Slurm User and Administrator Guide for Cray Systems Natively
This document describes the unique features of Slurm on Cray XC computers natively, or without the use of Cray's Application Level Placement Scheduler (ALPS). You should be familiar with the Slurm's mode of operation on Linux clusters before studying the differences in Cray system operation described in this document. When running Slurm in native mode a Cray system will function very similar to a Linux cluster.
Slurm is designed to operate as a workload manager on Cray XC systems (Cascade) without the use of ALPS. In addition to providing the same look and feel of a regular Linux cluster this also allows for many functionalities such as:
- Ability to run multiple jobs per node
- Ability to status running jobs with sstat
- Full accounting support for job steps
- Ability to run multiple jobs/steps in background from the same session
- Network Performance Counters
- Core Specialization
To access Cray's Network Performance Counters (NPC) you can use the --network option in sbatch/salloc/srun to request them. There are 2 different types of counters, system and blade.
For the system option (--network=system) only one job can use system at a time. Only nodes requested will be marked in use for the job allocation. If the job does not fill up the entire system the rest of the nodes are not able to be used by other jobs using NPC, if idle their state will appear as PerfCnts. These nodes are still available for other jobs not using NPC.
For the blade option (--network=blade) Only nodes requested will be marked in use for the job allocation. If the job does not fill up the entire blade(s) allocated to the job those blade(s) are not able to be used by other jobs using NPC, if idle their state will appear as PerfCnts. These nodes are still available for other jobs not using NPC.
To use set CoreSpecPlugin=core_spec/cray_aries. Ability to reserve a number of cores allocated to the job for system operations and not used by the application. The application will not use these cores, but will be charged for their allocation.
Many new plugins were added to utilize the Cray system without ALPS. These should be set up in your slurm.conf outside of your normal configuration.
Set BurstBufferPlugins=burst_buffer/datawarp to use. The burst buffer capability on Cray systems is also known by the name DataWarp. For more information, see Slurm Burst Buffer Guide.
To use set CoreSpecPlugin=core_spec/cray_aries.
Set JobSubmitPlugins=job_submit/cray_aries to use. This plugin is primarily used to set a gres=craynetwork value which is used to limit the number of applications that can run on a node at once. For a node without MICs on it that number at most is 4. Nodes with MICs the number drops to 2. This craynetwork gres needs to be set up in your slurm.conf to ensure proper functionality. In example...
... Grestypes=craynetwork NodeName=nid000[00-10] gres=craynetwork:4 #node without MIC NodeName=nid000[11-20] gres=craynetwork:2 #node with MIC ...
Set PowerPlugin=power/cray_aries to use. PowerParameters is also typically configured. For more information, see Slurm Power Management Guide.
Set ProctrackType=proctrack/cray_aries to use.
Set SelectType=select/cray_aries to use. This plugin is a layered plugin. Which means it enhances a lower layer select plugin. By default it is layered on top of the select/linear plugin. It can also be layered on top of the select/cons_res plugin by using the SelectTypeParameters=other_cons_res, doing this will allow you to run multiple jobs on a Cray node just like on a normal Linux cluster. Use additional SelectTypeParameters to identify the resources to allocate (e.g. cores, sockets, memory, etc.). See the slurm.conf man page for details.
Realm-Specific IP Addressing (RSIP) will automatically try to interact with anything opened on ports 8192 to 60000. Configure SlurmctldPort, SlurmdPort, and SrunPortRange to use ports above 60000. In the case of SrunPortRange, making 1000 or more ports available is recommended.
Set SwitchType=switch/cray_aries to use.
Set TaskPlugin=cray_aries,cgroup to use. Use of the task/cgroup plugin is required alongside task/cray_aries. You may also use the task/affinity plugin along with task/cray_aries,task/cgroup if desired (i.e. TaskPlugin=cray_aries,affinity,cgroup). Note that plugins are used in the order they are defined in the comma separated list, and that task/cray_aries must be listed before task/cgroup due to internal dependencies between the two plugins.
Some Slurm plugins (burst_buffer/datawarp and power/cray_aries) plugins parse JSON format data. These plugins are designed to make use of the JSON-C library for this purpose. See JSON-C installation instructions for details.
Some services on the system need to be set up to run correctly with Slurm. Below is how to restart the service and the nodes they run on. It is probably a good idea to set this up to happen automatically.
- boot node
- WLM_DETECT_ACTIVE=SLURM /etc/init.d/aeld restart
- sdb node
- WLM_DETECT_ACTIVE=SLURM /etc/init.d/ncmd restart
- WLM_DETECT_ACTIVE=SLURM /etc/init.d/apptermd restart
As with Linux clusters you will need to start a slurmd on each of your compute nodes. If you choose to use munge authentication, advised, you will also need munge installed and a munged running on each of your compute nodes as well. See the quick start guide for more info. Outside of the differences listed in this file it can be used to set up your Cray system to run Slurm natively.
On larger systems, you may wish to set the PMI_MMAP_SYNC_WAIT_TIME environment variable in your users' profiles to a larger value than the default (180 seconds) to prevent PMI from falsely detecting job launch failures.
A backup controller can be setup in or outside the Cray. However, when the backup is within the Cray, both the primary and the backup controllers will go down when the Cray is rebooted. It is best to setup the backup controller on a Cray external node so that the controller can still receive new jobs when the Cray is down. When the backup is configured on an external node the no_backup_scheduling SchedulerParameter should be specified in the slurm.conf. This allows new jobs to be submitted while the Cray is down and prevents any new jobs from being started.
Last modified 7 March 2019