One good way to appreciate the impact and significance of the multi-core revolution is to examine its effect on software packages that are widely familiar. The LAPACK/ScaLAPACK libraries for linear algebra fit that description. These libraries, which embody much of our work in the adaptation of block partitioned algorithms to parallel linear algebra software design, have served the HPC and Computational Science community remarkably well for 20 years. Both LAPACK and ScaLAPACK apply the idea of blocking in a consistent way to a wide range of algorithms in linear algebra (LA), including linear systems, least square problems, singular value decomposition, eigenvalue decomposition, etc., for problems with dense and banded coefficient matrices. ScaLAPACK also addresses the much harder problem of implementing these routines on top of distributed memory architectures. Yet it manages to keep close correspondence to LAPACK in the way the code is structured or organized. The design of these packages has had a major impact on how mathematical software has been written and used successfully during that time. Yet, when you look at how these foundational libraries can be expected to fair on large-scale multi-core systems, it becomes clear that we are on the verge of a transformation in software design, at least as potent as the change engendered a decade ago by message passing architectures, when the community had to rethink and rewrite many of its algorithms, libraries, and applications.
Historically, LA methods have put a strong emphasis on weak scaling or isoscaling of algorithms, where speed is achieved when the number of processors are increased while the problem size per processor is kept constant, effectively increasing the overall problem size.4 This measure tells us when we can exploit parallelism to solve larger problems. In this approach, increasing the speed of a single processing element should decrease the time to solution. But in the emerging era of multiprocessors, although the number of processing elements (i.e., cores) in systems will grow rapidly (exponentially, at least for a few generations), the computational power of individual processing units is likely to be reduced. Since the speed of individual processors will decline, we should expect that problems reaching their scaling limits on a certain number of processors will require increased time to solution on the next generation of architectures. In order to address the problem, emphasis has to be shifted from weak to strong scaling, where speed is achieved when the number of processors increased while the overall problem size is kept constant, which effectively decreases the problem size per processor. But to achieve this goal, we have to examine methods to exploit parallelization at much finer granularity than traditional approaches employ.
The standard approach to parallelization of numerical linear algebra algorithms for both shared and distributed memory systems, utilized by the LAPACK/ScaLAPACK libraries, is to rely on a parallel implementation of BLAS - threaded BLAS for shared memory systems and PBLAS for distributed memory systems. Historically, this approach made the job of writing hundreds of routines in a consistent and accessible manner doable. But although this approach solves numerous complexity problems, it also enforces a very rigid and inflexible software structure, where, at the level of LA, the algorithms are expressed in a serial way. This obviously inhibits the opportunity to exploit inherently parallel algorithms at finer granularity. This is shown by the fact that the traditional method is successful mainly in extracting parallelism from Level 3 BLAS; in the case of most of the Level 1 and 2 BLAS, however, it usually fails to achieve speedups and often results in slowdowns. It relies on the fact that, for large enough problems, the O(n3) cost of Level 3 BLAS dominates the computation and renders the remaining operations negligible. The problem with encapsulating parallelization in the BLAS/PBLAS in this way is that it requires a heavy synchronization model on a shared memory system and a heavily synchronous and blocking form of collective communication on distributed memory systems with message passing. This paradigm will break down on the next generation architectures, because it relies on coarse grained parallelization and emphasizes weak scaling, rather than strong scaling.