LevelDB with CUDA

LUDA: A Customized GPU Acceleration Scheme to Boost LSM Key-Value Store Compactions

• make sure you linux run LevelDB 1.22 (https://github.com/google/leveldb) properly

• CUDA 10.2

• moderngpu from https://github.com/moderngpu/moderngpu

• assume your nvcc path is /usr/local/cuda/bin/nvcc

• assume the path of LUDA is /home/abc/flying/LUDA

  • modify the path to moderngpu head filens in /home/abc/flying/LUDA/cuda/decode_kv.cu
  • modify the path to a test directory leveldb::DB::Open(options, "your_directory", &db); in /home/abc/flying/LUDA/db/paper_test.cc, or compile will fail

|-- AUTHORS
|-- build/
|-- CMakeLists.txt
|-- CONTRIBUTING.md
|-- cuda/
|   |-- cuda_common.h               # luda data structures
|   |-- decode_kv.cu                # luda kernels
|   |-- ...
|-- db/
|   |-- db_impl.cc                  # luda take over compaction
|   |-- ...
|-- doc/
|-- helpers/
|-- include/
|-- issues/
|-- LICENSE
|-- NEWS
|-- port/
|-- README.md
|-- table/
|-- TODO
|-- util/

cd /home/abc/flying/LUDA
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release

• modify the following files in the CMakeFiles to compile and link to CUDA kernels:

  • paper_test.dir/build.make

    • add NVCC=/usr/local/cuda/bin/nvcc to the beginning of this file

    • replace all /usr/bin/c++ with $(NVCC)

    • add the following at the end of the CMakeFiles/paper_test.dir/util/testharness.cc.o: target:

      $(NVCC) $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -o CMakeFiles/paper_test.dir/util/decode_kv.cu.o -c /home/abc/flying/LUDA/cuda/decode_kv.cu

  • paper_test.dir/flags.make

    • get the gencode of your GPU and the following at the end of CXX_FLAGS = ... :

      -std=c++11 -rdc=true -Xptxas -dlcm=ca --extended-lambda -arch=sm_61

  • paper_test.dir/link.txt

    • replace all /usr/bin/c++ with the following :

      /usr/local/cuda/bin/nvcc -arch=sm_61 CMakeFiles/paper_test.dir/util/decode_kv.cu.o

  • leveldb.dir/build.make

    • add NVCC=/usr/local/cuda/bin/nvcc to the beginning of this file

    • replace all /usr/bin/c++ with $(NVCC)

  • db_bench.dir/link.txt

    • replace /usr/bin/c++ with /usr/local/cuda/bin/nvcc -arch=sm_61 CMakeFiles/paper_test.dir/util/decode_kv.cu.o

make sure the above configurations a applied.

cd /home/abc/flying/LUDA/build
cmake --build . --target clean && cmake --build . --target paper_test && cmake --build . --target db_bench
cp ../build._static.sh && ./build_static.sh   # build libleveldb.a on demand