diff --git a/doc/txt/bgl.txt b/doc/txt/bgl.txt new file mode 100644 index 0000000000000000000000000000000000000000..b47c2b3651cf4a8e1d49ea1de0d3870132d45dca --- /dev/null +++ b/doc/txt/bgl.txt @@ -0,0 +1,35 @@ +Notes for SLURM use with Blue Gene/L +As of 17 August 2004 + +There is still much development work required, but some basic functionality is +in place and usable for development purposes. + + +TO BUILD +Download a current version of the SLURM code (version 0.4 or later). +Execute "configure". +Manually add "#define HAVE_BGL" to the file "config.h". This will + be fixed later with autoconf. +Execute "make". +Execute "make install". +Build a configuration file: + Set "InactiveLimit=0" (or leave to the default value of zero). + Define the base partitions using a three-digit suffix indicating its + position in the X, Y, and Z dimentions. Set their NodeAddr to the + location where the slurmd daemon will execute (the same node for + all of the base partitions). Note that SLURM itself is not aware + of the actual node location, but this will work. + + +TO RUN +Execute "slurmctld" and *one* copy of "slurmd". +The scontrol, scancel, squeue, and sinfo commands execute as desired. +The srun command works fine too, but BGL will not support proper job + steps. There will be a job allocation and mpirun will use the + allocated BGL resources by virtue of environment variables. Attempts + to launch job steps will function, but all tasks get started on the + one node where the slurmd daemon executes. +Most of the existing SLURM test suite runs, but some tests fail due to + the defined node names (e.g. "bgl123") not matching the real node + names reported by the hostname command. srun's attach option will + only work for batch jobs, which will be the normal BGL mode of operation.