Skip to content

Latest commit

 

History

History
78 lines (58 loc) · 1.37 KB

notes_vtune.md

File metadata and controls

78 lines (58 loc) · 1.37 KB

Intel VTune (GUI)

vtune-gui with VNC

  • Create docker image:
cd docker_vnc
docker build -t vtune-gui . -f Dockerfile
cd ..
  • Run docker container:
docker run --rm -itd \
    --cap-add=SYS_ADMIN \
    --cap-add=SYS_PTRACE \
    --device=/dev/dri \
    --name=vtune-gui-container \
    -v $PWD:/ws \
    -w /ws \
    -v /home/user/Documents/ml/pytorch/:/pytorch \
    --network=host \
    --security-opt seccomp:unconfined \
    --ipc=host \
    vtune-gui

docker exec -it vtune-gui-container /bin/bash
  • Run profiling as "Remote Linux (SSH)"
# in vtune container:
ssh-keygen -t rsa
cat ~/.ssh/id_rsa.pub

# in pth-dev container:
apt-get install openssh-server
nano ~/.ssh/authorized_keys

  • Run perf snapshot

    • python check_interp.py --use-perf
    • sampling interval 0.01 ms

  • Run hotspot analysis

Run linux perf

  • Set on host:
cat /proc/sys/kernel/perf_event_paranoid
> 4

sudo su
echo "-1" > /proc/sys/kernel/perf_event_paranoid

cat /proc/sys/kernel/perf_event_paranoid
> -1
  • Run linux perf:

// (https://www.slideshare.net/emBO_Conference/profiling-your-applications-using-the-linux-perf-tools) // perf record --call-graph dwarf -F 40 -- ./bench 0 0 0 1 // perf record --call-graph dwarf -F 40 -- ./bench 1 0 0 1 // perf report

perf record --call-graph dwarf -F 40 -- python -u check_interp.py --use-perf
perf report