Differences

This shows you the differences between two versions of the page.

--- gpu_resources [2017/04/27 13:19] – csteel
+++ gpu_resources [2024/03/26 13:52] (current) – external edit 127.0.0.1
@@ Line 2: / Line 2: @@
 This is a collaborative resource, please improve it. Login using your MCIN user name and ID and add your discoveries.
+===== Items of Interest / for Discussion? =====
+==== Resources ====
+* [ OpenACC - Tutorial - Steps to More Science ]( https://developer.nvidia.com/openacc/3-steps-to-more-science )
+"Here are three simple steps to start accelerating your code with GPUs. We will be using PGI OpenACC compiler for C, C++, FORTRAN, along with tools from the PGI Community Edition."
+* [ Performance Portability from GPUs to CPUs with OpenACC ](https://devblogs.nvidia.com/parallelforall/performance-portability-gpus-cpus-openacc/)
+* [ Data Center Management Tools ]( http://www.nvidia.com/object/data-center-managment-tools.html )
+    * The GPU Deployment Kit
+    * Ganglia
+    * Slurm
+    * NVIDIA Docker
+    * Others???
+"...performance on multicore CPUs for HPC apps using MPI + OpenACC is equivalent to MPI + OpenMP code. Compiling and running the same code on a Tesla K80 GPU can provide large speedups."
 ===== Preventing Job Clobbering =====
-Today I was training a model and inadvertently kicked Konrad's job off the GPU. I discovered how to configure TensorFlow so that it doesn't do this:
+There are currently 3 GPU's in ace-gpu-1. To select one of the three (0, 1, 2), set the CUDA_VISIBLE_DEVICES environment variable. This can be accomplished by adding the following line to your ~/.bash_profile file on ace-gpu-1, where X is either 0, 1 or 2:
+<code>
+export CUDA_VISIBLE_DEVICES=X
+</code>
+This will only take effect when you log in, so log out and back in and try the following to ensure that it worked:
+<code>
+echo $CUDA_VISIBLE_DEVICES
+</code>
+If it outputs the ID that you selected then you're ready to use the GPU.
+==== Sharing a single GPU ====
+To configure TensorFlow to not pre-allocate all GPU memory you can use the following Python code:
 <code>
@@ Line 15: / Line 53: @@
 </code>
-We should develop some kind of policy to run jobs on ace-gpu-1 so that we don't inadvertently ruin other peoples' processes.
+This has been found to work only to a certain extent, and when there are several jobs that use a significant amount of the GPU resources, jobs can still be ruined even when using the above code
 ===== GPU Info =====
@@ Line 47: / Line 84: @@
 nsight
 </code>
+Nvidia Visual Profiler (https://developer.nvidia.com/nvidia-visual-profiler) would be useful for GPU monitoring if we had X visualization, but we do not:
+<code>
+/usr/local/cuda/bin/nvvp
+</code>
 ===== GPU Accounting =====
@@ Line 60: / Line 103: @@
 </code>
+Output example:
+<code>
+==============NVSMI LOG==============
+Timestamp                           : Thu Apr 27 09:09:50 2017
+Driver Version                      : 375.39
+Attached GPUs                       : 1
+GPU 0000:01:00.0
+    Accounting Mode                 : Enabled
+    Accounting Mode Buffer Size     : 1920
+    Accounted Processes
+        Process ID                  : 15819
+            GPU Utilization         : 100 %
+            Memory Utilization      : 6 %
+            Max memory usage        : 187 MiB
+            Time                    : 3769 ms
+            Is Running              : 0
+...
+</code>
 Users: to check GPU stats per process:
 <code>
@@ Line 109: / Line 173: @@
                                 Doesn't work with -u or -x flags.
 </code>
+* [[http://docs.nvidia.com/deploy/driver-persistence/index.html#persistence-mode]]
+* [[http://docs.nvidia.com/deploy/driver-persistence/index.html#persistence-daemon]]
 ===== Deep Learning =====