Merge pull request #32 from AaltoSciComp/nvidia-performance-counters

Add functionality for enabling nvidia performance counters
fgci-org · Sep 18, 2020 · 00da97a · 00da97a
2 parents bd84791 + bba5000
commit 00da97a
Show file tree

Hide file tree

Showing 3 changed files with 10 additions and 0 deletions.
diff --git a/defaults/main.yml b/defaults/main.yml
@@ -20,4 +20,6 @@ cuda_bash_cuda_root: "/usr/local/cuda"
 cuda_bash_cuda_inc_dir: "/usr/local/cuda/bin"
 cuda_bash_cpath: "/usr/local/cuda/include"
 
+cuda_enable_perf_counters: False
+
 # vim:ft=ansible:
diff --git a/tasks/main.yml b/tasks/main.yml
@@ -35,6 +35,13 @@
   - include_tasks: cuda_init.yml
     when: cuda_init
 
+  - name: Enable performance counters for all users via modprobe
+    template:
+      src: nvidia.conf.j2
+      dest: /etc/modprobe.d/nvidia.conf
+      mode: 0644
+    when: cuda_enable_perf_counters
+
   # This is here because if we in the same playbook try to start slurmd without
   # having run the cuda_init.sh script then slurmd doesn't start and the play fails.
   # todo: reload nvidia modules/etc instead of restart

diff --git a/templates/nvidia.conf.j2 b/templates/nvidia.conf.j2
@@ -0,0 +1 @@
+options nvidia "NVreg_RestrictProfilingToAdminUsers=0"
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		options nvidia "NVreg_RestrictProfilingToAdminUsers=0"