CuSan ·

CuSan is tool to find data races between (asynchronous) CUDA calls and the host. To that end, we analyze and instrument CUDA codes to track CUDA domain-specific memory accesses and synchronization semantics during compilation using LLVM. Our runtime then passes these information appropriately to ThreadSanitizer (packaged with Clang/LLVM) for the final data race analysis.

Usage

Making use of CuSan consists of two phases:

1. Compile your code with Clang/LLVM (version 14) using one the CuSan compiler wrappers, e.g., cusan-clang++ or cusan-mpic++. This will (a) analyze and instrument the CUDA API appropriately, such as kernel calls and their particular memory access semantics (r/w), (b) add ThreadSanitizer instrumentation, and (c) finally link our runtime library.
2. Execute the target program for the data race analysis. Our runtime internally calls ThreadSanitizer to expose the CUDA synchronization and memory access semantics.

Checking CUDA-aware MPI applications

You need to use the MPI correctness checker MUST, or preload our (very) simple MPI interceptor libCusanMPIInterceptor.so for CUDA-aware MPI data race detection. These libraries call ThreadSanitizer with the particular access semantics of MPI. Therefore, the combined semantics of CUDA and MPI are properly exposed to ThreadSanitizer to detect data races of data dependent MPI and CUDA calls.

Example report

The following is an example report for 03_cuda_to_mpi.c of our test suite, where the necessary synchronization is not called:

L.23  __global__ void kernel(int* arr, const int N)
...
L.58  int* d_data;
L.59  cudaMalloc(&d_data, size * sizeof(int));
L.60
L.61  if (world_rank == 0) {
L.62    kernel<<<blocksPerGrid, threadsPerBlock>>>(d_data, size);
L.63  #ifdef CUSAN_SYNC
L.64    cudaDeviceSynchronize();  // CUSAN_SYNC needs to be defined
L.65  #endif
L.66    MPI_Send(d_data, size, MPI_INT, 1, 0, MPI_COMM_WORLD);

==================
WARNING: ThreadSanitizer: data race (pid=689288)
  Read of size 8 at 0x7fb51f200000 by main thread:
    #0 main cusan/test/pass/03_cuda_to_mpi.c:66:5 (03_cuda_to_mpi.c.exe+0x4e8448)

  Previous write of size 8 at 0x7fb51f200000 by thread T6:
    #0 __device_stub__kernel(int*, int) cusan/test/pass/03_cuda_to_mpi.c:23:47 (03_cuda_to_mpi.c.exe+0x4e81ef)

  Thread T6 'cuda_stream' (tid=0, running) created by main thread at:
    #0 __pool_create_fiber_dbg cusan/build/_deps/fiber_pool-src/fiberpool.cpp:538:16 (libCusanFiberpool-d.so+0x1c152)
    #1 main cusan/test/pass/03_cuda_to_mpi.c:59:3 (03_cuda_to_mpi.c.exe+0x4e8331)

SUMMARY: ThreadSanitizer: data race cusan/test/pass/03_cuda_to_mpi.c:66:5 in main
==================
ThreadSanitizer: reported 1 warnings

Building cusan

cusan requires LLVM version 14 and CMake version >= 3.20. Use CMake presets develop or release to build.

Dependencies

CuSan was tested with:

• System modules: 1) gcc/11.2.0 2) cuda/11.8 3) openmpi/4.1.6 4) git/2.40.0 5) python/3.10.10 6) clang/14.0.6
• External libraries: TypeART (https://github.com/tudasc/TypeART/tree/feat/cuda), FiberPool (optional, default off)
• Testing: llvm-lit, FileCheck
• GPU: Tesla T4 and Tesla V100 (mostly: arch=sm_70)

Build example

cusan uses CMake to build. Example build recipe (release build, installs to default prefix ${cusan_SOURCE_DIR}/install/cusan)

$> cd cusan
$> cmake --preset release
$> cmake --build build --target install --parallel

Build options

Option	Default	Description
CUSAN_FIBERPOOL	OFF	Use external library to efficiently manage fibers creation .
CUSAN_SOFTCOUNTER	OFF	Runtime stats for calls to ThreadSanitizer and CUDA-callbacks. Only use for stats collection, not race detection.
CUSAN_SYNC_DETAIL_LEVEL	ON	Analyze, e.g., memcpy and memcpyasync w.r.t. arguments to determine implicit sync.
CUSAN_LOG_LEVEL_RT	3	Granularity of runtime logger. 3 is most verbose, 0 is least. For release, set to 0.