First update for SC25

Author: AndiH

.archive.mk

@@ -6,16 +6,16 @@
 # Changelog:
 # * Nov 2022: The archive is extracted again, then slides.pdf is removed if a patched slides-sc22.pdf is found (which includes an SC22 slide 0 title slide); and then repackaged
 .PHONY: all
-all: tut105-multi-gpu.tar.gz
+all: tut113-multi-gpu.tar.gz
 
-SOURCES=$(shell gfind . -maxdepth 1 -mindepth 1 -not -path "./.*" -not -name "tut105-multi-gpu.tar.gz" -printf '%P\n' | sort -h)
+SOURCES=$(shell gfind . -maxdepth 1 -mindepth 1 -not -path "./.*" -not -name "tut113-multi-gpu.tar.gz" -printf '%P\n' | sort -h)
 
-tut105-multi-gpu.tar.gz: $(shell find . -not -name "tut105-multi-gpu.tar.gz")
+tut113-multi-gpu.tar.gz: $(shell find . -not -name "tut113-multi-gpu.tar.gz")
 	sed -i '1 i***Please check GitHub repo for latest version of slides: https://github.com/FZJ-JSC/tutorial-multi-gpu/ ***\n' README.md
-	tar czf $@ --transform 's,^,ISC25-tut105-Multi-GPU/,' --exclude=".*" $(SOURCES)
+	tar czf $@ --transform 's,^,SC25-tut113-Multi-GPU/,' --exclude=".*" $(SOURCES)
 	tar xf $@
 	rm $@
-	find ISC25-tut105-Multi-GPU/ -not -path './.*' -iname 'slides-*.pdf' -execdir rm slides.pdf \;
-	tar czf $@ ISC25-tut105-Multi-GPU
-	rm -rf ISC25-tut105-Multi-GPU
+	find SC25-tut113-Multi-GPU/ -not -path './.*' -iname 'slides-*.pdf' -execdir rm slides.pdf \;
+	tar czf $@ SC25-tut113-Multi-GPU
+	rm -rf SC25-tut113-Multi-GPU
 	sed -i '1,2d' README.md

.etc/.set-facl-permissions.sh

@@ -2,7 +2,7 @@
 
 set -x
 
-for user in haghighimood1 kraus1 hrywniak1 oden1 garciadegonzalo1 badwaik1 john2; do
+for user in haghighimood1 kraus1 hrywniak1 oden1 garciadegonzalo1 badwaik1 john2 appelhans1; do
 	setfacl -m u:$user:rwx -R $PROJECT_training2446/common/
 	setfacl -m u:$user:rwx -R $PROJECT_training2446/env.sh
 done

.etc/deploy-material.sh

@@ -1 +1 @@
-rsync --archive --exclude="*minified.pdf" --exclude="tut*" --exclude=".*" --exclude="*-sc*.pdf" --verbose ../ judac:/p/project1/training2526/common/material/
+rsync --archive --exclude="*minified.pdf" --exclude="tut*" --exclude=".*" --exclude="*-sc*.pdf" --verbose ../ judac:/p/project1/training2555/common/material/

.etc/deploy.sh

@@ -1 +1 @@
-rsync --archive --exclude="deploy.sh" --exclude="raw/" --exclude="sc2*-titleslides/" --verbose . judac:/p/project1/training2526/common/environment/
+rsync --archive --exclude="deploy.sh" --exclude="raw/" --exclude="sc2*-titleslides/" --verbose . judac:/p/project1/training2555/common/environment/

.etc/instructions-header.md

@@ -1,6 +1,6 @@
-# SC24 Tutorial: Efficient Distributed GPU Programming for Exascale
+# SC25 Tutorial: Efficient Distributed GPU Programming for Exascale
 
--   Sunday, November 17, 2024 8:30 AM to 5:30 PM
--   Location: B211, Atlanta Convention Center, Georgia, USA
+-   Sunday, November 16, 2025 8:30 AM to 5:00 PM
+-   Location: Room 127, St. Louis Convention Center, St. Louis, USA
 -   Program Link:
-    https://sc24.conference-program.com/presentation/?id=tut123&sess=sess412
+    https://sc25.conference-program.com/presentation/?id=tut113&sess=sess252

.etc/jsccourse-bashrc.sh

@@ -11,10 +11,10 @@
 # Andreas Herten, >2017
 ################################################
 if [ -z "$_JSCCOURSE_ENV_SOURCED" ]; then
-	project="training2526"
+	project="training2555"
 
 	export JSCCOURSE_DIR_GROUP=/p/project1/$project
-	export JSCCOURSE_DIR_LOCAL=${JSCCOURSE_DIR_LOCAL_BASE:-$HOME}/ISC25-Multi-GPU-Tutorial
+	export JSCCOURSE_DIR_LOCAL=${JSCCOURSE_DIR_LOCAL_BASE:-$HOME}/SC25-Multi-GPU-Tutorial
 
 	export _JSCCOURSE_ENV_SOURCED="$(date)"
 	export C_V_D="0,1,2,3"
@@ -24,7 +24,7 @@ if [ -z "$_JSCCOURSE_ENV_SOURCED" ]; then
 	res=""
 	currentday=$(date +%d)
 	if [[ "$currentday" == "13" ]]; then
-		res="--reservation isc25-mgpu"
+		res="--reservation sc25-mgpu"
 	fi
 
 	export SLURM_NTASKS=1
@@ -120,7 +120,7 @@ if [[ $- =~ "i" ]]; then
 
 	echo ""
 	echo "*******************************************************************************"
-	echo "       Welcome to the ISC25 Tutorial on Multi-GPU Computing for Exascale!       "
+	echo "       Welcome to the SC25 Tutorial on Multi-GPU Computing for Exascale!       "
 	# echo " A default call to get a batch system allocation is stored in \$JSC_ALLOC_CMD!"
 	# echo " Use it with \`eval \$JSC_ALLOC_CMD\`. The value of \$JSC_ALLOC_CMD is:"
 	# echo -n "  "

.etc/modules.sh

@@ -1,10 +1,10 @@
 module purge
 module load GCC/13.3.0
-module load CUDA/12
+module load CUDA/12 #12.6.0
 module load OpenMPI/5.0.5
 export MPI_HOME=$EBROOTOPENMPI
 #export MPI_HOME=$EBROOTPSMPI
-module load NCCL/default-CUDA-12
+module load NCCL/default-CUDA-12 #2.22.3-1
 module load NVSHMEM/3.1.7-CUDA-12
 module load Nsight-Systems/2025.3.1
 module load MPI-settings/CUDA

.zenodo.json

@@ -24,26 +24,31 @@
             "orcid": "0000-0002-9670-5296",
             "affiliation": "FernUni Hagen",
             "name": "Oden, Lena"
+        },
+        {   
+            "orcid": "0000-0003-1433-9198",
+            "affiliation": "NVIDIA",
+            "name": "Appelhans, David"
         }
     ],    
 
     "title": "Efficient Distributed GPU Programming for Exascale",
 
-    "publication_date": "2025-06-13",
+    "publication_date": "2025-11-16",
 
-    "description": "<p>Over the past decade, GPUs became ubiquitous in HPC installations around the world, delivering the majority of performance of some of the largest supercomputers (e.g. Summit, Sierra, JUWELS Booster). This trend continues in the recently deployed and upcoming Pre-Exascale and Exascale systems (JUPITER, LUMI, Leonardo; El Capitan, Frontier, Aurora): GPUs are chosen as the core computing devices to enter this next era of HPC.To take advantage of future GPU-accelerated systems with tens of thousands of devices, application developers need to have the proper skills and tools to understand, manage, and optimize distributed GPU applications.In this tutorial, participants will learn techniques to efficiently program large-scale multi-GPU systems. While programming multiple GPUs with MPI is explained in detail, also advanced tuning techniques and complementing programming models like NCCL and NVSHMEM are presented. Tools for analysis are shown and used to motivate and implement performance optimizations. The tutorial teaches fundamental concepts that apply to GPU-accelerated systems in general, taking the NVIDIA platform as an example. It is a combination of lectures and hands-on exercises, using a development system for JUPITER (JEDI), for interactive learning and discovery.</p>",
+    "description": "<p>Over the past decade, GPUs became ubiquitous in HPC installations around the world, delivering the majority of performance of some of the largest supercomputers, steadily increasing the available compute capacity. Finally, four exascale systems are deployed (Frontier, Aurora, El Capitan, JUPITER), using GPUs as the core computing devices for this era of HPC.  To take advantage of these GPU-accelerated systems with tens of thousands of devices, application developers need to have the proper skills and tools to understand, manage, and optimize distributed GPU applications.  In this tutorial, participants will learn techniques to efficiently program large-scale multi-GPU systems. While programming multiple GPUs with MPI is explained in detail, also advanced tuning techniques and complementing programming models like NCCL and NVSHMEM are presented. Tools for analysis are shown and used to motivate and implement performance optimizations. The tutorial teaches fundamental concepts that apply to GPU-accelerated systems of any vendor in general, taking the NVIDIA platform as an example. It is a combination of lectures and hands-on exercises, using the JUPITER system for interactive learning and discovery.</p>",
 
-    "notes": "Slides and exercises of tutorial presented at ISC High Performance 2025; https://isc.app.swapcard.com/widget/event/isc-high-performance-2025/planning/UGxhbm5pbmdfMjU4MTc5Ng==",
+    "notes": "Slides and exercises of tutorial presented at SC25 (The International Conference for High Performance Computing, Networking, Storage, and Analysis 2025); https://sc25.conference-program.com/presentation/?id=tut113&sess=sess252",
 
     "access_right": "open",
 
-    "conference_title": "ISC 2025",
-    "conference_acronym": "ISC25",
-    "conference_dates": "10 June-13 June 2025",
-    "conference_place": "Hamburg, Germany",
-    "conference_url": "https://www.isc-hpc.com/",
+    "conference_title": "SC 2025",
+    "conference_acronym": "SC25",
+    "conference_dates": "16 November-21 November 2025",
+    "conference_place": "St. Louis, Missouri, USA",
+    "conference_url": "https://sc25.supercomputing.org/",
     "conference_session": "Tutorials",
-    "conference_session_part": "Afternoon",
+    "conference_session_part": "Day 1",
 
     "upload_type": "lesson"
 }

03-H_Multi_GPU_Parallelization/.master/Instructions.ipynb

@@ -4,14 +4,13 @@
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "# SC24 Tutorial: Efficient Distributed GPU Programming for Exascale\n",
+        "# SC25 Tutorial: Efficient Distributed GPU Programming for Exascale\n",
         "\n",
-        "-   Sunday, November 17, 2024 8:30 AM to 5:30 PM\n",
-        "-   Location: B211, Atlanta Convention Center, Georgia, USA\n",
-        "-   Program Link:\n",
-        "    https://sc24.conference-program.com/presentation/?id=tut123&sess=sess412\n",
-        "\n",
-        "## Hands-On 3: Multi-GPU Parallelization with CUDA-aware MPI\n",
+        "- Sunday, November 16, 2025 8:30 AM to 5:00 PM\n",
+        "- Location: Room 127, St. Louis Convention Center, St. Louis, USA\n",
+        "- Program Link:\n",
+        "  https://sc25.conference-program.com/presentation/?id=tut113&sess=sess252\n",
+        "  \\## Hands-On 3: Multi-GPU Parallelization with CUDA-aware MPI\n",
         "\n",
         "### Task: Parallelize Jacobi Solver for Multiple GPUs using CUDA-aware MPI\n",
         "\n",
@@ -27,23 +26,23 @@
         "and `POP` macros). Once you are familiar with the code, please work on\n",
         "the `TODOs` in `jacobi.cu`:\n",
         "\n",
-        "-   Get the available GPU devices and use it and the local rank to set\n",
-        "    the active GPU for each process\n",
-        "-   Compute the top and bottom neigbhors. We are using\n",
-        "    reflecting/periodic boundaries on top and bottom, so rank0’s Top\n",
-        "    neighbor is (size-1) and rank(size-1) bottom neighbor is rank 0\n",
-        "-   Use MPI_Sendrecv to exchange data between the neighbors\n",
-        "    -   use CUDA-aware MPI, so the send - and the receive buffers are\n",
-        "        located in GPU-memory\n",
-        "    -   The first newly calculated row (‘iy_start’) is sent to the top\n",
-        "        neigbor and the bottom boundary row (`iy_end`) is received from\n",
-        "        the bottom process.\n",
-        "    -   The last calculated row (`iy_end-1`) is send to the bottom\n",
-        "        process and the top boundary (`0`) is received from the top\n",
-        "    -   Don’t forget to synchronize the computation on the GPU before\n",
-        "        starting the data transfer\n",
-        "    -   use the self-defined MPI_REAL_TYPE. This allows an easy switch\n",
-        "        between single- and double precision\n",
+        "- Get the available GPU devices and use it and the local rank to set the\n",
+        "  active GPU for each process\n",
+        "- Compute the top and bottom neigbhors. We are using reflecting/periodic\n",
+        "  boundaries on top and bottom, so rank0’s Top neighbor is (size-1) and\n",
+        "  rank(size-1) bottom neighbor is rank 0\n",
+        "- Use MPI_Sendrecv to exchange data between the neighbors\n",
+        "  - use CUDA-aware MPI, so the send - and the receive buffers are\n",
+        "    located in GPU-memory\n",
+        "  - The first newly calculated row (‘iy_start’) is sent to the top\n",
+        "    neigbor and the bottom boundary row (`iy_end`) is received from the\n",
+        "    bottom process.\n",
+        "  - The last calculated row (`iy_end-1`) is send to the bottom process\n",
+        "    and the top boundary (`0`) is received from the top\n",
+        "  - Don’t forget to synchronize the computation on the GPU before\n",
+        "    starting the data transfer\n",
+        "  - use the self-defined MPI_REAL_TYPE. This allows an easy switch\n",
+        "    between single- and double precision\n",
         "\n",
         "Compile with\n",
         "\n",
@@ -61,17 +60,17 @@
         "\n",
         "### Description\n",
         "\n",
-        "-   The work distribution of the first task is not ideal, because it can\n",
-        "    lead to the process with the last rank having to calculate\n",
-        "    significantly more than all the others. Therefore, the load\n",
-        "    distribution is to be optimized in this task.\n",
-        "-   Compute the `chunk_size` that each rank gets either (ny - 2) / size\n",
-        "    or (ny - 2) / size + 1 rows.\n",
-        "-   Compute how many processes get (ny - 2) / size resp (ny - 2) /\n",
-        "    size + 1 rows\n",
-        "-   Adapt the computation of (`iy_start_global`)"
+        "- The work distribution of the first task is not ideal, because it can\n",
+        "  lead to the process with the last rank having to calculate\n",
+        "  significantly more than all the others. Therefore, the load\n",
+        "  distribution is to be optimized in this task.\n",
+        "- Compute the `chunk_size` that each rank gets either (ny - 2) / size or\n",
+        "  (ny - 2) / size + 1 rows.\n",
+        "- Compute how many processes get (ny - 2) / size resp (ny - 2) / size +\n",
+        "  1 rows\n",
+        "- Adapt the computation of (`iy_start_global`)"
       ],
-      "id": "e42b5ab3-f626-4da5-b0c9-52a444cefde8"
+      "id": "8b73eab2-1e9f-42a8-b366-29ff21d469ea"
     }
   ],
   "nbformat": 4,

03-H_Multi_GPU_Parallelization/.master/Instructions.md

@@ -1,10 +1,9 @@
-# SC24 Tutorial: Efficient Distributed GPU Programming for Exascale
+# SC25 Tutorial: Efficient Distributed GPU Programming for Exascale
 
--   Sunday, November 17, 2024 8:30 AM to 5:30 PM
--   Location: B211, Atlanta Convention Center, Georgia, USA
+-   Sunday, November 16, 2025 8:30 AM to 5:00 PM
+-   Location: Room 127, St. Louis Convention Center, St. Louis, USA
 -   Program Link:
-    https://sc24.conference-program.com/presentation/?id=tut123&sess=sess412
-
+    https://sc25.conference-program.com/presentation/?id=tut113&sess=sess252
 ## Hands-On 3: Multi-GPU Parallelization with CUDA-aware MPI
 
 ### Task: Parallelize Jacobi Solver for Multiple GPUs using CUDA-aware MPI