Triad GPU Bandwidth

Test achievable triad kernel bandwidth using all available CUDA allocation/transfer methodologies:

• pageable allocation and explicit transfer: cudaMalloc / cudaMemcpy
• pinned allocatiion and explicit transfer: cudaHostAlloc / cudaMemcpy
• pinned allocation with mapping: cudaHostAlloc(..., cudaHostAllocMapped)
• unified memory: cudaMallocManaged
• unified memory with AccessedBy hint: cudaMallocManaged + cudaMemAdvise(..., cudaMemAdviseAccessedBy)
• unified memory with prefetching: cudaMallocManaged + cudaMemPrefetchAsync
• system allocator: new

The basic operation is a host-to-device transfer, followed by the triad kernel, followed by a device-to-host transfer.

The output is a CSV file with columns for the transfer time (zero for implicit transfers), the kernel time, and the total time.

• total time: the time from the start of the first transfer to the end of the last transfer.
• transfer time: the combined time for the host-to-device and the device-to-host transfers.
• kernel time: start of the kernel execution to end of the kernel execution

transfer + kernel may not equal total, though they should be close.

Examples

• Run benchmarks on GPU 0 from n = 1e5; n <= 2.5e8; n *= 1.3. Repeat each benchmark 5 times. Pin access and allocations to NUMA node 0. Show output on terminal and also pipe to triad.csv.

numactl -p 0 ./triad | tee triad.csv

• Run only the pinned benchmark

./triad --pinned

• Run only n = 1e9

./triad -n 1e9

• Run 3 iterations of each benchmark

./triad -i 3

• Show all options

./triad -h

Automatic Testing for Functional System Allocator

Forks a child process to test whether CUDA can use the system allocator. If CUDA cannot, this causes a sticky error that permanently damages the CUDA context, so we use a child process to fully isolate the context so it can be completely destroyed. The child process needs to be create before the parent does any CUDA activity at all.

The test is done in test_system_allocator.cu

Getting stable benchmark results

It is important to do 3 things:

1. Call cudaDeviceReset() before each benchmark. This ensures that any CUDA state is wiped between runs.
2. Call cudaFree(0) after cudaDeviceReset(). This initializes the GPU, ensuring that we don't actidentally time any lazy initialization.
3. Pin to a single NUMA region or CPU. This ensures that data copies always take a consistent route from CPU to GPU.

Building with a gcc that has std::regex

mkdir build && cd build
cmake ..

Building on Power9 with gcc 4.8.5

GCC 4.8.5 doesn't have a working std::regex (used for cxxopts), so install a supported version of clang. GCC 4.8.5 cannot build libcxx, so we use a clang without libcxx to build a clang with libcxx. Depending on your installed CUDA, you'll need a different version of clang.

CUDA	Clang	Installer
9.2	5.0.0	https://gist.github.com/cwpearson/c5521dfc50175b1d977643b2fc5a2bb1
10.1	5.0.0	https://gist.github.com/cwpearson/c13ac7c25bde8c8644300e211faf4e78

Add the clang to your path, and have CMake use clang in the build.

mkdir build && cd build
cmake .. -DCMAKE_TOOLCHAIN_FILE=`readlink -f ../toolchains/clang.toolchain`

The CUDA documentation claims that clang 8.0.0 is supported for CUDA 10.1, but if you actually try it says it requires clang>=3.2 and clang<8. Clang 7.1.0 fails on CUDA 10.0.0 with some errors about __fp16.