diff --git a/doc/user/CMakeLists.txt b/doc/user/CMakeLists.txt
index 2e60e1c2be040139a0b22e6d00343290cec38707..b0ced19e0c612ad32adc89a07b8780c23bdd9632 100644
--- a/doc/user/CMakeLists.txt
+++ b/doc/user/CMakeLists.txt
@@ -58,6 +58,7 @@ set(FIGURES_USERGUIDE
trace_qr.jpg
potri_async.png
chameleon_header.png
+ lstopo-sirocco24.png
)
foreach(_fig ${FIGURES_USERGUIDE})
configure_file(${CMAKE_CURRENT_SOURCE_DIR}/chapters/${_fig}
diff --git a/doc/user/chapters/parallel_worker.org b/doc/user/chapters/parallel_worker.org
index 6023210712983792ffa1cdd53c2a9f7b2396504e..b18937ae9b1f01008caf0586f3dc3488cafe5563 100644
--- a/doc/user/chapters/parallel_worker.org
+++ b/doc/user/chapters/parallel_worker.org
@@ -12,7 +12,7 @@ workers which execute a single parallel task (see [[https://files.inria.fr/starp
To use this functionnality:
- StarPU must be compiled with the configure option =--enable-parallel-worker=
-- Chameleon automatically detects if the StarPU parallel workers are available or not and does not need any specific cmake options to use it.
+- Chameleon automatically detects if the StarPU parallel workers are available or not and can exploit them, but you need to force Chameleon to be linked with multi-threaded BLAS library if you want the parallel workers to be able to do parallel BLAS calls. To do that, you must add =-DCHAMELEON_KERNELS_MT=ON= to your cmake line.
Below are given some examples to use the couple Chameleon/StarPU to enable parallel tasks to be run concurrently. For now, this is only available for a few subset of tasks that are used in the Cholesky decomposition (=POTRF=, =TRSM=, =SYRK=, =HERK=, and =GEMM=) but all other algorithms using these kernels benefit from it.
@@ -20,13 +20,10 @@ Below are given some examples to use the couple Chameleon/StarPU to enable paral
:PROPERTIES:
:CUSTOM_ID: environment-variables
:END:
-- =CHAMELEON_PARALLEL_WORKER_LEVEL=hardware-level[:number-of-parallel-workers]=
- : Specify the number of parallel workers per hardware-level - the
- default value is 1. Note that hardware-level must correspond to an
- hwloc machine level type (hwloc_obj_type_t) e.g. =L2=, =L3=, =SOCKET=,
- =MACHINE=.
-- =CHAMELEON_PARALLEL_WORKER_SHOW= : When defined, the parallel workers
- contents is displayed.
+
+ - =CHAMELEON_PARALLEL_WORKER_LEVEL=hardware-level[:number-of-parallel-workers]=
+Specify the number of parallel workers per hardware-level. The default value is 1. Note that hardware-level must correspond to an hwloc machine level type (hwloc_obj_type_t) e.g.: =L2=, =L3=, =SOCKET=, =MACHINE=.
+ - =CHAMELEON_PARALLEL_WORKER_SHOW= : When defined, the parallel workers contents is displayed.
**** Limitations
:PROPERTIES:
@@ -35,11 +32,10 @@ Below are given some examples to use the couple Chameleon/StarPU to enable paral
For now, there is still an issue of bad performances with the usage of the =lws= scheduler with the parallel workers.
-
**** Examples
In the following examples, =STARPU_MAIN_THREAD_BIND= is set to 1 to bind
-the main thread of StarPU to a dedicated CPU to a reserved CPU,
+the main thread of StarPU to a dedicated CPU,
subtracted from the CPU workers. This avoids using a whole parallel
worker to make the submission.
@@ -57,8 +53,18 @@ CPUs are dedicated to run the GPUs.
parallel worker does not have all the CPUs of the last L3 cache, as
there are 3 dedicated CPUs.
+#+begin_src sh
+CHAMELEON_PARALLEL_WORKER_LEVEL=L3 \
+CHAMELEON_PARALLEL_WORKER_SHOW=1 \
+STARPU_MAIN_THREAD_BIND=1 \
+STARPU_CALIBRATE=1 \
+STARPU_SCHED=dmdar \
+STARPU_NWORKER_PER_CUDA=2 \
+STARPU_SILENT=1 \
+$PTCHAMELEON/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 1440:3000:480 -g 2
+#+end_src
+
#+begin_example
-$ CHAMELEON_PARALLEL_WORKER_LEVEL=L3 CHAMELEON_PARALLEL_WORKER_SHOW=1 STARPU_MAIN_THREAD_BIND=1 STARPU_CALIBRATE=1 STARPU_SCHED=dmdar STARPU_NWORKER_PER_CUDA=2 STARPU_SILENT=1 ~/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 1440:3000:480 -g 2
Number of parallel workers created: 8
Parallel worker 0 contains the following logical indexes:
0 1 2 3 4 5 6 7
@@ -91,8 +97,18 @@ Id;Function;threads;gpus;P;Q;mtxfmt;nb;uplo;n;lda;seedA;tsub;time;gflops
:CUSTOM_ID: example-define-2-parallel-workers-per-socket-sirocco24
:END:
+#+begin_src sh
+CHAMELEON_PARALLEL_WORKER_LEVEL=socket:2 \
+CHAMELEON_PARALLEL_WORKER_SHOW=1 \
+STARPU_MAIN_THREAD_BIND=1 \
+STARPU_CALIBRATE=1 \
+STARPU_SCHED=dmdar \
+STARPU_NWORKER_PER_CUDA=2 \
+STARPU_SILENT=1 \
+$PTCHAMELEON/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 1440:3000:480 -g 2
+#+end_src
+
#+begin_example
-$ CHAMELEON_PARALLEL_WORKER_LEVEL=socket:2 CHAMELEON_PARALLEL_WORKER_SHOW=1 STARPU_MAIN_THREAD_BIND=1 STARPU_CALIBRATE=1 STARPU_SCHED=dmdar STARPU_NWORKER_PER_CUDA=2 STARPU_SILENT=1 ~/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 1440:3000:480 -g 2
Number of parallel workers created: 4
Parallel worker 0 contains the following logical indexes:
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
@@ -113,20 +129,20 @@ Id;Function;threads;gpus;P;Q;mtxfmt;nb;uplo;n;lda;seedA;tsub;time;gflops
:PROPERTIES:
:CUSTOM_ID: downloading
:END:
-#+begin_example
+#+begin_src sh
# Root directory
PTCHAMELEON=~/PTCHAMELEON
mkdir $PTCHAMELEON
cd $PTCHAMELEON
git clone git@gitlab.inria.fr:starpu/starpu.git
git clone --recursive git@gitlab.inria.fr:solverstack/chameleon.git
-#+end_example
+#+end_src
***** Setup on sirocco16 (2 cpu intel + 2 v100)
:PROPERTIES:
:CUSTOM_ID: setup-on-sirocco16-2-cpu-intel-2-v100
:END:
-#+begin_example
+#+begin_src sh
module load build/cmake/3.15.3 \
linalg/mkl/2022.0.2 \
trace/eztrace/1.1-8 \
@@ -143,28 +159,34 @@ cd $PTCHAMELEON/starpu
./autogen.sh
mkdir build && cd build
# In case you want to debug take the first line
-#../configure --enable-debug --enable-verbose --enable-parallel-worker --disable-opencl --disable-build-doc --enable-maxcpus=64 --disable-socl --prefix=$PTCHAMELEON/starpu/build/install --enable-fxt
-../configure --enable-parallel-worker --disable-opencl --disable-build-doc --enable-maxcpus=64 --disable-socl --prefix=$PTCHAMELEON/starpu/build/install
+#../configure --enable-debug --enable-verbose --enable-parallel-worker --disable-opencl \
+# --disable-build-doc --enable-maxcpus=64 --disable-socl \
+# --prefix=$PTCHAMELEON/starpu/build/install --enable-fxt
+#
+../configure --enable-parallel-worker --disable-opencl --disable-build-doc \
+ --enable-maxcpus=64 --disable-socl \
+ --prefix=$PTCHAMELEON/starpu/build/install
make -j install
source $PTCHAMELEON/starpu/build/install/bin/starpu_env
# Build Chameleon
cd $PTCHAMELEON/chameleon
mkdir build && cd build
-cmake .. -DBLA_VENDOR=Intel10_64lp -DCHAMELEON_KERNELS_MT=ON -DCHAMELEON_ENABLE_EXAMPLE=OFF -DCHAMELEON_USE_CUDA=ON
+cmake .. -DBLA_VENDOR=Intel10_64lp -DCHAMELEON_KERNELS_MT=ON \
+ -DCHAMELEON_ENABLE_EXAMPLE=OFF -DCHAMELEON_USE_CUDA=ON
make -j
# test
-STARPU_SILENT=1
-STARPU_SCHED=dmdar
-CHAMELEON_PARALLEL_WORKER_LEVEL=L3
-CHAMELEON_PARALLEL_WORKER_SHOW=1
-STARPU_MAIN_THREAD_BIND=1
-STARPU_CUDA_PIPELINE=2
-STARPU_NWORKER_PER_CUDA=4
-STARPU_CALIBRATE=1
-$PTCHAMELEON/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 960:3000:480 -g 2
-#+end_example
+STARPU_SILENT=1 \
+ STARPU_SCHED=dmdar \
+ CHAMELEON_PARALLEL_WORKER_LEVEL=L3 \
+ CHAMELEON_PARALLEL_WORKER_SHOW=1 \
+ STARPU_MAIN_THREAD_BIND=1 \
+ STARPU_CUDA_PIPELINE=2 \
+ STARPU_NWORKER_PER_CUDA=4 \
+ STARPU_CALIBRATE=1 \
+ $PTCHAMELEON/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 960:3000:480 -g 2
+#+end_src
***** Setup on sirocco24 (2 cpu amd + 2 a100)
:PROPERTIES:
@@ -172,7 +194,7 @@ $PTCHAMELEON/chameleon/build/testing/chameleon_dtesting -o potrf -n 59520 -b 960
:END:
Identical to sirocco16 except for the Intel MKL library:
-#+begin_example
+#+begin_src sh
module load build/cmake/3.15.3 \
linalg/mkl/2020_update4 \
trace/eztrace/1.1-8 \
@@ -183,11 +205,11 @@ module load build/cmake/3.15.3 \
trace/fxt/0.3.14 \
trace/eztrace/1.1-9 \
language/python
-#+end_example
+#+end_src
Strangely the execution requires the creation of links:
-#+begin_example
+#+begin_src sh
cd $PTCHAMELEON
for lib in libmkl_gf_lp64.so libmkl_gnu_thread.so libmkl_intel_lp64.so libmkl_sequential.so ;
do
@@ -195,5 +217,5 @@ do
done
LD_LIBRARY_PATH=$PTCHAMELEON:$LD_LIBRARY_PATH
-#+end_example
+#+end_src