Skip to content

Latest commit

 

History

History
61 lines (47 loc) · 3.36 KB

TEST.md

File metadata and controls

61 lines (47 loc) · 3.36 KB
Raw

Testing and Benchmarks

Testing

Various applications have been ported to use FPGU API. These are present in programs/ (These are deprecated now as tests use an obselete FGPU feature though these test work functionally. For both testing and benchmarks, see benchmarks/). For an example, refer to benchmarks/cudaSDK/vectorAdd/vectorAdd.cu. File doc/PORT.md explains how to port an application to use FGPU API.

The header file include/fractional_gpu_testing.hpp contains some helper functions that are particularly useful for test/benchmark application.

The method for testing currently is to compare the functional output of a program with and without FGPU API. Since FGPU API is only intended to limit resources available to an application, the functional output should be same.

Adding test applications

To add a new test application to build system, refer to [CMakelists.txt][../CMakelists.txt].

Benchmarks

In micro-benchmark evaluation, we have taken codes from CUDA SDK and Rodinia suite and modified them to use FGPU API. These benchmarks can be found in benchmarks. The methodology for benchmark is as follows:

  1. Configure FGPU into a specific mode (Refer to doc/BUILD.md to see all modes that FGPU supports)
  2. Run a single application on the first partition of GPU. All other partitions are empty. Measure time taken by kernels of the application.
  3. Run an application on the first partition of GPU. On all other partitions, run an interfering application (copies of same application is ran across all the partitions). Measure time taken by kernels of the application of interest.

Amount of variation between 2) and 3) cases highlight the amount of isolation between partitions. For more detailed explanation of the methodology and results of micro-benchmarks, please refer to doc/FGPU-RTAS-2019.pdf.

Configuring benchmarks

The file benchmarks/config_benchmark.sh allows to modify some paramters of benchmarks. Below we list the paramters of importance:

  • benchmarks - This array lists all the applications that we are interested in.
  • interference - This array lists all the interfering application to be ran in parallel along the application of interest.
  • NUM_COLORS - This variable defines how many partitions has the FGPU been split into. The upper limit for this depends on the number of memory colors available in hardware.
  • ENABLE_VOLTA_MPS_PARTITION - On Volta architectures, MPS can be used to restrict amount of computation resources allocated to an application. Enable this option when you wish to evaluate effectiveness of MPS on Volta GPU. The script uses MPS to split the computation resources equally between all partitions of GPU. Note: FGPU should be set in No Partitioning mode when using this option as we rely on MPS to do the partition and not FGPU.

Running benchmarks

After selecting the mode for FGPU and building the code (Refer to doc/BUILD.md), run the following commands:

cd $PROJ_DIR/benchmarks
sudo ./run_benchmark.sh

Note: The benchmark script requires schedtool and taskset to be installed. They can be installed on popular Linux distributions.