User Tools

Site Tools


cluster:211

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Next revision
Previous revision
cluster:211 [2022/02/28 14:47] – created hmeij07cluster:211 [2026/10/05 19:14] (current) – hmeij07
Line 3: Line 3:
  
 ===== DMTCP CRAC ===== ===== DMTCP CRAC =====
 +
 +**Update** 05/10/2026
 +
 +Retrying this on node n91 with Rocky 8.10 and latest Blackwell GPUs and cuda-13.0.
 +
 + * Local install at /projects/dmtcp-crac/3.0.0 read the 3.0.0.install file
 +
 +Runs, but fails tests. Not sure why. Besides that the contrib/split-cuda fails to compile.
 +
 +Look at this page for an NVIDIA maintained checkpoint utility; [[cluster:234|cuda-checkpoint]]
 +
 +
  
 This is a new DMTCP(https://github.com/dmtcp/dmtcp.git) plugin to checkpoint- restart CUDA application with noval split-process architecture.  This is a new DMTCP(https://github.com/dmtcp/dmtcp.git) plugin to checkpoint- restart CUDA application with noval split-process architecture. 
Line 9: Line 21:
   * http://dmtcp.sourceforge.net/   * http://dmtcp.sourceforge.net/
  
-CRAC consists of the plugin on top of DMTCP.+CRAC consists of the plugin on top of DMTCP.\\ 
 +This software runs in the original directory
  
 Compilation needs ''gcc'' version 8 or later (using 9.2.0 on CentOS 7, compute node ''n79'') Compilation needs ''gcc'' version 8 or later (using 9.2.0 on CentOS 7, compute node ''n79'')
  
 +<code>
 +
 +# env on node n79 CRAC-early-developmennt-master.zip
 +
 +# download
 +wget https://github.com/DMTCP-CRAC/CRAC-early-development/archive/master.zip
 +
 +# unzip
 +unzip ../dmtcp-crac-master.zip
 +mv CRAC-early-development-master /share/apps/CENTOS7/dmtcp/3.0.0.b
 +
 +# gcc 
 + export PATH=/share/apps/CENTOS7/gcc/9.2.0/bin:$PATH
 + export LD_LIBRARY_PATH=/share/apps/CENTOS7/gcc/9.2.0/lib64:\
 +/share/apps/CENTOS7/gcc/9.2.0/libexec:\
 +/share/apps/CENTOS7/gcc/9.2.0/lib:\
 +$LD_LIBRARY_PATH
 +
 +
 + export PATH=/share/apps/CENTOS7/openmpi/4.0.4/bin:$PATH
 + export LD_LIBRARY_PATH=/share/apps/CENTOS7/openmpi/4.0.4/lib:$LD_LIBRARY_PATH
 + export PATH=/share/apps/CENTOS7/python/3.8.3/bin:$PATH
 + export LD_LIBRARY_PATH=/share/apps/CENTOS7/python/3.8.3/lib:$LD_LIBRARY_PATH
 +
 + export CUDA_HOME=/usr/local/cuda
 + export PATH=/usr/local/cuda/bin:$PATH
 + export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
 + export LD_LIBRARY_PATH=/usr/local/cuda/lib:$LD_LIBRARY_PATH
 +
 +
 +# make in place
 +cd /share/apps/CENTOS7/dmtcp/3.0.0.b/
 +./configure
 +make  # no errors
 +$ ls bin
 +dmtcp_command      dmtcp_discover_rm  dmtcp_nocheckpoint  dmtcp_rm_loclaunch  dmtcp_ssh   mtcp_restart
 +dmtcp_coordinator  dmtcp_launch       dmtcp_restart       dmtcp_srun_helper   dmtcp_sshd
 +
 +make check  # all failed, msg: checkpoint error ???
 +make check2 # /bin/sh: -c: line 11: syntax error near unexpected token `&'
 +make check3 # /bin/sh: -c: line 11: syntax error near unexpected token `&'
 +
 +cd contrib/split-cuda
 +# edit Makefile set to gcc/g++ in PATH
 +make # no errors, but missing lib
 +
 +$ ls             
 +libdmtcp_split-cuda.so
 +kernel-loader.exe  
 +libcuda_wrappers.so 
 +
 +# -lcuda -lcusparse -lcusolver -lcublas
 +# my 10.2 toolkit does not have cublas v11
 +# so linking against lowest version in hpc_sdk
 +
 +# seems to have worked
 +$ ldd kernel-loader.exe 
 + libcublas.so.11 => not found
 +# now
 + libcublas.so.11 => /usr/local/cuda/lib64/libcublas.so.11 (0x00007fc3b877a000)
 +
 +
 +</code>
 +
 +Next gobble together a gpu program like lammps/amber and test on gpu. Or you may have to wait on new compute nodes to arrive with latest toolkit and redo. The libcublas links needs to be made on each node in  this queue ''n79'' belongs to.
 +
 +<code>
 +
 +# oh well, nice try
 +cudaGetDeviceCount failed CUDA driver version is insufficient for CUDA runtime version
 +
 +# manual with amber20 example
 +
 +dmtcp_launch --new-coordinator \
 +--coord-port 0 --port-file /sanscratch/checkpoints/111/port.txt \
 +--ckptdir /sanscratch/checkpoints/111 --interval 600   pmemd.cuda \
 +-O -o mdout.$LSB_JOBID -inf mdinfo.1K10 -x mdcrd.1K10 -r restrt.1K10 -ref inpcrd
  
  
 +</code>
  
  
 +Put your checkpoints in /sanscratch/checkpoints$LSF_JOBPID. You need to make that directory but do not have to clean it up. Checkpoints older than 120 days are automatically removed.
  
 \\ \\
cluster/211.1646059665.txt.gz · Last modified: by hmeij07