among the threads to prevent conflict and still achieve good performance. More difficult, but certainly not impossible. For these situations, perhaps it is more natural for each thread to maintain its own private data structure that is later merged into a common data structure. But to do this, one has to introduce a great deal of private state associated with the thread, i.e., it becomes more like a "light-weight process".

In conclusion, at least for the PETSc package, the concept of being thread-safe is not simple. It has major ramifications about its performance and how it would be used; it is not a simple matter of throwing a few locks around and then everything is honky-dory.

References: The Problem with Threads, Edward A. Lee, Technical Report No. UCB/EECS-2006-1 January 10, 2006

If you have any comments/brickbats on this summary, please direct them to [email protected]; we are interested in alternative viewpoints.