here - Stefan-Marr.de

2.4. Common Approaches to Concurrent and Parallel Programming
ations, but also includes collective operations such as scatter and gather. In
general, these operations have rendezvous semantics. Thus, every sending operation
requires a matching receive operation and vice versa. MPI-2 [Message
Passing Interface Forum, 2009] added operations for one-sided communication
to introduce a restricted notion of shared memory in addition to the main
message passing, for situations where rendezvous semantics are not flexible
enough.
While MPI is widely used, and still considered the standard in HPC applications
when it comes to performance, it is also criticized for not being
designed with productivity in mind [Lusk and Yelick, 2007]. Its low-level API
does not fit well with common HPC applications. In addition to offering comparably
low-level APIs and data types, there is a potential misfit between
message-passing-based programming and data-centric applications.
APGAS (Asynchronous PGAS) languages were conceived with the same
goals as PGAS languages to solve the problems MPI has with regard to programmer
productivity. The distinguishing feature of APGAS languages compared
to PGAS languages is the notion that all operations on remote memory
need to be realized via an asynchronous task that executes on the remote
memory. Thus, APGAS language try to make the cost of such remote memory
accesses more explicit than in PGAS languages. This programming model
is supposed to guide developers to utilize data locality efficiently and to structure
data and communication to reduce costly operations.
Languages such as X10 [Charles et al., 2005] and Habanero [Cavé et al.,
2010] realize that idea by making locality explicit as part of the type system.
X10’s specification [Saraswat et al., 2012] goes so far as to define remote accesses
as having a by-value semantics for the whole lexical scope. This results
in a programming model very similar to message-passing. X10 combines this
with a fork/join-like task-based parallelism model, which makes is a hybrid
language in terms of our categorization. X10 differentiates between different
places as its notion of locality. Across places, it enforces isolation, but inside a
single place it provides a programming model that corresponds to our definition
of communicating threads.
2.4.5. Data Parallelism
Approaches for data parallelism provide abstractions to handle data dependencies.
In general, the tendency in these approaches is to move from control
33