deal-on-gpu

This is the project repo of the deal-on-gpu team at EuroHack19 in Lugano, Switzerland.

The presentation of the final results can be found here.

Team members

The team deal-on-gpu consisted of (in alphabetical order):

• Momme Allalen
• Paddy Ó Conbhuı́ (mentor)
• Prashanth Kanduri (mentor)
• Martin Kronbichler
• Peter Munch

We profited from the dedicated work by:

• Daniel Arndt
• Karl Ljungkvist
• Bruno Turcksin

Installation on Piz Daint

Scripts for building on Piz Daint with gcc can be found in scripts/daint-gcc/.

To build dealii and step-64, just run

./scripts/daint-gcc/make_dealii.sh --download --build-p4est
./scripts/daint-gcc/make_step-64.sh

To resume a build of dealii, or build after a changing the source in build/dealii/src, just run

./scripts/daint-gcc/make_dealii.sh

If you have the dealii source in a different directory, use the --dealii-source-dir=<dealii source> option when running make_dealii.sh. Change the build root with the --build-root=<build root> option for both make_dealii.sh and make_step-64.sh.

Note: LAPACK on Piz Daint is missing a needed linker flag in its config. This problem will manifest in a failure to link the dealii shared library and programs. Add the option -DLAPACK_LINKER_FLAGS="${ATP_POST_LINK_OPTS}" to the dealii cmake command to fix it.

Using Nvprof + NVVP:

We compile on Daint with cudatoolkit 9.1 due to some transitive dependencies from pre-installed modules. However, to profile P100 GPUs with nvprof, we need nvprof from cudatoolkit 9.2.

The following module setup should set the needed environment

module load daint-gpu
module swap cudatoolkit/9.2.148_3.19-6.0.7.1_2.1__g3d9acc8

First, generate a timeline:

srun nvprof -f -o profile-timeline.nvvp ./step-64

And then generate metrics and analysis-metrics for a kernel. To analyze the apply_kernel_shmem kernel, for example, we can run

nvprof -f -o profile-metrics-apply_kernel_shmem.metrics --kernels ::apply_kernel_shmem: --analysis-metrics --metrics all ./step-64

The --kernels syntax is [context]:[nvtx range]:kernel_id:[invocation]. You can leave the optional values blank to match all instances.

From there, you can open the profiles in NVVP. You need to "import...", and then choose the .nvvp file for the timeline, the .metrics file for the metrics, and include the kernel syntax in the kernels panel.

To generate source-level statistics to see stalls, memory accesses, branching etc., add the -lineinfo flag to nvcc, and the --source-level-analysis flags to nvprof e.g.

nvprof -f -o profile-metrics-apply_kernel_shmem.metrics --kernels ::apply_kernel_shmem: --analysis-metrics --metrics all --source-level-analysis global_access,shared_access,branch,instruction_execution,pc_sampling ./step-64

Note the source level analysis will significantly slow down the execution time!

Displaying source-level info in nvvp requires nvdisasm is installed, which should be available in the cuda toolkit.