romnn / microgpusim Goto Github PK

GPUcachesim is a cycle-level, trace-driven, parallel GPU simulator written in Rust.

As of now, the simulator is validated for the NVIDIA Pascal architecture but extensible to model various hardware configurations.

• provide a modular and extensible simulation framework
• support for fast, multi-threaded simulation powered by Rust
• provide pre-configured base configurations for hardware
• usability-first: we aim to improve UX and DX over existing simulators

Note: GPUcachesim is evolving rapidly at the moment, hence API's and code may undergo large changes in the near future. For that reason, we restrain from publishing versioned packages to https://crates.io just yet. However, it is absolutely possible to clone or fork this repository to try things out.

• Step 0: Build GPUcachesim from source

  git clone https://github.com/romnn/gpucachesim
  cd gpucachesim
  cargo build --release # build the simulator
  cargo build -p trace --release # build the tracer

• Step 1: Trace an application

  GPUcachesim is a trace-driven simulator, hence we must first trace an input application. Any compiled CUDA application should work!

  TRACES_DIR=./traces/ LD_PRELOAD=./target/release/libtrace.so <executable> [args]

  We do provide a few test applications. Assuming a working CUDA compilation toolchain, you can build our simple vectoradd_l1_enabled application for testing:

  make -Bj -C ./test-apps/vectoradd/
  TRACES_DIR=./traces/ LD_PRELOAD=./target/release/libtrace.so ./test-apps/vectoradd/vectoradd_l1_enabled 100 32
  ls ./traces/ # allocations.json, commands.json, kernel-0.msgpack

  After tracing, the ./traces directory will contain the following files:

  • allocations.json contains a list of traced memory allocations.
  • commands.json contains all traced CUDA commands, such as CUDA memory transfers and kernel launches.
  • kernel-<ID>.msgpack contains the binary encoded instruction trace for each kernel based on its unique kernel launch ID.

• Step 2: Simulate the trace

  To simulate the traced application, just pass commands.json to GPUcachesim:

  ./target/release/gpucachesim ./traces/commands.json

  To use deterministic parallel simulation, use the --parallel flag. For maximum performance, try --nondeterministic 10. For more available options, see gpucachesim --help.

todo

todo