liangminghuang/PROMPT

PROMPT: A Fast and Extensible Memory Profiling Framework

Build

Need to use LLVM/Clang as the compiler for the support of LLVM. Need to make sure LLVMgold is installed.

cd PROMPT
export PATH=${llvm install bin directory}:$PATH
make build # By default install to PROMPT/build
make install # By default install to PROMPT/install

The .rc file is generated in PROMPT/install

Use PROMPT

Preprocessing

1. Generate a single LLVM bitcode file
2. (No need to do it manually if specifying a pair of function and basic block in slamp-driver) Annoteate Metadata ID for functions, basic blocks, and instructions with -metadata-namer pass
3. (optional) Run loop profile to generate loopProf.out

Run PROMPT

1. Source PROMPT.rc generated in the install directory
2. If the program requires command line arguments, export PROFILEARGS
3. Run slamp-driver with one or three command line arguments, (1) the bitcode file name, and optionally (2) function name, and (3) basic block name. When only providing one argument, slamp-driver will look for loop profile output loopProf.out and use it to get all hot loops to profile.

Example:

source PROMPT/install/PROMPT.rc
# Or slamp-driver benchmark.bc if there is proper loop profile
PROFILEARGS="aminos 1234 123" slamp-driver benchmark.plain.bc md for.cond219

Add a New Profiling Module

• Event src/runtime/Events/configs
  • Add a new config file in the events
• Frontend src/runtime/frontend
  • Add in the CMakeLists.txt
  • Modify the FrontendGenerator.py to add a new module
• Logic src/runtime/ProfilingModules
  • Implement and add header .cpp and .h file
  • Register the .cpp file in the CMakeLists.txt
• Backend
  • Currently in SLAMPcustom/consumer/consumer.cpp
    • Add a new loop to process the events, import the header and instantiate the module
  • Will be moved to src/runtime/backend

Make Sense of the Output

The default output is at result.slamp.profile.

When used as a memory dependence profiler, PROMPT generates a output file where each dependence consists of six numbers, following this format:

[loop id, source id, bare destination id, destination id, is loop carried, count]

where:

• loop id: the basic block metadata id of the loop
• source id: the instruction id of the source instruction (have to be an instruction in the loop, can be a function call)
• bare destination id: the instruction id where store happens (can be outside of the loop, not a function call)
• destination id: the instruction id of the source instruction (have to be an instruction in the loop, can be a function call)
• is loop carried: 0 for intra-iteration; 1 for loop-carried
• count: always 1 if counting is turned off, otherwise the dynamic count of the dependence

Note that the first dependence of a loop is always [loop id, 0, 0, 0, 0, 0], showing that a loop has been profiled.

TODOs

Decoupling

• Compile with NOELLE and SCAF
• Compile with NOELLE and SCAF (without Speculation Modules)
• Seperate all profiling modules out
• Make LTO optional
• Convert producer library to be configurable

Implementation

• Replace malloc hook (removed in glibc 2.35)
• External function handling: currently disabled in SLAMP.cpp by not doing the replacements
• External function handling with correct report of allocation events
• Multithreaded profiling?
• Multiple loops at the same time
• Package the components better
  • The queue
  • The container
• Parallel Containers
  • Replace the vector with a lower-latency memory region
• 16 bytes load and store handling

Integration

• Replace SpecPriv profiler and LAMP with PROMPT
  • Check the problem with the failed and long-running benchmarks

Debug & Testing

• Check the slowdown problem with multiple backend running together
• Check the slowdown problem when multiple containers running together
• Test with newer version of LLVM
• Test with way bigger benchmarks

Documentation

• Preprocessing tasks
• Simple demo
• Get started doc
• Extending with new profiler doc

Modules