Tail Latency Debugger

Quick start

System Build

1. Clone this repository

• git clone git@github.com:inhocho89/llvm14-ldb.git

2. Compile LLVM

• llvm14-ldb> mkdir build
• llvm14-ldb> cd build
• llvm14-ldb/build> cmake -DLLVM_TARGETS_TO_BUILD=X86 -DLLVM_TARGET_ARCH=X86 -DCMAKE_BUILD_TYPE="Release" -DLLVM_BUILD_EXAMPLES=1 -DLLVM_INCLUDE_EXAMPLES=1 -DLLVM_ENABLE_PROJECTS="clang" -G "Unix Makefiles" ../llvm
• llvm14-ldb/build> cmake --build . -j

3. Compile LDB library

• llvm14-ldb/build> cd ../libldb
• llvm14-ldb/libldb> make

Running sample application (Please take a look apps/demo/Makefile to see how to write Makefile)

1. Compile sample application

• llvm14-ldb> cd apps/demo
• llvm14-ldb/apps/demo> make

2. (optional) run perf to get Linux scheduling data (in a separate terminal)

• llvm14-ldb/apps/demo> sudo perf sched record -k CLOCK_MONOTONIC_RAW

3. Run application with preloading LDB shim layer.

• llvm14-ldb/apps/demo> LD_PRELOAD=/path/to/llvm14-ldb/libldb/libshim.so ./demo

4. When enough data is collected (when the application exits), stop the perf

Parsing the data

The raw data is in ldb.data (by LDB) and perf.data (by perf).

1. Parsing latency distribution for each function (this will only use LDB data)

• llvm14-ldb/apps/demo> python3 ../../scripts/parse_summary.py demo > summary

2. Parsing timeline of a request

• llvm14-ldb/apps/demo> python3 ../../scripts/parse_req.py demo 100 > req100