Database Reference
In-Depth Information
High
Back-end approach
Data parallelism
C/C++/Fortran with MPI
RScaLAPACK (Samatova et al, 2005)
pR
Automatic parallelization
Task parallelism
Task-pR (Samatova et al, 2004)
Embarrassing parallelism
Data parallelism
Snow (Tierney, Rossini,
Li, Sevcikova, 2006)
Manual parallelization
Message passing
Rmpi (Hao Yu, 2006)
rpvm (Na Li & Tony Rossini, 2006)
Compiled approach
MATLAB C automatic parallelization
→
→
Low
High
Parallel Performance
Figure 8.8
Evolution of parallel computing with R.
between
R
processes. It essentially provides wrappers for message passing
interface (MPI) and parallel virtual machine (PVM) routines realized through
R add-on libraries like
Rmpi
89
and
rpvm
.
90
While flexible for writing almost
any parallel program in
R
, these approaches lack eciency due to the inter-
preted nature of the underlying programming language. These libraries are
not transparent to the users with limited knowledge of parallel computing.
The second approach, such as the
snow
library,
91
offers a capability to ad-
dress embarrassingly parallel statistical computations. While simple to use, it
is limited to computations that require no coordination between R processes:
All processes work on their local data, perform exactly the same function, and
return the result to the parent process.
In contrast, our approach with
pR
addresses the above issues by pro-
viding a framework that aims to automatically and eciently exploit both
data parallelism
(e.g., cluster analysis, principal component analysis, correla-
tion) and
task parallelism
(e.g., likelihood maximization, bootstrap sampling,
Markov Chain Monte Carlo, animations) (Figure 8.9). With a data parallel
approach, the given data is divided among various processes that perform
more or less the same task on their data chunks to obtain the desired re-
sult. Demonstration of this capability is realized through an
RScaLAPACK
library,
92
,
93
which provides hooks to the ScaLAPACK
94
routines for paral-
lel, optimized, and portable linear algebra solvers.
RScaLAPACK
manages
