starpu-max/doc/chapters/introduction.texi

@c -*-texinfo-*-

@c This file is part of the StarPU Handbook.
@c Copyright (C) 2009--2011  Universit@'e de Bordeaux 1
@c Copyright (C) 2010, 2011, 2012  Centre National de la Recherche Scientifique
@c Copyright (C) 2011 Institut National de Recherche en Informatique et Automatique
@c See the file starpu.texi for copying conditions.

@menu
* Motivation::                  Why StarPU ?
* StarPU in a Nutshell::        The Fundamentals of StarPU
@end menu

@node Motivation
@section Motivation

@c complex machines with heterogeneous cores/devices
The use of specialized hardware such as accelerators or coprocessors offers an
interesting approach to overcome the physical limits encountered by processor
architects. As a result, many machines are now equipped with one or several
accelerators (e.g. a GPU), in addition to the usual processor(s). While a lot of
efforts have been devoted to offload computation onto such accelerators, very
little attention as been paid to portability concerns on the one hand, and to the
possibility of having heterogeneous accelerators and processors to interact on the other hand.

StarPU is a runtime system that offers support for heterogeneous multicore
architectures, it not only offers a unified view of the computational resources
(i.e. CPUs and accelerators at the same time), but it also takes care of
efficiently mapping and executing tasks onto an heterogeneous machine while
transparently handling low-level issues such as data transfers in a portable
fashion.

@c this leads to a complicated distributed memory design
@c which is not (easily) manageable by hand

@c added value/benefits of StarPU
@c   - portability
@c   - scheduling, perf. portability

@node StarPU in a Nutshell
@section StarPU in a Nutshell

@menu
* Codelet and Tasks::           
* StarPU Data Management Library::  
* Glossary::
* Research Papers::
@end menu

From a programming point of view, StarPU is not a new language but a library
that executes tasks explicitly submitted by the application.  The data that a
task manipulates are automatically transferred onto the accelerator so that
the programmer does not have to take care of complex data movements.  StarPU
also takes particular care of scheduling those tasks efficiently and allows
scheduling experts to implement custom scheduling policies in a portable
fashion. The target audience is typically developers of compilers or computation
libraries which want to seamlessly extend them to support heterogeneous
architectures.

@c explain the notion of codelet and task (i.e. g(A, B)
@node Codelet and Tasks
@subsection Codelet and Tasks

One of the StarPU primary data structures is the @b{codelet}. A codelet describes a
computational kernel that can possibly be implemented on multiple architectures
such as a CPU, a CUDA device or a Cell's SPU.

@c TODO insert illustration f : f_spu, f_cpu, ...

Another important data structure is the @b{task}. Executing a StarPU task
consists in applying a codelet on a data set, on one of the architectures on
which the codelet is implemented. A task thus describes the codelet that it
uses, but also which data are accessed, and how they are
accessed during the computation (read and/or write).
StarPU tasks are asynchronous: submitting a task to StarPU is a non-blocking
operation. The task structure can also specify a @b{callback} function that is
called once StarPU has properly executed the task. It also contains optional
fields that the application may use to give hints to the scheduler (such as
priority levels).

By default, task dependencies are inferred from data dependency (sequential
coherence) by StarPU. The application can however disable sequential coherency
for some data, and dependencies be expressed by hand.
A task may be identified by a unique 64-bit number chosen by the application
which we refer as a @b{tag}.
Task dependencies can be enforced by hand either by the means of callback functions, by
submitting other tasks, or by expressing dependencies
between tags (which can thus correspond to tasks that have not been submitted
yet).

@c TODO insert illustration f(Ar, Brw, Cr) + ..

@c DSM
@node StarPU Data Management Library
@subsection StarPU Data Management Library

Because StarPU schedules tasks at runtime, data transfers have to be
done automatically and ``just-in-time'' between processing units,
relieving the application programmer from explicit data transfers.
Moreover, to avoid unnecessary transfers, StarPU keeps data
where it was last needed, even if was modified there, and it
allows multiple copies of the same data to reside at the same time on
several processing units as long as it is not modified.

@node Glossary
@subsection Glossary

A @b{codelet} records pointers to various implementations of the same
theoretical function.

A @b{memory node} can be either the main RAM or GPU-embedded memory.

A @b{bus} is a link between memory nodes.

A @b{data handle} keeps track of replicates of the same data (@b{registered} by the
application) over various memory nodes. The data management library manages
keeping them coherent.

The @b{home} memory node of a data handle is the memory node from which the data
was registered (usually the main memory node).

A @b{task} represents a scheduled execution of a codelet on some data handles.

A @b{tag} is a rendez-vous point. Tasks typically have their own tag, and can
depend on other tags. The value is chosen by the application.

A @b{worker} execute tasks. There is typically one per CPU computation core and
one per accelerator (for which a whole CPU core is dedicated).

A @b{driver} drives a given kind of workers. There are currently CPU, CUDA,
OpenCL and Gordon drivers. They usually start several workers to actually drive
them.

A @b{performance model} is a (dynamic or static) model of the performance of a
given codelet. Codelets can have execution time performance model as well as
power consumption performance models.

A data @b{interface} describes the layout of the data: for a vector, a pointer
for the start, the number of elements and the size of elements ; for a matrix, a
pointer for the start, the number of elements per row, the offset between rows,
and the size of each element ; etc. To access their data, codelet functions are
given interfaces for the local memory node replicates of the data handles of the
scheduled task.

@b{Partitioning} data means dividing the data of a given data handle (called
@b{father}) into a series of @b{children} data handles which designate various
portions of the former.

A @b{filter} is the function which computes children data handles from a father
data handle, and thus describes how the partitioning should be done (horizontal,
vertical, etc.)

@b{Acquiring} a data handle can be done from the main application, to safely
access the data of a data handle from its home node, without having to
unregister it.


@node Research Papers
@subsection Research Papers

Research papers about StarPU can be found at

@indicateurl{http://runtime.bordeaux.inria.fr/Publis/Keyword/STARPU.html}

Notably a good overview in the research report

@indicateurl{http://hal.archives-ouvertes.fr/inria-00467677}