index.html 24.9 KB
Newer Older
AUMAGE Olivier's avatar
AUMAGE Olivier committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN"
            "http://www.w3.org/TR/REC-html40/loose.dtd">
<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 hands-on session</TITLE>
<link rel="stylesheet" type="text/css" href="../../style.css" />
<link rel="Shortcut icon" href="http://www.inria.fr/extension/site_inria/design/site_inria/images/favicon.ico" type="image/x-icon" />
</HEAD>

<body>

<div class="title">
<h1><a href="../../">StarPU</a></h1>
<h2>ComPAS 2018</h2>
<h3>StarPU Tutorial - Toulouse, July 2018</h3>
</div>

<div class="menu">
      <a href="../">Back to the main page</a>
</div>

<div class="section">

<p>
Other materials (talk slides, links) are available at the
<a href="index.html#other">bottom</a> of this page.
</p>
</div>

<div class="section">
<h2>Setup</h2>

<div class="section">
<h3>Connection to the Platform</h3>
<p>
<!--
The lab works are going to be done on
the <a href="https://groupes.renater.fr/wiki/poincare/public/description_de_poincare">MDS</a> platform.
-->
<!--
A subset of machines has been specifically booked for our own usage.
-->
You should have received information on how to connect to the
platform.
</p>

<!--
<P>
To use StarPU on the machines, you need to load the following modules
</p>

<tt>
<pre>
module load cuda
module load openmpi
module load hwloc/1.6.2_gnu47
module load openblas/v0.2.8_gnu48
</pre>
</tt>
-->

<p>
  The following variables need to be set to use StarPU.
</p>

<tt>
<pre>
AUMAGE Olivier's avatar
AUMAGE Olivier committed
69
70
71
72
73
74
75
76
77
78
79
80
81
export TP_DIR=/mnt/n7fs/ens/tp_abuttari/TP_StarPU/

export HWLOC_PATH=$TP_DIR/hwloc-1.11.10
export PATH=$HWLOC_PATH/bin:$PATH
export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:$HWLOC_PATH/lib/pkgconfig
export LD_LIBRARY_PATH=$HWLOC_PATH/lib:$LD_LIBRARY_PATH

export FXT_PATH=$TP_DIR/fxt-0.3.8
export PATH=$FXT_PATH/bin:$PATH
export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:$FXT_PATH/lib/pkgconfig
export LD_LIBRARY_PATH=$FXT_PATH/lib:$LD_LIBRARY_PATH

export STARPU_PATH=$TP_DIR/starpu
AUMAGE Olivier's avatar
AUMAGE Olivier committed
82
83
84
export PATH=$STARPU_PATH/bin:$PATH
export PKG_CONFIG_PATH=$PKG_CONFIG_PATH:$STARPU_PATH/lib/pkgconfig
export LD_LIBRARY_PATH=$STARPU_PATH/lib:$LD_LIBRARY_PATH
AUMAGE Olivier's avatar
AUMAGE Olivier committed
85
export STARPU_IDLE_FILE=$HOME/starpu_idle_microsec.log
AUMAGE Olivier's avatar
AUMAGE Olivier committed
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896

export LIBRARY_PATH=$LD_LIBRARY_PATH

</pre>
</tt>

<p>
  You can either add the previous lines to your
file <tt>$HOME/.bash_profile</tt>, or use the script
file <tt>/gpfslocal/pub/training/runtime_june2016/starpu_env.sh</tt>
</p>

</div>

<!--
<div class="section">
<h3>Job Submission</h3>
<p>
Jobs can be submitted to the platform to reserve a set of nodes and to
execute a application on those nodes.
Here is a script to submit your
first StarPU application. It calls the
tool <tt>starpu_machine_display</tt> which shows the processing units
that StarPU can use, and the bandwidth and affinity measured between
the memory nodes.
</p>

<p>
MdS nodes are accessed through queues which represent
machines with similar characteristics. For our lab works, we have 2
sets of machines:
<ul>
  <li> GPU nodes accessed with the queue <tt>clgpu</tt>. </li>
  <li> Non-GPU nodes accessed with the queue <tt>clallmds</tt>. </li>
</ul>
</p>

<tt>
<pre>
#!/bin/bash
# @ class            = clgpu
# @ job_name = job_starpu_machine_display
# @ total_tasks = 10
# @ node = 1
# @ wall_clock_limit = 00:10:00
# @ output = $(HOME)/starpu/$(job_name).$(jobid).out
# @ error = $(HOME)/starpu/$(job_name).$(jobid).err
# @ job_type = mpich
# @ queue

source /gpfslocal/pub/training/runtime_june2016/starpu_env.sh
starpu_machine_display
</pre>
</tt>

<P>
  You will find a copy of the script in <tt>/gpfslocal/pub/training/runtime_june2016/starpu_machine_display.sh</tt>.
  To submit the script, simply call:
</p>

<tt>
<pre>
llsubmit starpu_machine_display.sh
</pre>
</tt>

<p>
The state of the job can be queried by calling the command <tt>llq | grep $USER</tt>.
Once finished, the standard output and the standard error generated by
the script execution are available in the files:
<ul>
<li>${HOME}/starpu/<b>jobname</b>.<b>sequence_number</b>.out</li>
<li>${HOME}/starpu/<b>jobname</b>.<b>sequence_number</b>.err</li>
</ul>
</p>

<p>
Note that the first time <tt>starpu_machine_display</tt> is executed,
it calibrates the performance model of the bus, the results are then
stored in different files in the
directory <tt>$HOME/.starpu/sampling/bus</tt>. If you run the command
several times, you will notice that StarPU may calibrate the bus speed
several times. This is because the cluster's batch scheduler may assign a
different node each time, and StarPU does not know that the local
cluster we use is homogeneous, and thus assumes that all nodes of the
cluster may be different. The following line could be added to the
script file to force StarPU to use the same machine ID
for the whole cluster:
</p>

<tt>
<pre>
$ export STARPU_HOSTNAME=poincaregpu
</pre>
</tt>

<p>
Of course, on a heterogeneous cluster, the cluster launcher script
should set various hostnames for the different node classes, as
appropriate.
</p>
</div>
-->

<!--
<div class="section">
<h3>Tutorial Material</h3>

<p>
All files needed for the lab works are available on the machine in the
directory <tt>/gpfslocal/pub/training/runtime_june2016</tt>.
</p>

</div>

</div>
-->

<div class="section">
<h2>Session Part 1: Task-based Programming Model</h2>

<div class="section">
<h3>Application Example: Vector Scaling</h3>

<h4>Making it and Running it</h4>

<p>
A typical <a href="files/Makefile"><tt>Makefile</tt></a> for
applications using StarPU is the following:
</p>

<tt>
<pre>
CFLAGS += $(shell pkg-config --cflags starpu-1.2)
LDFLAGS += $(shell pkg-config --libs starpu-1.2)
%.o: %.cu
	nvcc $(CFLAGS) $< -c $

vector_scal_task_insert: vector_scal_task_insert.o vector_scal_cpu.o vector_scal_cuda.o vector_scal_opencl.o
</pre>
</tt>

<p>
Here are the source files for the application:
<ul>
<li><a href="files/vector_scal_task_insert.c">The main application</a></li>
<li><a href="files/vector_scal_cpu.c">The CPU implementation of the codelet</a></li>
<li><a href="files/vector_scal_cuda.cu">The CUDA implementation of the codelet</a></li>
<li><a href="files/vector_scal_opencl.c">The OpenCL host implementation of the codelet</a></li>
<li><a href="files/vector_scal_opencl_kernel.cl">The OpenCL device implementation of the codelet</a></li>
</ul>

Run <tt>make</tt>, and run the
resulting <tt>vector_scal_task_insert</tt> executable using the batch
scheduler using the <a href="files/vector_scal.sh">given script
  vector_scal.sh</a>. It should be working: it simply scales a given
vector by a given factor.
</p>

<tt>
<pre>
#!/bin/bash
# @ class            = clgpu
# @ job_name = job_vector_scal
# @ total_tasks = 10
# @ node = 1
# @ wall_clock_limit = 00:10:00
# @ output = $(HOME)/starpu/$(job_name).$(jobid).out
# @ error = $(HOME)/starpu/$(job_name).$(jobid).err
# @ job_type = mpich
# @ queue

source /gpfslocal/pub/training/runtime_june2016/starpu_env.sh

make vector_scal_task_insert

./vector_scal_task_insert
</pre>
</tt>

<h4>Computation Kernels</h4>
<p>
Examine the source code, starting from <tt>vector_scal_cpu.c</tt> : this is
the actual computation code, which is wrapped into a <tt>vector_scal_cpu</tt>
function which takes a series of DSM interfaces and a non-DSM parameter. The
code simply gets the factor value from the non-DSM parameter,
an actual pointer from the first DSM interface,
and performs the vector scaling.
</p>

<p>
The GPU implementation, in <tt>vector_scal_cuda.cu</tt>, is basically
the same, with the host part (<tt>vector_scal_cuda</tt>) which extracts the
actual CUDA pointer from the DSM interface, and passes it to the device part
(<tt>vector_mult_cuda</tt>) which performs the actual computation.
</p>

<p>
The OpenCL implementation in <tt>vector_scal_opencl.c</tt> and
<tt>vector_scal_opencl_kernel.cl</tt>is more hairy due to the low-level aspect
of the OpenCL standard, but the principle remains the same.
</p>

<p>
Modify the source code of the different implementations (CPU, CUDA and
OpenCL) and see which ones gets executed. You can force the execution
of one the implementations simply by disabling a type of device when
running your application, e.g.:
</p>

<tt>
<pre>
# to force the implementation on a GPU device, by default, it will enable CUDA
STARPU_NCPUS=0 vector_scal_task_insert

# to force the implementation on a OpenCL device
STARPU_NCPUS=0 STARPU_NCUDA=0 vector_scal_task_insert
</pre>
</tt>

<p>
You can set the environment variable STARPU_WORKER_STATS to 1 when
running your application to see the number of tasks executed by each
device. You can see the whole list of environment
variables <a href="http://starpu.gforge.inria.fr/doc/html/ExecutionConfigurationThroughEnvironmentVariables.html">here</a>.
</p>

<tt>
<pre>
STARPU_WORKER_STATS=1 vector_scal_task_insert
</pre>
</tt>

<h4>Main Code</h4>
<p>
Now examine <tt>vector_scal_task_insert.c</tt>: the <tt>cl</tt>
(codelet) structure simply gathers pointers on the functions
mentioned above.
</p>

<p>
The <tt>main</tt> function
<ul>
<li>Allocates an <tt>vector</tt> application buffer and fills it.</li>
<li>Registers it to StarPU, and gets back a DSM handle. From now on, the
application is not supposed to access <tt>vector</tt> directly, since its
content may be copied and modified by a task on a GPU, the main-memory copy then
being outdated.</li>
<li>Submits a (asynchronous) task to StarPU.</li>
<li>Waits for task completion.</li>
<li>Unregisters the vector from StarPU, which brings back the modified version
to main memory.</li>
</ul>
</p>

</div>

<div class="section">
<h3>Data Partitioning</h3>

<p>
In the previous section, we submitted only one task. We here discuss how to
<i>partition</i> data so as to submit multiple tasks which can be executed in
parallel by the various CPUs and GPUs.
</p>

<p>
Let's examine <a href="files/mult.c">mult.c</a>.

<ul>
<li>
The computation kernel, <tt>cpu_mult</tt> is a trivial matrix multiplication
kernel, which operates on 3 given DSM interfaces. These will actually not be
whole matrices, but only small parts of matrices.
</li>
<li>
<tt>init_problem_data</tt> initializes the whole A, B and C matrices.
</li>
<li>
<tt>partition_mult_data</tt> does the actual registration and partitioning.
Matrices are first registered completely, then two partitioning filters are
declared. The first one, <tt>vert</tt>, is used to split B and C vertically. The
second one, <tt>horiz</tt>, is used to split A and C horizontally. We thus end
up with a grid of pieces of C to be computed from stripes of A and B.
</li>
<li>
<tt>launch_tasks</tt> submits the actual tasks: for each piece of C, take
the appropriate piece of A and B to produce the piece of C.
</li>
<li>
The access mode is interesting: A and B just need to be read from, and C
will only be written to. This means that StarPU will make copies of the pieces
of A and B along the machines, where they are needed for tasks, and will give to
the tasks some
uninitialized buffers for the pieces of C, since they will not be read
from.
</li>
<li>
The <tt>main</tt> code initializes StarPU and data, launches tasks, unpartitions data,
and unregisters it. Unpartitioning is an interesting step: until then the pieces
of C are residing on the various GPUs where they have been computed.
Unpartitioning will collect all the pieces of C into the main memory to form the
whole C result matrix.
</li>
</ul>
</p>

<p>
Run the application with the <a href="files/mult.sh">batch scheduler</a>, enabling some statistics:
</p>

<tt>
<pre>
</pre>
</tt>

<p>
Figures show how the computation were distributed on the various processing
units.
</p>
</div>

<div class="section">
<h3>Other example</h3>

<p>
<a href="files/gemm/xgemm.c"><tt>gemm/xgemm.c</tt></a> is a very similar
matrix-matrix product example, but which makes use of BLAS kernels for
much better performance. The <tt>mult_kernel_common</tt> functions
shows how we call <tt>DGEMM</tt> (CPUs) or <tt>cublasDgemm</tt> (GPUs)
on the DSM interface.
</p>

<p>
Let's execute it.
</p>

<tt>
<pre>
#!/bin/bash
# @ class            = clgpu
# @ job_name = job_gemm
# @ total_tasks = 10
# @ node = 1
# @ wall_clock_limit = 00:10:00
# @ output = $(HOME)/starpu/$(job_name).$(jobid).out
# @ error = $(HOME)/starpu/$(job_name).$(jobid).err
# @ job_type = mpich
# @ queue

source /gpfslocal/pub/training/runtime_june2016/starpu_env.sh

make gemm/sgemm
STARPU_WORKER_STATS=1 ./gemm/sgemm
</pre>
</tt>

<!--
<p>
We can notice that StarPU gave much more tasks to the GPU. You can also try
to set <tt>num_gpu=2</tt> to run on the machine which has two GPUs (there is
only one of them, so you may have to wait a long time, so submit this in
background in a separate terminal), the interesting thing here is that
with <b>no</b> application modification beyond making it use a task-based
programming model, we get multi-GPU support for free!
</p>
-->

</div>

<!--
<div class="section">
<h3>More Advanced Examples</h3>
<p>
<tt>examples/lu/xlu_implicit.c</tt> is a more involved example: this is a simple
LU decomposition algorithm. The <tt>dw_codelet_facto_v3</tt> is actually the
main algorithm loop, in a very readable, sequential-looking way. It simply
submits all the tasks asynchronously, and waits for them all.
</p>

<p>
<tt>examples/cholesky/cholesky_implicit.c</tt> is a similar example, but which makes use
of the <tt>starpu_insert_task</tt> helper. The <tt>_cholesky</tt> function looks
very much like <tt>dw_codelet_facto_v3</tt> of the previous paragraph, and all
task submission details are handled by <tt>starpu_insert_task</tt>.
</p>

<p>
Thanks to being already using a task-based programming model, MAGMA and PLASMA
have been easily ported to StarPU by simply using <tt>starpu_insert_task</tt>.
</p>
</div>
-->

<div class="section">
<h3>Exercise</h3>
<p>
Take the vector example again, and add partitioning support to it, using the
matrix-matrix multiplication as an example. Here we will use the
<a href="http://starpu.gforge.inria.fr/doc/html/group__API__Data__Partition.html#ga212189d3b83dfa4e225609b5f2bf8461"><tt>starpu_vector_filter_block()</tt></a> filter function. You can see the list of
predefined filters provided by
StarPU <a href="http://starpu.gforge.inria.fr/doc/html/starpu__data__filters_8h.html">here</a>.
Try to run it with various numbers of tasks.
</p>
</div>
</div>

<div class="section">
<h2>Session Part 2: Optimizations</h2>


<p>
This is based on StarPU's documentation
<a href="http://starpu.gforge.inria.fr/doc/html/HowToOptimizePerformanceWithStarPU.html">optimization
  chapter</a>
</p>

<div class="section">
<h3>Data Management</h3>

<p>
We have explained how StarPU can overlap computation and data transfers
thanks to DMAs. This is however only possible when CUDA has control over the
application buffers. The application should thus use <a href="http://starpu.gforge.inria.fr/doc/html/group__API__Standard__Memory__Library.html#ga49603eaea3b05e8ced9ba1bd873070c3"><tt>starpu_malloc()</tt></a>
when allocating its buffer, to permit asynchronous DMAs from and to
it.
</p>

<p>
Take the vector example again, and fix the allocation, to make it use
<a href="http://starpu.gforge.inria.fr/doc/html/group__API__Standard__Memory__Library.html#ga49603eaea3b05e8ced9ba1bd873070c3"><tt>starpu_malloc()</tt></a>.
</p>

</div>

<div class="section">
<h3>Task Submission</h3>

<p>
To let StarPU reorder tasks, submit data transfers in advance, etc., task
submission should be asynchronous whenever possible. Ideally, the application
should behave like that: submit the
whole graph of tasks, and wait for termination.
</p>

</div>

<div class="section">
<h3>Task Scheduling Policy</h3>
<p>
By default, StarPU uses the <tt>eager</tt> simple greedy scheduler. This is
because it provides correct load balance even if the application codelets do not
have performance models: it uses a single central queue, from which workers draw
tasks to work on. This however does not permit to prefetch data, since the
scheduling decision is taken late.
</p>

<p>
If the application codelets have performance models, the scheduler should be
changed to take benefit from that. StarPU will then really take scheduling
decision in advance according to performance models, and issue data prefetch
requests, to overlap data transfers and computations.
</p>

<p>
For instance, compare the <tt>eager</tt> (default) and <tt>dmda</tt> scheduling
policies:
</p>

<tt>
<pre>
STARPU_BUS_STATS=1 STARPU_WORKER_STATS=1 gemm/sgemm -x 1024 -y 1024 -z 1024
</pre>
</tt>

<p>
with:
</p>

<tt>
<pre>
STARPU_BUS_STATS=1 STARPU_WORKER_STATS=1 STARPU_SCHED=dmda gemm/sgemm -x 1024 -y 1024 -z 1024
</pre>
</tt>

<p>
You can see most (all?) the computation have been done on GPUs,
leading to better performances.
</p>

<p>
Try other schedulers, use <tt>STARPU_SCHED=help</tt> to get the
list.
</p>

<p>
Also try with various sizes and draw curves.
</p>

<p>
You can also try the double version, <tt>dgemm</tt>, and notice that GPUs get
less great performance.
</p>

</div>


<div class="section">
<h3>Performance Model Calibration</h3>

<p>
Performance prediction is essential for proper scheduling decisions, the
performance models thus have to be calibrated.  This is done automatically by
StarPU when a codelet is executed for the first time.  Once this is done, the
result is saved to a file in <tt>$STARPU_HOME</tt> for later re-use.  The
<tt>starpu_perfmodel_display</tt> tool can be used to check the resulting
performance model.
</p>

<tt>
<pre>
$ starpu_perfmodel_display -l
file: &lt;starpu_sgemm_gemm.mirage&gt;
$ starpu_perfmodel_display -s starpu_sgemm_gemm
performance model for cpu_impl_0
# hash		size		flops		mean (us)	stddev (us)		n
8bd4e11d	2359296        	0.000000e+00   	1.848856e+04   	4.026761e+03   	12
performance model for cuda_0_impl_0
# hash		size		flops		mean (us)	stddev (us)		n
8bd4e11d	2359296        	0.000000e+00   	4.918095e+02   	9.404866e+00   	66
...
</pre>
</tt>

<p>
This shows that for the sgemm kernel with a 2.5M matrix slice, the average
execution time on CPUs was about 18ms, with a 4ms standard deviation, over
12 samples, while it took about 0.049ms on GPUs, with a 0.009ms standard
deviation. It is a good idea to check this before doing actual performance
measurements. If the kernel has varying performance, it may be a good idea to
force StarPU to continue calibrating the performance model, by using <tt>export
STARPU_CALIBRATE=1</tt>
</p>

<p>
If the code of a computation kernel is modified, the performance changes, the
performance model thus has to be recalibrated from start. To do so, use
<tt>export STARPU_CALIBRATE=2</tt>
</p>

<p>
The performance model can also be drawn by using <tt>starpu_perfmodel_plot</tt>,
which will emit a gnuplot file in the current directory.
</p>

</div>
</div>

<div class="section">
<h2>Sessions Part 3: MPI Support</h2>

<p>
StarPU provides support for MPI communications. It does so in two ways. Either the
application specifies MPI transfers by hand, or it lets StarPU infer them from
data dependencies.
</p>

<div class="section">
<h3>Manual MPI transfers</h3>

<p>Basically, StarPU provides
equivalents of <tt>MPI_*</tt> functions, but which operate on DSM handles
instead of <tt>void*</tt> buffers. The difference is that the source data may be
residing on a GPU where it just got computed. StarPU will automatically handle
copying it back to main memory before submitting it to MPI.
</p>

<p>
<a href="files/mpi/ring_async_implicit.c"><tt>ring_async_implicit.c</tt></a>
shows an example of mixing MPI communications and task submission. It
is a classical ring MPI ping-pong, but the token which is being passed
on from neighbour to neighbour is incremented by a starpu task at each
step.
</p>

<p>
This is written very naturally by simply submitting all MPI
communication requests and task submission asynchronously in a
sequential-looking loop, and eventually waiting for all the tasks to
complete.
</p>

<tt>
<pre>
#!/bin/bash
# @ class            = clgpu
# @ job_name = job_ring
# @ total_tasks = 10
# @ node = 1
# @ wall_clock_limit = 00:10:00
# @ output = $(HOME)/starpu/$(job_name).$(jobid).out
# @ error = $(HOME)/starpu/$(job_name).$(jobid).err
# @ job_type = mpich
# @ queue

source /gpfslocal/pub/training/runtime_june2016/starpu_env.sh

make ring_async_implicit
mpirun -np 2 $PWD/ring_async_implicit
</pre>
</tt>
</div>

<div class="section">
<h3>starpu_mpi_insert_task</h3>

<p>
<a href="files/mpi/stencil5.c">A stencil application</a> shows a basic MPI
task model application. The data distribution over MPI
nodes is decided by the <tt>my_distrib</tt> function, and can thus be changed
trivially.
It also shows how data can be migrated to a
new distribution.
</p>

<tt>
<pre>
#!/bin/bash
# @ class            = clgpu
# @ job_name = job_stencil
# @ total_tasks = 10
# @ node = 1
# @ wall_clock_limit = 00:10:00
# @ output = $(HOME)/starpu/$(job_name).$(jobid).out
# @ error = $(HOME)/starpu/$(job_name).$(jobid).err
# @ job_type = mpich
# @ queue

source /gpfslocal/pub/training/runtime_june2016/starpu_env.sh

make stencil5
mpirun -np 2 $PWD/stencil5 -display
</pre>
</tt>

</div>

<div class="section">
<h2>Session Part 4: OpenMP Support</h2>

<div class="section">
<h3>The Klang-Omp OpenMP Compiler</h3>

<p>
The <b>Klang-Omp</b> OpenMP compiler converts C/C++ source codes annotated with OpenMP 4 directives into StarPU enabled codes. Klang-Omp is source-to-source compiler based on the LLVM/CLang compiler framework.
</p>

<p>
The following shell sequence shows an example of an OpenMP version of the Cholesky decomposition compiled into StarPU code.
</p>

<tt>
<pre>
cd
source /gpfslocal/pub/training/runtime_june2016/openmp/environment
cp -r /gpfslocal/pub/training/runtime_june2016/openmp/Cholesky .
cd Cholesky
make
./cholesky_omp4.starpu
</pre>
</tt>

<p>
Homepage of the Klang-Omp OpenMP compiler: <a href="http://kstar.gforge.inria.fr/">Klang-Omp</a>
</p>

</div>
</div>


<div class="section" id="contact">
<h2>Contact</h2>
<p>
For any questions regarding StarPU, please contact the StarPU developers mailing list.
<a href="mailto:starpu-devel@lists.gforge.inria.fr?subject=StarPU">starpu-devel@lists.gforge.inria.fr</a>
</p>
</div>

<div class="section">
<h2>More Performance Optimizations</h2>
<p>
The StarPU
documentation <a href="http://starpu.gforge.inria.fr/doc/html/PerformanceFeedback.html">performance
    feedback chapter</a> provides more optimization tips for further
reading after this tutorial.
</p>

<!--
<div class="section">
<h3>FxT Tracing Support</h3>

<p>
In addition to online profiling, StarPU provides offline profiling tools,
based on recording a trace of events during execution, and analyzing it
afterwards.
</p>

<p>
To use the version of StarPU compiled with FxT support, you need to reload the
module StarPU after loading the module FxT.
</p>

<tt>
<pre>
module unload runtime/starpu/1.1.4
module load trace/fxt/0.2.13
module load runtime/starpu/1.1.4
</pre>
</tt>

<p>
The trace file is stored in <tt>/tmp</tt> by default. Since execution will
happen on a cluster node, the file will not be reachable after execution,
we need to tell StarPU to store output traces in the home directory, by
setting:
</p>

<tt>
<pre>
$ export STARPU_FXT_PREFIX=$HOME/
</pre>
</tt>

<p>
do not forget the add the line in your file <tt>.bash_profile</tt>.
</p>

<p>
The application should be run again, and this time a <tt>prof_file_XX_YY</tt>
trace file will be generated in your home directory. This can be converted to
several formats by using:
</p>

<tt>
<pre>
$ starpu_fxt_tool -i ~/prof_file_*
</pre>
</tt>

<p>
This will create
<ul>
<li>
a <tt>paje.trace</tt> file, which can be opened by using the <a
href="http://vite.gforge.inria.fr/">ViTE</a> tool. This shows a Gant diagram of
the tasks which executed, and thus the activity and idleness of tasks, as well
as dependencies, data transfers, etc. You may have to zoom in to actually focus
on the computation part, and not the lengthy CUDA initialization.
</li>
<li>
a <tt>dag.dot</tt> file, which contains the graph of all the tasks
submitted by the application. It can be opened by using Graphviz.
</li>
<li>
an <tt>activity.data</tt> file, which records the activity of all processing
units over time.
</li>
</ul>
</p>
</div>
</div>
-->

<div class="section" id="other">
<h2>Other Materials: Talk Slides and Website Links</h2>
<p>
<h3>General Session Introduction</h3>
<ul>
<li> <a href="slides/00_intro_runtimes.pdf">Slides: Introduction to Runtime Systems</a>
</li>
</ul>
<h3>The Hardware Locality Library (hwloc)</h3>
<ul>
<li> <a href="http://www.open-mpi.org/projects/hwloc/tutorials/">Tutorial:
    hwloc</a>. For questions regarding hwloc, please
  contact <a href="mailto:brice.goglin@inria.fr">brice.goglin@inria.fr</a>.
</li>
</ul>
<h3>The StarPU Runtime System</h3>
<ul>
<Li> <a href="slides/01_introducing_starpu.pdf">Slides: StarPU - Part. 1 – Introducing StarPU</a></li>
<Li> <a href="slides/02_mastering_starpu.pdf">Slides: StarPU - Part. 2 – Mastering StarPU</a></li>
</ul>
<h3>The EZtrace Performance Debugging Framework</h3>
<ul>
<li> <a href="http://eztrace.gforge.inria.fr/tutorials/index.html">Tutorial:
  EzTrace</a>. For questions regarding EzTrace, please
  contact <a href="mailto:eztrace-devel@lists.gforge.inria.fr">eztrace-devel@lists.gforge.inria.fr</a>.
</li>
</ul>


</p>
</div>

<div class="section bot">
<p class="updated">
  Last updated on 2016/06/17.
</p>
</body>
</html>