IBM RS/6000 SP2 ¼öÆÛÄÄÇ»ÅÍ |
---|
2002/03/05(23:27) from 218.149.65.204 | |
ÀÛ¼ºÀÚ : °ÁÙ±â (jkkang65@hanmail.net) | Á¶È¸¼ö : 1927 , ÁÙ¼ö : 66 |
Re: [LoadLeveler] Job Command File ¼³¸í |
---|
LoadLeveler The shared memory nature of the P3 processors on Tempest require new LoadLeveler commands to enable the user to request specific CPU and node combinations. Those commands are listed below. In addition, because each node on Tempest contains sixteen processors which share 8 GB of memory, MHPCC recommends that users submit Tempest jobs in sixteen-processor increments. New commands in the LoadLeveler command files are described in the following paragraphs. tasks_per_node This is number of tasks per node. The maximum number of tasks per node is sixteen. Use this command in conjunction with the node command (see below). You cannot specify BOTH tasks_per_node and total_tasks in the same LoadLeveler command file. SYNTAX: #@ tasks_per_node = 16 network.MPI This command replaces some of the #@ environment line variables in previous LoadLeveler command files. This command will set the communications library to either IP or US, the communications adapter to either the SP switch or ethernet adapter, and whether or not to share the adapter. Please set this line as listed below. For batch jobs SYNTAX: #@ network.MPI = css0,shared,US node_usage This command file variable allows sharing of the node. This variable command MUST be set to the same option that the network.MPI variable is set. Both MUST be shared OR not_shared. You cannot MIX the two or your job will NOT run as expected. SYNTAX: #@ node_usage = shared node This command specifes the number of nodes you wish use. This command replaces the min_processor and max_processor commands. SYNTAX: #@ node = 24 total_tasks This command may be used when you wish to run with less than four tasks per node. For example, to run two tasks on each of 12 available nodes for a total of 24 tasks, specify the following: SYNTAX: #@ total_tasks = 24 #@ node = 12 comment QOS This command may be used by HPCMP users to run in background mode. Hours accumulated by jobs run in background mode are not counted against users' allocations. SYNTAX: #@ comment = QOS=0 environment MP_SHARED_MEMORY=yes This environment variable will ensure communications between processors on a node is via shared memory rather than via message passing over the switch adapter. This should be included in all .cmd files. Example Command File For A 64 CPU Job On 4 Nodes #!/bin/csh #@ job_name = my_job #@ initialdir = /u/my_dir/program_resides #@ output = $(job_name).out #@ error = $(job_name).err #@ job_type = parallel #@ notification = complete #@ notify_user = my_email_address #@ network.MPI = css0,shared,US #@ node_usage = shared #@ node = 4 #@ tasks_per_node = 16 #@ wall_clock_limit = 40000 #@ environment = MP_INFOLEVEL=2;MP_LABELIO=yes;MP_SAVEHOSTFILE=hosts;MP_SHARED_MEMORY=yes; #@ account_no = ABCDE-1234-F56 #@ queue # If procs are greater that 128....delete (comment out) next line echo $LOADL_PROCESSOR_LIST > my_prog.hosts my_prog_executable |