AdaptMemBench is configurable benchmarking framework to measure memory system performance of compute kernels extracted from applications. This framework can be used to explore the design landscape of a target architecture and to guide application-specific optimization decisions. AdaptMemBench supports reproducibility in experimental results and facilitates sharing results. Accommodation of the polyhedral model in the build system provides a convenient testbed for potential code optimizations. Access to hardware performance counters offers insights on the interaction of the application with the memory subsystem.
This framework requires the following to be installed:
- GNU GCC (version 6.4 or later)
- Python (version 2.7 or later)
- PAPI (The Performance API)
- ISL (Integer Set Library)
- matplotlib library (version 3.0.3)
Generating Benchmark Drivers:
Benchmark drivers can be generated for kernels implemented with unified data spaces or distinct (independent) data spaces - use adaptMemBench_gen_unifiedDS and adaptMemBench_gen_distinctDS generation scripts respectively. Drivers instrumented with PAPI can be generated using adaptMemBench_gen_papi_unifiedDS and adaptMemBench_gen_papi_distinctDS
cd src
python ../generators/adaptMemBench_gen_<option> <kernel>
This generates a benchmark driver named adapt_membench_<option>_<kernel>.cpp
To compile the driver:
cd -
make adapt_membench_<option>_<kernel>
Executing the benchmark driver:
./bin/adapt_membench_<option>_<kernel> --wss <data size value> --num_threads <number of threads> --ntimes <number of repetitions> ...
To generate a stencil 1D benchmark implemented with unified data spaces,
cd src
python ../generators/adaptMemBench_gen_unifiedDS stencil_1d
cd -
make adaptMemBench_gen_unifiedDS_stencil_1d
./bin/adapt_membench_unifiedDS_stencil_1d --wss 1048576 --num_threads 8 --ntimes 1000
For detailed instructions, please refer documentation