README.md

DataCrumbs

A multi-level dataflow profiler to capture I/O calls from workflows.

Dependency installation

Requirements

• Newer Kernel > 5.6 see kernel doc
• BCC > v0.30.0
• cmake
• python 3.10
  • hydra-core>=1.2.0

Test Requirements

• OpenMPI

Chameleon

We have a Image that can be used to do development on Chameleon called HARI-UBUNTU-22.04.04-BCC.

Compiling the code

Clone the repo

git clone https://github.com/hariharan-devarajan/datacrumbs.git

cd datacrumbs
mkdir build
cd build
cmake ..
make -j

Build dependency for profiler

sudo pip install -r requirements.txt

Running the code

The profiler tool need to run as root

sudo su
export PYTHONPATH=<PATH to datacrumbs>
cd <PATH to datacrumbs Root>
python3 datacrumbs/main.py

Once the profiler is loaded, it will wait for applictaions to connect.

Environment variable to consider

Check ulimit -n for open file handlers

ulimit -n 1048576

Increase probe limit within BCC

export BCC_PROBE_LIMIT=1048576

Running the test

Once the profiler has started u can run the application code.

cd <PATH to datacrumbs Root>
cd tests/scripts
./run.sh

Checking the profiler output.

The profiler output is created in the directory where the profiler runs.

cc@ebpf:~/datacrumbs$ head -n 5 profile.pfw 
[
{"pid": 30545, "tid": 30545, "name": "__libc_malloc [libc.so.6]", "cat": "[libc.so.6]", "ph": "C", "ts": 0.0, "args": {"count": 21, "time": 0.000198116}}
{"pid": 30545, "tid": 30545, "name": "cfree [libc.so.6]", "cat": "[libc.so.6]", "ph": "C", "ts": 0.0, "args": {"count": 2, "time": 1.9788e-05}}
{"pid": 30545, "tid": 30545, "name": "[unknown]", "cat": "unknown", "ph": "C", "ts": 0.0, "args": {"count": 1, "time": 1.3094e-05}}
{"pid": 30545, "tid": 30545, "name": "__libc_malloc [libc.so.6]", "cat": "[libc.so.6]", "ph": "C", "ts": 24000000.0, "args": {"count": 149, "time": 0.000503765}}

Visualizing the counters

The output format used is Chrome Tracing format and can be viewed using perfetto.

Analyzing counters using queries

The profiler output can be analyzed using Dask distributed analysis. Please refer to the notebook.