cluster:234
This is an old revision of the document!
Cuda Checkpoint
- source and user manual at
- located in n91:/projects/cuda-checkpoint-main/580.65.06
- utility linked in n91:/usr/local/bin/cuda-checkpoint
- do not find a way to download 550 features version
- for our mwgpu256 gpu compute nodes
Put your checkpoints in /sanscratch/checkpoints/$SLURM_JOB_ID directory. You need to make that directory in your submit script but do not have to clean it up. Checkpoints older than 120 days are automatically removed.
[root@n91 ~]# rpm -qa criu criu-3.18-5.module+el8.10.0+2001+6a33db9f.x86_64 [root@n91 ~]# rpm -ql criu /run/criu /usr/lib/.build-id /usr/lib/.build-id/9b /usr/lib/.build-id/9b/603c4fff3830838f46feaf4bd1204adc49421e /usr/lib/tmpfiles.d/criu.conf /usr/sbin/criu /usr/share/doc/criu /usr/share/doc/criu/COPYING /usr/share/doc/criu/README.md /usr/share/man/man8/criu.8.gz
An example of performing a checkpoint/restore is
/projects/cuda-checkpoint-main/580.65.06/src/example.sh
the counter does not spit out output so hard to test….
Actually works a little different.
To get these output files, the process follows these standard steps: 1. Compile and Run:bash nvcc counter.cu -o counter ./counter & PID=$! 2. Suspend GPU Work: You use cuda-checkpoint to freeze and extract the GPU memory layout into host memory space cuda-checkpoint --toggle --pid $PID 3. Dump to Files: You use CRIU to write the combined state out into your output directory criu dump --shell-job --images-dir demo --tree $PID Once completed, the target counter process is terminated, and your demo/ directory will be packed with the binary image (.img) files representing the exact frozen state of your application.
cluster/234.1791297222.txt.gz · Last modified: by hmeij07
