HuntKTm

HuntKTm: Hybrid Scheduling and Automatic Management for Efficient Kernel Execution on Modern GPUs

Requirements

• cmake (version 3.27.7)
• llvm (version 14.0.6)
• cuda (version 12.3)

Note: The operating system we used is Debian 10.2.1 and spark is used to manage softwares with different versions.

Building

git clone https://github.com/Gemini321/HuntKTm.git
cd HuntKTm
./build.sh

Note: Ensure libstatus.so is on your LD_LIBRARY_PATH.

Compilation

Compiling the benchmarks with script:

./compile.sh

If you want to compile your program manunally, use libSchedulerPass.so and libWrapperPass.so to transform the LLVM bitcode (either with opt or clang). Then compile and link runtime library to your program. Refer to the Makefile located in ./benchmarks/multi_stream/b1 for guidance.

Running

Running multi-task or single-task experiment as below:

# run multi-task experiment
cd driver
./run_multi_task.sh

# run single-task experiment
./run_single_task.sh

If you want to run a single task without the scripts we provides, there are two programs to be run: RuntimeScheduler and your_program. Note that RuntimeScheduler should be run before any of your programs.