index.html 64.2 KB
Newer Older
1
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN" "http://www.w3.org/TR/REC-html40/loose.dtd">
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
2 3 4 5 6
<html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en">
<HEAD>
<meta http-equiv="content-type" content="text/html; charset=UTF-8" />
<TITLE>StarPU</TITLE>
<link rel="stylesheet" type="text/css" href="style.css" />
7
<link rel="Shortcut icon" href="http://www.inria.fr/extension/site_inria/design/site_inria/images/favicon.ico" type="image/x-icon" />
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
8 9 10 11
</HEAD>

<body>

12
<div class="title">
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
13
<h1><a href="./">StarPU</a></h1>
14 15
<h2>A Unified Runtime System for Heterogeneous Multicore Architectures</h2>
</div>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
16

17
<div class="menu">
18
<a href="https://team.inria.fr/storm/">STORM TEAM</a> |
19
&nbsp; &nbsp; &nbsp;
20
|
21
<a href="#overview">Overview</a> |
Nathalie Furmento's avatar
Nathalie Furmento committed
22
<a href="#news">News</a> |
23
<a href="#contact">Contact</a> |
Nathalie Furmento's avatar
Nathalie Furmento committed
24
<a href="people/">People</a> |
25
<a href="#features">Features</a> |
26
<a href="#software">Software</a> |
THIBAULT Samuel's avatar
THIBAULT Samuel committed
27
<a href="#tryit">Try it!</a> |
Nathalie Furmento's avatar
Nathalie Furmento committed
28
<a href="help/">Help</a> |
29
<a href="#publications">Publications</a> |
30
<a href="internships/">Jobs/Interns</a> |
31
<a href="files/">Download</a> |
THIBAULT Samuel's avatar
THIBAULT Samuel committed
32
<a href="market/">Market</a> |
Nathalie Furmento's avatar
Nathalie Furmento committed
33
<a href="tutorials/">Tutorials</a> |
34
<a href="http://starpu.gforge.inria.fr/testing/morse/master/">Benchmarks</a> |
35
<a href="https://gforge.inria.fr/plugins/mediawiki/wiki/starpu/index.php/Main_Page">Intranet</a>
Nathalie Furmento's avatar
Nathalie Furmento committed
36
</div>
37

38 39
<div class="section" id="overview">
<h3>Overview</h3>
40 41 42 43 44 45
  <p>
<span class="important">StarPU is a task programming library for hybrid architectures</span>
<ol>
<li><b>The application provides algorithms and constraints</b>
    <ul>
    <li>CPU/GPU implementations of tasks</li>
46
    <li>A graph of tasks, using either the StarPU's high level <b>GCC plugin</b> pragmas, StarPU's rich <b>C/C++ API</b>, or <b>OpenMP pragmas</b>.</li>
47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
    </ul>
<br>
</li>
<li><b>StarPU handles run-time concerns</b>
    <ul>
    <li>Task dependencies</li>
    <li>Optimized heterogeneous scheduling</li>
    <li>Optimized data transfers and replication between main memory and discrete memories</li>
    <li>Optimized cluster communications</li>
    </ul>
</li>
</ol>
</p>
<p>
<span class="important">Rather than handling low-level issues, <b>programmers can concentrate on algorithmic concerns!</b></span>
</p>

<p>
65 66 67 68 69 70
<span class="note">The StarPU documentation is available in
<a href="./doc/starpu.pdf">PDF</a> and in <a href="./doc/html/">HTML</a>.</span>
Please note that these documents are up-to-date with the latest release of
StarPU.
</p>
<p>
71 72
The latest documentation in <a href="./testing/master/doc/starpu.pdf">PDF</a>
and <a href="./testing/master/doc/html">HTML</a> is updated everyday, but covers
73
the latest developments which may not be available in the latest release.
74 75 76 77
</p>
</div>

<div class="section emphasize newslist" id="news">
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
78
<h3>News</h3>
79 80 81 82 83 84 85
<p>
  November
  2019 <b>&raquo;&nbsp;</b>.
  A <a href="/tutorials/2019-11-HPNS-Inria/">StarPU tutorial</a> will
  be given as part of the Inria automn school "High Performance
  Numerical Simulation".
</p>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
86
<p>
87 88 89 90 91 92 93 94 95 96 97 98
October
2019 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      release 1.3.3 of StarPU is now
      available!</b></a> The 1.3 release brings among other
      functionalities a MPI master-slave support, a tool to replay
      execution through SimGrid, a HDF5 implementation of the
      Out-of-core, a new implementation of StarPU-MPI on top of
      NewMadeleine, implicit support for asynchronous partition
      planning, a resource management module to share processor cores
      and accelerator devices with other parallel runtime systems, ...
</p>
<p>
99 100 101 102 103 104 105 106 107 108 109 110
June
2019 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      release 1.3.2 of StarPU is now
      available!</b></a> The 1.3 release brings among other
      functionalities a MPI master-slave support, a tool to replay
      execution through SimGrid, a HDF5 implementation of the
      Out-of-core, a new implementation of StarPU-MPI on top of
      NewMadeleine, implicit support for asynchronous partition
      planning, a resource management module to share processor cores
      and accelerator devices with other parallel runtime systems, ...
</p>
<p>
111 112 113 114 115 116 117
May 2019 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      v1.1.8 release of StarPU is now available!</b></a>. This release notably brings the concept of
      scheduling contexts which allows to separate computation
      resources. This is really intented to be the last release for the
      branch 1.1.
</p>
<p>
Nathalie Furmento's avatar
Nathalie Furmento committed
118 119 120 121 122 123 124 125 126 127 128 129
April
2019 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      release 1.3.1 of StarPU is now
      available!</b></a> The 1.3 release brings among other
      functionalities a MPI master-slave support, a tool to replay
      execution through SimGrid, a HDF5 implementation of the
      Out-of-core, a new implementation of StarPU-MPI on top of
      NewMadeleine, implicit support for asynchronous partition
      planning, a resource management module to share processor cores
      and accelerator devices with other parallel runtime systems, ...
</p>
<p>
130
March
131
2019 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
132
      release 1.3.0 of StarPU is now
133 134 135 136 137 138 139 140 141
      available!</b></a> The 1.3 release brings among other
      functionalities a MPI master-slave support, a tool to replay
      execution through SimGrid, a HDF5 implementation of the
      Out-of-core, a new implementation of StarPU-MPI on top of
      NewMadeleine, implicit support for asynchronous partition
      planning, a resource management module to share processor cores
      and accelerator devices with other parallel runtime systems, ...
</p>
<p>
142 143 144 145 146 147 148 149
February 2019 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      1.2.8 release of StarPU is now available!</b></a>.
	The 1.2 release serie notably brings an out-of-core support, a MIC Xeon
	Phi support, an OpenMP runtime support, and a new internal
	communication system for MPI.
	(The release 1.2.7 is broken and should not be used)
</p>
<p>
150 151 152 153 154 155 156
September 2018 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      1.2.6 release of StarPU is now available!</b></a>.
      The 1.2 release serie notably brings an out-of-core support, a MIC Xeon
      Phi support, an OpenMP runtime support, and a new internal
      communication system for MPI.
</p>
<p>
157 158 159 160 161 162
August 2018 <b>&raquo;&nbsp;</b><a href="http://gforge.inria.fr/frs/?group_id=1570"><b>The
      1.2.5 release of StarPU is now available!</b></a>.
      The 1.2 release serie notably brings an out-of-core support, a MIC Xeon
      Phi support, an OpenMP runtime support, and a new internal
      communication system for MPI.
</p>
163
</div>
164 165

<div class="section emphasizebot" style="text-align: right; font-style: italic;">
166
Get the latest StarPU news by subscribing to the <a href="http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/starpu-announce">starpu-announce mailing list</a>.
167
See also the full <a href="news/">news</a>.
168 169
</div>

170 171 172 173 174 175 176 177 178 179 180
<div class="section" id="video">
<h3>Video Conference</h3>
<p>
A video recording (26') of a <a href=http://www.x.org/wiki/Events/XDC2014/XDC2014ThibaultStarPU/>presentation at the XDC2014 conference</a> gives an overview of StarPU
(<a href=http://www.x.org/wiki/Events/XDC2014/XDC2014ThibaultStarPU/xdc_starpu.pdf>slides</a>):
</p>
<center>
<iframe width="420" height="315" src="https://www.youtube.com/embed/frsWSqb8UJU" frameborder="0" allowfullscreen></iframe>
</center>
</div>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
181
<div class="section" id="tutorial">
182 183 184 185 186 187 188 189 190 191
<h3>Tutorial material</h3>
<p>
The latest tutorial material for StarPU is composed of two parts:
<ul>
<li><a href="http://starpu.gforge.inria.fr/tutorials/2016-06-PATC/slides/01_introducing_starpu.pdf">Introducing StarPU</a></li>
<li><a href="http://starpu.gforge.inria.fr/tutorials/2016-06-PATC/slides/02_mastering_starpu.pdf">Mastering StarPU</a></li>
</ul>
</p>
</div>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
192 193 194 195 196 197 198 199
<div class="section" id="slides">
<h3>Set of slides</h3>
<p>
A <a href="slides.pdf">set of slides</a> is also available to get an overview of
StarPU.
</p>
</div>

200 201
<div class="section" id="contact">
<h3>Contact</h3>
202
<p>For any questions regarding StarPU, please contact the StarPU developers mailing list.</p>
203 204 205
<pre>
<a href="mailto:starpu-devel@lists.gforge.inria.fr?subject=StarPU">starpu-devel@lists.gforge.inria.fr</a>
</pre>
206
<p>Details of the <a href="people/">StarPU team people</a> are also available.</p>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
207 208
</div>

209 210
<div class="section" id="features">
<h3>Features</h3>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
211

212
<h4>Portability</h4>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
213
  <p>
214 215 216 217 218 219 220 221
Portability is obtained by the means of a unified abstraction of the machine.
StarPU offers a unified offloadable task abstraction named <em>codelet</em>. Rather
than rewriting the entire code, programmers can encapsulate existing functions
within codelets. In case a codelet can run on heterogeneous architectures, <b>it
is possible to specify one function for each architectures</b> (e.g. one function
for CUDA and one function for CPUs). StarPU takes care of scheduling and
executing those codelets as efficiently as possible over the entire machine, include
multiple GPUs.
222 223 224 225
One can even specify <b>several functions for each architecture</b> (new in
v1.0) as well as
<b>parallel implementations</b> (e.g. in OpenMP), and StarPU will
automatically determine which version is best for each input size (new in v0.9).
226 227 228
StarPU can execute them concurrently, e.g. one per socket, provided that the
task implementations support it (which is the case for MKL, but unfortunately
most often not for OpenMP).
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
229 230
  </p>

231 232 233
<h4>Genericity</h4>
  <p>
The StarPU programming interface is very generic. For intance, various data
234
structures are supported mainline (vectors, dense matrices, CSR/BCSR/COO sparse matrices, ...),
235 236 237 238 239 240
but application-specific data structures can also be supported, provided that
the application describes how data is to be transfered (e.g. a series of
contiguous blocks). That was for instance used for hierarchically-compressed
matrices (h-matrices).
  </p>

241
<h4>Data transfers</h4>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
242
  <p>
243
To relieve programmers from the burden of explicit data transfers, a high-level
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
244
data management library enforces memory coherency over the machine: before a
245 246
codelet starts (e.g. on an accelerator), all its <b>data are automatically made
available on the compute resource</b>. Data are also kept on e.g. GPUs as long as
THIBAULT Samuel's avatar
THIBAULT Samuel committed
247 248
they are needed for further tasks. When a device runs out of memory, StarPU uses
an LRU strategy to <b>evict unused data</b>. StarPU also takes care of <b>automatically
249
prefetching</b> data, which thus permits to <b>overlap data transfers with computations</b>
THIBAULT Samuel's avatar
update  
THIBAULT Samuel committed
250
(including <b>GPU-GPU direct transfers</b>) to achieve the most of the architecture.
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
251 252
  </p>

253
<h4>Dependencies</h4>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
254
  <p>
255
Dependencies between tasks can be given either of several ways, to provide the
256 257
programmer with best flexibility:
  <ul>
258
    <li><b>implicitly</b> from RAW, WAW, and WAR data dependencies.</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
259
    <li>explicitly through <b>tags</b> which act as rendez-vous points between
260
    tasks (thus including tasks which have not been created yet),</li>
261
    <li><b>explicitly</b> between pairs of tasks,</li>
262
  </ul>
263 264
  </p>
  <p>
265 266 267
  These dependencies are computed in a completely decentralized way, and can be
  introduced completely dynamically as tasks get submitted by the application
  while tasks previously submitted are being executed.
268 269
  </p>
  <p>
270 271 272 273
StarPU also supports an OpenMP-like <a href="doc/html/DataManagement.html#DataReduction">reduction</a> access mode (new in v0.9).
  </p>
  <p>
It also supports a <a href="doc/html/DataManagement.html#DataCommute">commute</a> access mode to allow data access commutativity (new in v1.2).
274 275
  </p>

276 277 278 279 280
  <p>
It also supports transparent dependencies tracking between hierarchical subpieces of data
through asynchronous partitioning (new in v1.3).
  </p>

281 282 283
<h4>Heterogeneous Scheduling</h4>
  <p>
StarPU obtains
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
284
portable performances by efficiently (and easily) using all computing resources
285
at the same time. StarPU also takes advantage of the <b>heterogeneous</b> nature of a
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
286
machine, for instance by using scheduling strategies based on auto-tuned
287 288 289
performance models. These determine the relative performance achieved
by the different processing units for the various kinds of task, and thus
permits to <b>automatically let processing units execute the tasks they are the best for</b>.
290
Various strategies and variants are available. Some of them are centralized, but
291
most of them are <b>completely distributed</b>. <i>dmdas</i> (a data-locality-aware MCT strategy,
292
thus similar to heft but starts executing tasks before the whole task graph is
293 294 295
submitted, thus allowing dynamic task submission and a decentralized scheduler,
as well as an energy optimizing extension), <i>eager</i> (dumb centralized
queue), <i>lws</i> (decentralized locality-aware work-stealing), ...
296
The overhead per task is typically around the order of
297 298 299
magnitude of a microsecond. Tasks should thus be a few orders of magnitude
bigger, such as 100 microseconds or 1 millisecond, to make the overhead
negligible.
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
300 301
  </p>

302 303
<h4>Clusters</h4>
  <p>
304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325
To deal with clusters, StarPU can nicely integrate with <a
	href="doc/html/MPISupport.html">MPI</a>, through explicit or implicit
support, according to the application's preference.

    <ul>
        <li>Explicit network communication requests can be emitted, which will
then be <b>automatically combined and overlapped</b> with the intra-node data
transfers and computation,
        <li>The application can also just provide the whole task graph, a
data distribution over MPI nodes, and StarPU will automatically determine which
MPI node should execute which task, and <b>automatically generate all required
MPI communications</b> accordingly (new in v0.9). We have gotten excellent
scaling on a 256-node cluster with GPUs, we have not yet had the opportunity
to test on a yet larger cluster. We have however measured that with naive task
submission, it should scale to a thousand nodes, and with pruning-tuned task
submission, it should scale to about a <b>million nodes</b>.
        <li>Starting with v1.3, the application can also just provide the
whole task graph, and let StarPU decide the data distribution and task
distribution, thanks to a master-slave mechanism. This will however by nature
have a more limited scalability than the fully distributed paradigm mentioned
above.
    </ul>
326 327 328 329 330
  </p>

<h4>Out of core</h4>
  <p>
When memory is not big enough for the working set, one may have to resort to
Nathalie Furmento's avatar
Nathalie Furmento committed
331
using disks. StarPU makes this seamless thanks to its <a href="doc/html/OutOfCore.html">out of core support</a> (new in v1.2).
THIBAULT Samuel's avatar
update  
THIBAULT Samuel committed
332 333
StarPU will <b>automatically evict</b> data from the main memory in advance, and
<b>prefetch back</b> required data before it is needed for tasks.
334 335
  </p>

336
  <!--
337 338 339
<h4>Extensions to the C Language</h4>
<p>
  StarPU comes with a GCC plug-in
340
  that <a href="doc/html/cExtensions.html">extends the C programming
341 342
  language</a> with pragmas and attributes that make it easy
  to <b>annotate a sequential C program to turn it into a parallel
343 344
  StarPU program</b> (new in v1.0).
</p>
345 346 347 348 349 350
  -->

<h4>Fortran interface</h4>
<p>
  StarPU comes with native Fortran bindings and examples.
</p>
351

THIBAULT Samuel's avatar
THIBAULT Samuel committed
352 353
<h4>OpenMP 4 -compatible interface</h4>
<p>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
354
  <a href=http://kstar.gforge.inria.fr/>K'Star</a> provides an OpenMP
THIBAULT Samuel's avatar
THIBAULT Samuel committed
355 356 357 358 359 360 361 362 363
  4 -compatible interface on top of StarPU. This allows to just rebuild OpenMP
  applications with the K'Star source-to-source compiler, then build it with the
  usual compiler, and the result will use the StarPU runtime.
</p>
<p>
  K'Star also provides some extensions to the OpenMP 4 standard, to let the
  StarPU runtime perform online optimizations.
</p>

364 365 366 367
<h4>OpenCL-compatible interface</h4>
<p>
  StarPU provides an <a href="doc/html/SOCLOpenclExtensions.html">OpenCL-compatible interface, SOCL</a>
  which allows to simply run OpenCL applications on top of StarPU (new in v1.0).
368 369
</p>

370 371 372 373
<h4>Simulation support</h4>
<p>
  StarPU can very accurately simulate an application execution
  and measure the resulting performance thanks to using the
374
  <a href="http://simgrid.gforge.inria.fr">SimGrid simulator</a> (new in v1.1).  This allows
375 376 377 378 379
  to quickly experiment with various scheduling heuristics, various application
  algorithms, and even various platforms (available GPUs and CPUs, available
  bandwidth)!
</p>

380 381
<h4>All in all</h4>
  <p>
382
All that means that, with the help
383
of <a href="doc/html/cExtensions.html">StarPU's extensions to the C
384 385
language</a>, the following sequential source code of a tiled version of
the classical Cholesky factorization algorithm using BLAS is also valid
THIBAULT Samuel's avatar
THIBAULT Samuel committed
386
StarPU code, possibly running on all the CPUs and GPUs, and given a data
Nathalie Furmento's avatar
Nathalie Furmento committed
387
distribution over MPI nodes, it is even a distributed version!
388 389 390 391 392 393 394 395 396 397 398 399 400
  </p>

  <tt><pre>
for (k = 0; k < tiles; k++) {
  potrf(A[k,k])
  for (m = k+1; m < tiles; m++)
    trsm(A[k,k], A[m,k])
  for (m = k+1; m < tiles; m++)
    syrk(A[m,k], A[m, m])
  for (m = k+1, m < tiles; m++)
    for (n = k+1, n < m; n++)
      gemm(A[m,k], A[n,k], A[m,n])
}</pre></tt>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
401

402
<h4>Supported Architectures</h4>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
403
<ul>
404
<li>SMP/Multicore Processors (x86, PPC, ARM, ... all Debian architecture have been tested) </li>
Nathalie Furmento's avatar
Nathalie Furmento committed
405
<li>NVIDIA GPUs (e.g. heterogeneous multi-GPU), with pipelined and concurrent kernel execution support (new in v1.2) and GPU-GPU direct transfers (new in v1.1)</li>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
406 407
<li>OpenCL devices</li>
<li>Cell Processors (experimental)</li>
Nathalie Furmento's avatar
Nathalie Furmento committed
408 409
<li>Intel SCC (experimental, new in v1.2)</li>
<li>Intel MIC / Xeon Phi (new in v1.2)</li>
410
</ul>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
411

412
<h4>Supported Operating Systems</h4>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
413
<ul>
Ludovic Courtès's avatar
Ludovic Courtès committed
414 415
<li>GNU/Linux</li>
<li>Mac OS X</li>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
416 417 418
<li>Windows</li>
</ul>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
419 420 421 422 423 424 425 426 427 428 429
<h4>Stability</h4>
<p>
StarPU is checked every night with
<ul>
<li>Valgrind / Helgrind</li>
<li>gcc' Address/Leak/Thread/Undefined Sanitizers</li>
<li>cppcheck</li>
<li>Coverity</li>
</ul>
</p>

430
<h4>Performance analysis tools</h4>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
431 432 433 434 435 436 437 438 439
  <p>
In order to understand the performance obtained by StarPU, it is helpful to
visualize the actual behaviour of the applications running on complex
heterogeneous multicore architectures.  StarPU therefore makes it possible to
generate Pajé traces that can be visualized thanks to the <a
href="http://vite.gforge.inria.fr/"><b>ViTE</b> (Visual Trace Explorer) open
source tool.</a>
  </p>

440
<p>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
441 442 443 444 445 446
<b>Example:</b> LU decomposition on 3 CPU cores and a GPU using a very simple
greedy scheduling strategy. The green (resp. red) sections indicate when the
corresponding processing unit is busy (resp. idle). The number of ready tasks
is displayed in the curve on top: it appears that with this scheduling policy,
the algorithm suffers a certain lack of parallelism. <b>Measured speed: 175.32
GFlop/s</b>
447
<center><a href="./images/greedy-lu-16k-fx5800.png"> <img src="./images/greedy-lu-16k-fx5800.png" alt="LU decomposition (greedy)" width="75%"></a></center>
448 449
</p>

Nathalie Furmento's avatar
website  
Nathalie Furmento committed
450 451 452 453 454 455
<p>
This second trace depicts the behaviour of the same application using a
scheduling strategy trying to minimize load imbalance thanks to auto-tuned
performance models and to keep data locality as high as possible. In this
example, the Pajé trace clearly shows that this scheduling strategy outperforms
the previous one in terms of processor usage. <b>Measured speed: 239.60
456
GFlop/s</b>
457
<center><a href="./images/dmda-lu-16k-fx5800.png"><img src="./images/dmda-lu-16k-fx5800.png" alt="LU decomposition (dmda)" width="75%"></a></center>
Nathalie Furmento's avatar
website  
Nathalie Furmento committed
458 459
</p>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
460 461
<p>
<a href="http://www.hlrs.de/temanejo">Temanejo</a> can be used to debug the task
462
graph, as shown below (new in v1.1).
THIBAULT Samuel's avatar
THIBAULT Samuel committed
463 464 465
</p>

<center>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
466
<a href="images/temanejo.png"><img src="images/temanejo.png" width="50%"/></a>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
467 468
</center>

Nathalie Furmento's avatar
website  
Nathalie Furmento committed
469 470
</div>

471 472 473 474 475
<div class="section" id="software">
<h3>Software using StarPU</h3>

<p>
Some software is known for being able to use StarPU to tackle heterogeneous
476
architectures, here is a non-exhaustive list (feel free to ask to be added to the
477
list!):
478 479 480
</p>

<ul>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
481
	<li>AL4SAN, dense linear algebra library</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
482
	<li><a href="https://project.inria.fr/chameleon/">Chameleon</a>, dense linear algebra library</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
483
	<li><a href="http://exa2pro.eu">Exa2pro</a>, Enhancing Programmability and boosting Performance Portability for Exascale Computing Systems</li>
THIBAULT Samuel's avatar
fix URL  
THIBAULT Samuel committed
484
	<li><a href="http://github.com/ecrc/exageostat">ExaGeoStat</a>, Machine learning framework for Climate/Weather prediction applications</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
485
	<li><a href="https://hal.inria.fr/hal-01507613">FLUSEPA</a>, Navier-Stokes Solver for Unsteady Problems with Bodies in Relative Motion</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
486
	<li><a href="http://github.com/ecrc/hicma">HiCMA</a>, Low-rank general linear algebra library</li>
487
	<li>hmat, hierarchical matrix C/C++ library</li>
488
        <li><a href=http://kstar.gforge.inria.fr/>K'Star</a>, OpenMP 4 - compatible interface on top of StarPU.</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
489
	<li><a href="http://github.com/ecrc/ksvd">KSVD</a>, dense SVD on distributed-memory manycore systems</li>
490 491
	<li><a href="http://icl.cs.utk.edu/magma/">MAGMA</a>, dense linear algebra library, starting from version 1.1</li>
	<li><a href="https://gitlab.inria.fr/solverstack/maphys">MaPHyS</a>, Massively Parallel Hybrid Solver</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
492
	<li><a href="http://github.com/ecrc/moao">MOAO</a>, HPC framework for computational astronomy, servicing the European Extremely Large Telescope and the Japanese Subaru Telescope</li>
493
	<li><a href="http://pastix.gforge.inria.fr/">PaStiX</a>, sparse linear algebra library, starting from version 5.2.1</li>
494
	<li>PEPPHER, Performance Portability and Programmability for Heterogeneous Many-core Architectures</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
495
	<li><a href="http://github.com/ecrc/qdwh">QDWH</a>, QR-based Dynamically Weighted Halley</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
496
	<li><a href="http://buttari.perso.enseeiht.fr/qr_mumps/">qr_mumps</a>, sparse linear algebra library</li>
THIBAULT Samuel's avatar
update  
THIBAULT Samuel committed
497
	<li><a href="http://scalfmm-public.gforge.inria.fr/doc/">ScalFMM</a>, N-body interaction simulation using the Fast Multipole Method. </li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
498
	<li><a href="https://tel.archives-ouvertes.fr/tel-01410049/">SCHNAPS</a>, Solver for Conservative Hypebolic Non-linear systems Applied to PlasmaS. </li>
499 500
	<li><a href="https://hal.archives-ouvertes.fr/hal-01086246">SignalPU</a>, a Dataflow-Graph-specific programming model. </li>
	<li><a href="http://www.ida.liu.se/~chrke/skepu/">SkePU</a>, a skeleton programming framework.</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
501
	<li><a href="http://github.com/ecrc/stars-h">STARS-H</a>, HPC low-rank matrix market</li>
502
	<li><a href="http://www.xcalablemp.org/">XcalableMP</a>, Directive-based language eXtension for Scalable and performance-aware Parallel Programming</li>
503 504
</ul>

505
<p>
506
You can find <a href="#PublicationsOnApplications">below</a> the list of publications related to applications using StarPU.
507 508
</p>

509 510
</div>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
511 512 513 514
<div class="section" id="tryit">
<h3>Give it a try!</h3>
<p>
You can easily try the performance on the Cholesky factorization for
THIBAULT Samuel's avatar
THIBAULT Samuel committed
515 516 517
instance. Make sure to have the pkg-config and
<a href="http://www.open-mpi.org/projects/hwloc/">hwloc</a>
software installed for
THIBAULT Samuel's avatar
THIBAULT Samuel committed
518 519 520
proper CPU control and BLAS kernels for your computation units and configured in
your environment (e.g. MKL for CPUs and CUBLAS for GPUs).
</p>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
521 522

<tt><pre>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
523 524 525 526 527 528
$ wget http://starpu.gforge.inria.fr/files/starpu-someversion.tar.gz
$ tar xf starpu-someversion.tar.gz
$ cd starpu-someversion
$ ./configure
$ make -j 12
$ STARPU_SCHED=dmdas ./examples/cholesky/cholesky_implicit -size $((960*40)) -nblocks 40
THIBAULT Samuel's avatar
THIBAULT Samuel committed
529
$ STARPU_SCHED=dmdas mpirun -np 4 -machinefile mymachines ./mpi/examples/matrix_decomposition/mpi_cholesky_distributed -size $((960*40*4)) -nblocks $((40*4))</pre></tt>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
530 531 532 533 534

<p>Note that the dmdas scheduler uses performance models, and thus needs
calibration execution before exhibiting optimized performance (until the "model
something is not calibrated enough" messages go away).</p>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
535 536 537 538 539 540 541 542 543 544 545 546 547 548
<p>To get a glimpse at what happened, you can get an execution trace by
installing
<a href="http://savannah.nongnu.org/projects/fkt">FxT</a>
and <a href="http://vite.gforge.inria.fr/">ViTE</a>, and enabling traces:
</p>

<tt><pre>
$ ./configure --with-fxt
$ make -j 12
$ STARPU_SCHED=dmdas ./examples/cholesky/cholesky_implicit -size $((960*40)) -nblocks 40
$ ./tools/starpu_fxt_tool -i /tmp/prof_file_${USER}_0
$ vite paje.trace
</pre></tt>

THIBAULT Samuel's avatar
THIBAULT Samuel committed
549 550 551 552 553 554 555 556 557 558
<p>
Starting with StarPU 1.1, it is also possible to reproduce the performance that
we show in our articles on our machines, by installing simgrid, and then using
the simulation mode of StarPU using the performance models of our machines:
</p>
  <tt><pre>
$ ./configure --enable-simgrid
$ make -j 12
$ STARPU_PERF_MODEL_DIR=$PWD/tools/perfmodels/sampling STARPU_HOSTNAME=mirage STARPU_SCHED=dmdas ./examples/cholesky/cholesky_implicit -size $((960*40)) -nblocks 40
# size	ms	GFlops
THIBAULT Samuel's avatar
THIBAULT Samuel committed
559
38400	9915	1903.7</pre></tt>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
560 561
<p>(MPI simulation is not supported yet)</p>

562 563 564 565
<div class="section" id="publications">
<h3>Publications</h3>
<p>
All StarPU related publications are also
566
listed <a href="./publications">here</a>
567 568 569
with the corresponding Bibtex entries.
</p>

570 571
<p>
A good overview is available in
572 573 574
the following <a href="http://hal.archives-ouvertes.fr/inria-00467677">Research Report</a>.
</p>

575 576
<p>
If you need to cite StarPU, please
577
reference <a href="publications/Year/2011.html#AugThiNamWac11CCPE">[StarPU: A Unified Platform
578 579 580 581 582
    for Task Scheduling on Heterogeneous Multicore Architectures]</a>
for a general presentation. Other sub-sections below will give you
references for more specific aspects of StarPU.
</p>

583
<h4>General Presentations</h4> 
584
<a name="PublicationsGeneralPresentations"></a>
585 586
<ol>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
587 588 589 590 591 592 593
<a name="thibault:tel-01959127"></a>Samuel Thibault<br/>
<strong>On Runtime Systems for Task-based Programming on Heterogeneous Platforms</strong><br/>
Habilitation à diriger des recherches, Université de Bordeaux, December 2018<br/>
[<a href="https://hal.inria.fr/tel-01959127">WWW</a>]
[<a href="https://hal.inria.fr/tel-01959127/file/hdr.pdf">PDF</a>]
</li>
<li>
594 595 596
<a name="Aug11Thesis"></a>Cédric Augonnet<br/>
<strong>Scheduling Tasks over Multicore machines enhanced with Accelerators: a Runtime System's Perspective</strong><br/>
PhD thesis, Université Bordeaux 1, 351 cours de la Libération --- 33405 TALENCE cedex, December 2011<br/>
597
[<a href="http://tel.archives-ouvertes.fr/tel-00777154">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
598
[<a href="http://tel.archives-ouvertes.fr/tel-00777154/document">PDF</a>]
599 600
</li>
<li>
601 602
<a name="AugThiNamWac11CCPE"></a>Cédric Augonnet, Samuel Thibault, Raymond Namyst,  and Pierre-André Wacrenier<br/>
<strong>StarPU: A Unified Platform for Task Scheduling on Heterogeneous Multicore Architectures</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
603
<em>CCPE - Concurrency and Computation: Practice and Experience, Special Issue: Euro-Par 2009</em>, 23:187-198, February 2011<br/>
604
[<a href="http://hal.inria.fr/inria-00550877">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
605
[<a href="http://hal.inria.fr/inria-00550877/document">PDF</a>]
606
[doi:<a href="http://dx.doi.org/10.1002/cpe.1631">10.1002/cpe.1631</a>]
607
</li>
608
<li>
609 610
<a name="AugThiNamWac10RR7240"></a>Cédric Augonnet, Samuel Thibault,  and Raymond Namyst<br/>
<strong>StarPU: a Runtime System for Scheduling Tasks over Accelerator-Based Multicore Machines</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
611
Research Report RR-7240, INRIA, March 2010<br/>
612
[<a href="http://hal.inria.fr/inria-00467677">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
613
[<a href="http://hal.inria.fr/inria-00467677/document">PDF</a>]
614 615
</li>
<li>
616 617 618 619
<a name="Aug09Renpar19"></a>Cédric Augonnet<br/>
<strong>StarPU: un support exécutif unifié pour les architectures multicoeurs hétérogènes</strong><br/>
In <em>19èmes Rencontres Francophones du Parallélisme</em>, Toulouse / France, September 2009<br/>
Note: Best Paper Award<br/>
620
[<a href="http://hal.inria.fr/inria-00411581">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
621
[<a href="http://hal.inria.fr/inria-00411581/document">PDF</a>]
622 623
</li>
<li>
624 625
<a name="AugThiNamWac09Europar"></a>Cédric Augonnet, Samuel Thibault, Raymond Namyst,  and Pierre-André Wacrenier<br/>
<strong>StarPU: A Unified Platform for Task Scheduling on Heterogeneous Multicore Architectures</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
626
In <em>Euro-Par - 15th International Conference on Parallel Processing</em>, volume 5704 of <em>Lecture Notes in Computer Science</em>, Delft, The Netherlands, pages 863-874, August 2009<br/>
627
Springer<br/>
628
[<a href="http://hal.inria.fr/inria-00384363">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
629
[<a href="http://hal.inria.fr/inria-00384363/document">PDF</a>]
630 631 632 633 634 635 636 637
[doi:<a href="http://dx.doi.org/10.1007/978-3-642-03869-3_80">10.1007/978-3-642-03869-3_80</a>]
</li>
<li>
<a name="AugNam08HPPC"></a>Cédric Augonnet and Raymond Namyst<br/>
<strong>A unified runtime system for heterogeneous multicore architectures</strong><br/>
In <em>Proceedings of the International Euro-Par Workshops 2008, HPPC'08</em>, volume 5415 of <em>Lecture Notes in Computer Science</em>, Las Palmas de Gran Canaria, Spain, pages 174-183, August 2008<br/>
Springer<br/>
<strong>ISBN:</strong> 978-3-642-00954-9<br/>
638
[<a href="http://hal.inria.fr/inria-00326917">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
639
[<a href="http://hal.inria.fr/inria-00326917/document">PDF</a>]
640 641 642 643 644 645
[doi:<a href="http://dx.doi.org/10.1007/978-3-642-00955-6_22">10.1007/978-3-642-00955-6_22</a>]
</li>
<li>
<a name="Aug08Master"></a>Cédric Augonnet<br/>
<strong>Vers des supports d'exécution capables d'exploiter les machines multicoeurs hétérogènes</strong><br/>
Mémoire de DEA, Université Bordeaux 1, June 2008<br/>
646
[<a href="http://hal.inria.fr/inria-00289361">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
647
[<a href="http://hal.inria.fr/inria-00289361/document">PDF</a>]
648 649
</li>
</ol>
650
<h4>On Composability</h4> 
651
<a name="PublicationsOnComposability"></a>
Nathalie Furmento's avatar
Nathalie Furmento committed
652 653
<ol>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
654 655 656 657 658 659 660
<a name="hugo:tel-01162975"></a>Andra-Ecaterina Hugo<br/>
<strong>Composability of parallel codes on heterogeneous architectures</strong><br/>
Theses, Université de Bordeaux, December 2014<br/>
[<a href="https://tel.archives-ouvertes.fr/tel-01162975">WWW</a>]
[<a href="https://tel.archives-ouvertes.fr/tel-01162975/file/HUGO_ANDRA_2014.pdf">PDF</a>]
</li>
<li>
661 662 663
<a name="AH13Renpar"></a>Andra Hugo<br/>
<strong>Le problème de la composition parallèle : une approche supervisée</strong><br/>
In <em>21èmes Rencontres Francophones du Parallélisme (RenPar'21)</em>, Grenoble, France, January 2013<br/>
664
[<a href="http://hal.inria.fr/hal-00773610">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
665
[<a href="http://hal.inria.fr/hal-00773610/document">PDF</a>]
Nathalie Furmento's avatar
Nathalie Furmento committed
666 667
</li>
<li>
668 669 670
<a name="hugo:hal-00824514"></a>Andra Hugo, Abdou Guermouche, Raymond Namyst,  and Pierre-André Wacrenier<br/>
<strong>Composing multiple StarPU applications over heterogeneous machines: a supervised approach</strong><br/>
In <em>Third International Workshop on Accelerators and Hybrid Exascale Systems</em>, Boston, USA, May 2013<br/>
671
[<a href="http://hal.inria.fr/hal-00824514">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
672
[<a href="http://hal.inria.fr/hal-00824514/document">PDF</a>]
673 674 675 676 677
</li>
<li>
<a name="AH11Master"></a>Andra Hugo<br/>
<strong>Composabilité de codes parallèles sur architectures hétérogènes</strong><br/>
Mémoire de Master, Université Bordeaux 1, June 2011<br/>
678
[<a href="http://hal.inria.fr/inria-00619654/en/">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
679
[<a href="http://hal.inria.fr/inria-00619654/document">PDF</a>]
Nathalie Furmento's avatar
Nathalie Furmento committed
680 681
</li>
</ol>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
682 683 684 685 686 687 688 689 690 691 692
<h4>On Parallel Tasks</h4> 
<a name="PublicationsOnParallelTasks"></a>
<ol>
<li>
<a name="cojean:tel-01816341"></a>Terry Cojean<br/>
<strong>Programmation of heterogeneous architectures using moldable tasks</strong><br/>
Theses, Université de Bordeaux, March 2018<br/>
[<a href="https://tel.archives-ouvertes.fr/tel-01816341">WWW</a>]
[<a href="https://tel.archives-ouvertes.fr/tel-01816341/file/COJEAN_TERRY_2018.pdf">PDF</a>]
</li>
</ol>
693
<h4>On Scheduling</h4> 
694
<a name="PublicationsOnScheduling"></a>
695
<ol>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
696
<li>
697 698 699 700 701 702 703 704
<a name="bramas:hal-02120736"></a>Bérenger Bramas<br/>
<strong>Impact study of data locality on task-based applications through the Heteroprio scheduler</strong><br/>
<em>PeerJ Computer Science</em>, May 2019<br/>
[<a href="https://hal.inria.fr/hal-02120736">WWW</a>]
[<a href="https://hal.inria.fr/hal-02120736/file/peerj-cs-190.pdf">PDF</a>]
[doi:<a href="http://dx.doi.org/10.7717/peerj-cs.190">10.7717/peerj-cs.190</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
705 706 707 708 709 710 711 712 713
<a name="leandronesi:hal-02275363"></a>Lucas Leandro Nesi, Samuel Thibault, Luka Stanisic,  and Lucas Mello Schnorr<br/>
<strong>Visual Performance Analysis of Memory Behavior in a Task-Based Runtime on Hybrid Platforms</strong><br/>
In <em>2019 19th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)</em>, Larnaca, Cyprus, pages 142-151, May 2019<br/>
IEEE<br/>
[<a href="https://hal.inria.fr/hal-02275363">WWW</a>]
[<a href="https://hal.inria.fr/hal-02275363/file/CCGRID_camera_ready.pdf">PDF</a>]
[doi:<a href="http://dx.doi.org/10.1109/CCGRID.2019.00025">10.1109/CCGRID.2019.00025</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
714 715 716 717 718 719 720
<a name="alias:hal-02421327"></a>Christophe Alias, Samuel Thibault,  and Laure Gonnord<br/>
<strong>A Compiler Algorithm to Guide Runtime Scheduling</strong><br/>
Research Report RR-9315, INRIA Grenoble ; INRIA Bordeaux, December 2019<br/>
[<a href="https://hal.inria.fr/hal-02421327">WWW</a>]
[<a href="https://hal.inria.fr/hal-02421327/file/RR-9315.pdf">PDF</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
721 722
<a name="garciapinto:hal-01616632"></a>Vinicius Garcia Pinto, Lucas Mello Schnorr, Luka Stanisic, Arnaud Legrand, Samuel Thibault,  and Vincent Danjean<br/>
<strong>A Visual Performance Analysis Framework for Task-based Parallel Applications running on Hybrid Clusters</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
723
<em>CCPE - Concurrency and Computation: Practice and Experience</em>, 30, April 2018<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
724 725 726 727 728
[<a href="https://hal.inria.fr/hal-01616632">WWW</a>]
[<a href="https://hal.inria.fr/hal-01616632/file/CCPE_article_submitted_2018_02_06.pdf">PDF</a>]
[doi:<a href="http://dx.doi.org/10.1002/cpe.4472">10.1002/cpe.4472</a>]
</li>
<li>
729
<a name="pinto:hal-01842038"></a>Vinicius Garcia Pinto, Lucas Mello Schnorr, Arnaud Legrand, Samuel Thibault, Luka Stanisic,  and Vincent Danjean<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
730
<strong>Detecção de Anomalias de Desempenho em Aplicações de Alto Desempenho baseadas em Tarefas em Clusters Hìbridos</strong><br/>
THIBAULT Samuel's avatar
typo  
THIBAULT Samuel committed
731
In <em>WPerformance - 17o Workshop em Desempenho de Sistemas Computacionais e de Comunicação</em>, Natal, Brazil, July 2018<br/>
732 733 734 735
[<a href="https://hal.inria.fr/hal-01842038">WWW</a>]
[<a href="https://hal.inria.fr/hal-01842038/file/181587_1.pdf">PDF</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
736 737
<a name="kumar:tel-01538516"></a>Suraj Kumar<br/>
<strong>Scheduling of Dense Linear Algebra Kernels on Heterogeneous Resources</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
738
PhD thesis, Université de Bordeaux, April 2017<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
739 740 741 742
[<a href="https://tel.archives-ouvertes.fr/tel-01538516">WWW</a>]
[<a href="https://tel.archives-ouvertes.fr/tel-01538516/file/KUMAR_SURAL_2017.pdf">PDF</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
743 744 745 746 747 748 749 750
<a name="beaumont:hal-01386174"></a>O. Beaumont, L. Eyraud-Dubois,  and S. Kumar<br/>
<strong>Approximation Proofs of a Fast and Efficient List Scheduling Algorithm for Task-Based Runtime Systems on Multicores and GPUs</strong><br/>
In <em>2017 IEEE International Parallel and Distributed Processing Symposium (IPDPS)</em>, pages 768-777, May 2017<br/>
[<a href="https://hal.inria.fr/hal-01386174">WWW</a>]
[<a href="https://hal.inria.fr/hal-01386174/file/heteroPrioApproxProofsRR.pdf">PDF</a>]
[doi:<a href="http://dx.doi.org/10.1109/IPDPS.2017.71">10.1109/IPDPS.2017.71</a>]
</li>
<li>
751 752 753 754 755 756 757 758
<a name="agullo:hal-01223573"></a>Emmanuel Agullo, Olivier Beaumont, Lionel Eyraud-Dubois,  and Suraj Kumar<br/>
<strong>Are Static Schedules so Bad ? A Case Study on Cholesky Factorization</strong><br/>
In <em>IPDPS'16</em>, Proceedings of the 30th IEEE International Parallel & Distributed Processing Symposium, IPDPS'16, Chicago, IL, United States, May 2016<br/>
IEEE<br/>
[<a href="https://hal.inria.fr/hal-01223573">WWW</a>]
[<a href="https://hal.inria.fr/hal-01223573/file/heteroprioCameraReady-ieeeCompatiable.pdf">PDF</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
759 760 761 762
<a name="beaumont:hal-01361992"></a>Olivier Beaumont, Terry Cojean, Lionel Eyraud-Dubois, Abdou Guermouche,  and Suraj Kumar<br/>
<strong>Scheduling of Linear Algebra Kernels on Multiple Heterogeneous Resources</strong><br/>
In <em>International Conference on High Performance Computing, Data, and Analytics (HiPC)</em>, Hyderabad, India, December 2016<br/>
[<a href="https://hal.inria.fr/hal-01361992">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
763
[<a href="https://hal.inria.fr/hal-01361992v2/document">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
764 765
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
766 767
<a name="cojean:hal-01181135"></a>Terry Cojean, Abdou Guermouche, Andra Hugo, Raymond Namyst,  and Pierre-André Wacrenier<br/>
<strong>Resource aggregation for task-based Cholesky Factorization on top of heterogeneous machines</strong><br/>
768
In <em>HeteroPar'2016 workshop of Euro-Par</em>, Grenoble, France, August 2016<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
769 770
[<a href="https://hal.inria.fr/hal-01181135">WWW</a>]
[<a href="https://hal.inria.fr/hal-01181135/file/papier%20%281%29.pdf">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
771 772
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
773 774
<a name="garciapinto:hal-01353962"></a>Vinicius Garcia Pinto, Luka Stanisic, Arnaud Legrand, Lucas Mello Schnorr, Samuel Thibault,  and Vincent Danjean<br/>
<strong>Analyzing Dynamic Task-Based Applications on Hybrid Platforms: An Agile Scripting Approach</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
775
In <em>VPA - 3rd Workshop on Visual Performance Analysis</em>, Salt Lake City, United States, November 2016<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
776 777
Note: Held in conjunction with SC16<br/>
[<a href="https://hal.inria.fr/hal-01353962">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
778 779
[<a href="https://hal.inria.fr/hal-01353962v2/document">PDF</a>]
[doi:<a href="http://dx.doi.org/10.1109/VPA.2016.008">10.1109/VPA.2016.008</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
780
</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
781
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
782 783 784
<a name="JaBlHU2016a"></a>Johan Janzén, David Black-Schaffer,  and Andra Hugo<br/>
<strong>Partitioning GPUs for Improved Scalability</strong><br/>
In <em>IEEE 28th International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD)</em>, October 2016<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
785
[<a href="http://ieeexplore.ieee.org/abstract/document/7789322/">WWW</a>]
786 787 788
[doi:<a href="http://dx.doi.org/10.1109/SBAC-PAD.2016.14">10.1109/SBAC-PAD.2016.14</a>]
</li>
<li>
789 790 791 792 793 794 795
<a name="cojean:hal-01409965"></a>Terry Cojean, Abdou Guermouche, Andra Hugo, Raymond Namyst,  and Pierre-André Wacrenier<br/>
<strong>Resource aggregation for task-based Cholesky Factorization on top of modern architectures</strong><br/>
Note: This paper is submitted for review to the Parallel Computing special issue for HCW and HeteroPar 16 workshops, November 2016<br/>
[<a href="https://hal.inria.fr/hal-01409965">WWW</a>]
[<a href="https://hal.inria.fr/hal-01409965/file/submission.pdf">PDF</a>]
</li>
<li>
796 797
<a name="agullo:hal-01120507"></a>Emmanuel Agullo, Olivier Beaumont, Lionel Eyraud-Dubois, Julien Herrmann, Suraj Kumar, Loris Marchal,  and Samuel Thibault<br/>
<strong>Bridging the Gap between Performance and Bounds of Cholesky Factorization on Heterogeneous Platforms</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
798
In <em>HCW'2015 - Heterogeneity in Computing Workshop of IPDPS</em>, Hyderabad, India, May 2015<br/>
799
[<a href="https://hal.inria.fr/hal-01120507">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
800
[<a href="https://hal.inria.fr/hal-01120507/document">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
801
[doi:<a href="http://dx.doi.org/10.1109/IPDPSW.2015.35">10.1109/IPDPSW.2015.35</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
802
</li>
803
<li>
804 805 806
<a name="sergent:hal-00978364"></a>Marc Sergent and Simon Archipoff<br/>
<strong>Modulariser les ordonnanceurs de tâches : une approche structurelle</strong><br/>
In <em>Compas'2014</em>, Neuchâtel, Suisse, April 2014<br/>
807 808
[<a href="http://hal.inria.fr/hal-00978364">WWW</a>]
[<a href="http://hal.inria.fr/hal-00978364/PDF/ordonnanceurs_modulaires.pdf">PDF</a>]
809
</li>
810 811 812 813 814 815 816 817
<li>
<a name="AugCleThiNam10ICPADS"></a>Cédric Augonnet, Jérôme Clet-Ortega, Samuel Thibault,  and Raymond Namyst<br/>
<strong>Data-Aware Task Scheduling on Multi-Accelerator based Platforms</strong><br/>
In <em>The 16th International Conference on Parallel and Distributed Systems (ICPADS)</em>, Shanghai, China, December 2010<br/>
[<a href="http://hal.inria.fr/inria-00523937">WWW</a>]
[<a href="http://hal.inria.fr/inria-00523937/document">PDF</a>]
[doi:<a href="http://dx.doi.org/10.1109/ICPADS.2010.129">10.1109/ICPADS.2010.129</a>]
</li>
818
</ol>
819
<h4>On The C Extensions</h4> 
820
<a name="PublicationsOnTheCExtensions"></a>
821
<ol>
822 823 824 825
<li>
<a name="LC13Report"></a>Ludovic Courtès<br/>
<strong>C Language Extensions for Hybrid CPU/GPU Programming with StarPU</strong><br/>
Research Report RR-8278, INRIA, April 2013<br/>
826 827
[<a href="http://hal.inria.fr/hal-00807033">WWW</a>]
[<a href="http://hal.inria.fr/hal-00807033/PDF/RR-8278.pdf">PDF</a>]
828 829
</li>
</ol>
830
<h4>On OpenMP Support on top of StarPU</h4> 
831
<a name="PublicationsOnOpenMPSupportontopofStarPU"></a>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
832
<ol>
833
<li>
Nathalie Furmento's avatar
Nathalie Furmento committed
834 835 836 837 838 839 840 841
<a name="agullo:hal-01517153"></a>Emmanuel Agullo, Olivier Aumage, Berenger Bramas, Olivier Coulaud,  and Samuel Pitoiset<br/>
<strong>Bridging the gap between OpenMP and task-based runtime systems for the fast multipole method</strong><br/>
<em>IEEE Transactions on Parallel and Distributed Systems</em>, April 2017<br/>
[<a href="https://hal.inria.fr/hal-01517153">WWW</a>]
[<a href="https://hal.inria.fr/hal-01517153/file/tpds_kstar_scalfmm_print.pdf">PDF</a>]
[doi:<a href="http://dx.doi.org/10.1109/TPDS.2017.2697857">10.1109/TPDS.2017.2697857</a>]
</li>
<li>
842
<a name="agullo:hal-01372022"></a>Emmanuel Agullo, Olivier Aumage, Berenger Bramas, Olivier Coulaud,  and Samuel Pitoiset<br/>
843
<strong>Bridging the gap between OpenMP 4.0 and native runtime systems for the fast multipole method</strong><br/>
844 845 846 847 848
Research Report RR-8953, Inria, March 2016<br/>
[<a href="https://hal.inria.fr/hal-01372022">WWW</a>]
[<a href="https://hal.inria.fr/hal-01372022/file/RR-8953.pdf">PDF</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
849
<a name="virouleau:hal-01081974"></a>Philippe Virouleau, Pierrick Brunet, François Broquedis, Nathalie Furmento, Samuel Thibault, Olivier Aumage,  and Thierry Gautier<br/>
850
<strong>Evaluation of OpenMP Dependent Tasks with the KASTORS Benchmark Suite</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
851
In <em>IWOMP2014 - 10th International Workshop on OpenMP</em>, 10th International Workshop on OpenMP, IWOMP2014, Salvador, Brazil, France, pages 16 - 29, September 2014<br/>
852
Springer<br/>
853
[<a href="https://hal.inria.fr/hal-01081974">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
854
[<a href="https://hal.inria.fr/hal-01081974/document">PDF</a>]
855
[doi:<a href="http://dx.doi.org/10.1007/978-3-319-11454-5_2">10.1007/978-3-319-11454-5_2</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
856 857
</li>
</ol>
858
<h4>On MPI Support</h4> 
859
<a name="PublicationsOnMPISupport"></a>
860 861
<ol>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
862 863 864 865 866 867 868
<a name="lion:hal-02296118"></a>Romain Lion<br/>
<strong>Tolérance aux pannes dans l'exécution distribuée de graphes de tâches</strong><br/>
In <em>Conférence d'informatique en Parallélisme, Architecture et Système</em>, Anglet, France, June 2019<br/>
[<a href="https://hal.inria.fr/hal-02296118">WWW</a>]
[<a href="https://hal.inria.fr/hal-02296118/file/Compas_Romain_LION_submitted_final.pdf">PDF</a>]
</li>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
869
<a name="agullo:hal-01618526"></a>Emmanuel Agullo, Olivier Aumage, Mathieu Faverge, Nathalie Furmento, Florent Pruvost, Marc Sergent,  and Samuel Thibault<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
870
<strong>Achieving High Performance on Supercomputers with a Sequential Task-based Programming Model</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
871
<em>TPDS - IEEE Transactions on Parallel and Distributed Systems</em>, December 2017<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
872 873
[<a href="https://hal.inria.fr/hal-01618526">WWW</a>]
[<a href="https://hal.inria.fr/hal-01618526/file/tpds14.pdf">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
874
[doi:<a href="http://dx.doi.org/10.1109/TPDS.2017.2766064">10.1109/TPDS.2017.2766064</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
875 876 877 878
</li>
<li>
<a name="sergent:tel-01483666"></a>Marc Sergent<br/>
<strong>Scalability of a task-based runtime system for dense linear algebra applications</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
879
PhD thesis, Université de Bordeaux, December 2016<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
880 881
[<a href="https://tel.archives-ouvertes.fr/tel-01483666">WWW</a>]
[<a href="https://tel.archives-ouvertes.fr/tel-01483666/file/SERGENT_MARC_2016.pdf">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
882 883
</li>
<li>
884 885 886 887 888 889 890
<a name="agullo:hal-01283949"></a>Emmanuel Agullo, Olivier Aumage, Mathieu Faverge, Nathalie Furmento, Florent Pruvost, Marc Sergent,  and Samuel Thibault<br/>
<strong>Harnessing clusters of hybrid nodes with a sequential task-based programming model</strong><br/>
In <em>8th International Workshop on Parallel Matrix Algorithms and Applications</em>, July 2014<br/>
[<a href="https://hal.inria.fr/hal-01283949">WWW</a>]
[<a href="https://hal.inria.fr/hal-01283949/file/pmaa14.pdf">PDF</a>]
</li>
<li>
891 892
<a name="augonnet:hal-00992208"></a>Cédric Augonnet, Olivier Aumage, Nathalie Furmento, Samuel Thibault,  and Raymond Namyst<br/>
<strong>StarPU-MPI: Task Programming over Clusters of Machines Enhanced with Accelerators</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
893
Research Report RR-8538, INRIA, May 2014<br/>
894 895 896 897
[<a href="http://hal.inria.fr/hal-00992208">WWW</a>]
[<a href="http://hal.inria.fr/hal-00992208/PDF/RR-8538.pdf">PDF</a>]
</li>
<li>
898 899 900 901 902
<a name="AugAumFurNamThi2012EuroMPI"></a>Cédric Augonnet, Olivier Aumage, Nathalie Furmento, Raymond Namyst,  and Samuel Thibault<br/>
<strong>StarPU-MPI: Task Programming over Clusters of Machines Enhanced with Accelerators</strong><br/>
In Siegfried Benkner Jesper Larsson Träff and Jack Dongarra, editors, <em>EuroMPI 2012</em>, volume 7490 of <em>LNCS</em>, September 2012<br/>
Springer<br/>
Note: Poster Session<br/>
903
[<a href="http://hal.inria.fr/hal-00725477">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
904
[<a href="http://hal.inria.fr/hal-00725477/document">PDF</a>]
905
</li>
906
</ol>
Nathalie Furmento's avatar
Nathalie Furmento committed
907
<h4>On Memory Control</h4> 
908
<a name="PublicationsOnMemoryControl"></a>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
909 910
<ol>
<li>
911 912 913 914 915 916 917
<a name="chevalier:hal-01718280"></a>Arthur Chevalier<br/>
<strong>Critical resources management and scheduling under StarPU</strong><br/>
Master's thesis, Université de Bordeaux, September 2017<br/>
[<a href="https://hal.inria.fr/hal-01718280">WWW</a>]
[<a href="https://hal.inria.fr/hal-01718280/file/Memoire.pdf">PDF</a>]
</li>
<li>
Nathalie Furmento's avatar
Nathalie Furmento committed
918
<a name="sergent:hal-01284004"></a>Marc Sergent, David Goudin, Samuel Thibault,  and Olivier Aumage<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
919
<strong>Controlling the Memory Subscription of Distributed Applications with a Task-Based Runtime System</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
920
In <em>HIPS - 21st International Workshop on High-Level Parallel Programming Models and Supportive Environments</em>, Chicago, United States, May 2016<br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
921 922
[<a href="https://hal.inria.fr/hal-01284004">WWW</a>]
[<a href="https://hal.inria.fr/hal-01284004/file/PID4127657.pdf">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
923
[doi:<a href="http://dx.doi.org/10.1109/IPDPSW.2016.105">10.1109/IPDPSW.2016.105</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
924 925
</li>
</ol>
926
<h4>On Performance Model Tuning</h4> 
927
<a name="PublicationsOnPerformanceModelTuning"></a>
928 929
<ol>
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
930 931 932 933 934 935 936
<a name="agullo:hal-01474556"></a>Emmanuel Agullo, Bérenger Bramas, Olivier Coulaud, Luka Stanisic,  and Samuel Thibault<br/>
<strong>Modeling Irregular Kernels of Task-based codes: Illustration with the Fast Multipole Method</strong><br/>
Research Report RR-9036, INRIA Bordeaux, February 2017<br/>
[<a href="https://hal.inria.fr/hal-01474556">WWW</a>]
[<a href="https://hal.inria.fr/hal-01474556/file/rapport.pdf">PDF</a>]
</li>
<li>
937 938
<a name="AugThiNam09HPPC"></a>Cédric Augonnet, Samuel Thibault,  and Raymond Namyst<br/>
<strong>Automatic Calibration of Performance Models on Heterogeneous Multicore Architectures</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
939
In <em>HPPC - Proceedings of the International Euro-Par Workshops, Highly Parallel Processing on a Chip</em>, volume 6043 of <em>Lecture Notes in Computer Science</em>, Delft, The Netherlands, pages 56-65, August 2009<br/>
940
Springer<br/>
941
[<a href="http://hal.inria.fr/inria-00421333">WWW</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
942
[<a href="http://hal.inria.fr/inria-00421333/document">PDF</a>]
943
[doi:<a href="http://dx.doi.org/10.1007/978-3-642-14122-5_9">10.1007/978-3-642-14122-5_9</a>]
944 945
</li>
</ol>
946
<h4>On The Simulation Support through SimGrid</h4> 
947
<a name="PublicationsOnTheSimulationSupportthroughSimGrid"></a>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
948
<ol>
THIBAULT Samuel's avatar
update  
THIBAULT Samuel committed
949
<li>
950 951
<a name="stanisic:hal-01147997"></a>Luka Stanisic, Samuel Thibault, Arnaud Legrand, Brice Videau,  and Jean-François Méhaut<br/>
<strong>Faithful Performance Prediction of a Dynamic Task-Based Runtime System for Heterogeneous Multi-Core Architectures</strong><br/>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
952
<em>CCPE - Concurrency and Computation: Practice and Experience</em>, pp 16, May 2015<br/>
953 954
[<a href="https://hal.inria.fr/hal-01147997">WWW</a>]
[<a href="https://hal.inria.fr/hal-01147997/file/CCPE14_article.pdf">PDF</a>]
THIBAULT Samuel's avatar
THIBAULT Samuel committed
955
[doi:<a href="http://dx.doi.org/10.1002/cpe.3555">10.1002/cpe.3555</a>]
THIBAULT Samuel's avatar
update  
THIBAULT Samuel committed
956
</li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
957
<li>
THIBAULT Samuel's avatar
THIBAULT Samuel committed
958 959 960 961 962 963 964
<a name="stanisic:hal-01180272"></a>Luka Stanisic, Emmanuel Agullo, Alfredo Buttari, Abdou Guermouche, Arnaud Legrand, Florent Lopez,  and Brice Videau<br/>
<strong>Fast and Accurate Simulation of Multithreaded Sparse Linear Algebra Solvers</strong><br/>
In <em>The 21st IEEE International Conference on Parallel and Distributed Systems</em>, Melbourne, Australia, December 2015<br/>
[<a href="https://hal.inria.fr/hal-01180272">WWW</a>]
[<a href="https://hal.inria.fr/hal-01180272/file/QRMSTARSG_article.pdf">PDF</a>]
</li>
<li>