II Slurm Overview

Slurm Overview
Felip Moll - felip.moll@schedmd.com
SchedMD LLC
www.schedmd.com
Copyright 2018 SchedMD LLC

www.schedmd.com
SLUG Sept. 25&26th, 2018
Outline
● Roles of resource manager and job scheduler

● Slurm design and architecture
● Submitting and running jobs
● Managing jobs
● Accounting

www.schedmd.com
Outline

● Managing jobs
● Accounting

www.schedmd.com
Role of a Resource Manager
● The “glue” for a parallel computer to execute parallel jobs

On a PC: On a cluster:
Execute program Execute 8 copies of
"a.out": "a.out":
a.out srun -n8 a.out
● MPI or UPC would typically be used to manage

communications within the parallel program

www.schedmd.com
Role of a Resource Manager
● Allocate resources within a cluster Requires extensive knowledge about the

hardware and system software (e.g. to alter
network routing or manage switch window)
● Launch and otherwise manage jobs

www.schedmd.com
Role of a Job Scheduler
● When there is more work than resources, the job scheduler

manages queue(s) of work:
○ Supports complex scheduling algorithms
■ Optimized for network topology, fair share scheduling, advanced

reservations, preemption, gang scheduling (time-slicing jobs), etc.
○ Supports resource limits, QoS: by queue, user, group, etc.

www.schedmd.com
Examples
Resource Managers Schedulers
ALPS (Cray) Maui
Torque Moab
LoadLeveler (IBM)
Slurm
LSF
PBS Pro
Many span both roles

Slurm started as a resource manager (the “rm”
in “Slurm”) and added scheduling logic later
www.schedmd.com
Outline

● Managing jobs
● Accounting

www.schedmd.com
What is Slurm?
● Historically, Slurm was an acronym of:

– Simple Linux Utility for Resource Management
● Dev. started in 2002 @ Lawrence Livermore National Lab as a
resource manager for Linux clusters
● Sophisticated scheduling plugins added in 2008
● About 550,000 lines of C code today
● Supports Linux and limited support for other Unix variants
● Used on many of the world's largest computers
● Active global user community
www.schedmd.com
Who is using Slurm?

www.schedmd.com
Slurm Design Goals
● Highly scalable
– Managing 3.1 million core Tianhe-2
– Tested to much larger systems using emulation
● Open source GPLv2, available on Github: https://github.com/SchedMD/
● System administrator friendly
● Secure
● Fault-tolerant (no single point of failure)
● Portable - Targeting POSIX2008.1 and C99
www.schedmd.com
Slurm Portability
● Autoconf configuration engine adapts to environment
● Provides scheduling framework with general-purpose plugin

mechanism. Extensively customize using a building-block approach.
● Various system-specific plugins available (e.g. select/cray)
● Huge range of use cases:

○ Sophisticated workload management at HPC sites
○ Scalable HTC environments (14k jobs/minute sustained)
www.schedmd.com
Plugins
● Dynamically linked objects loaded at run time based upon

configuration file and/or user options
● 110+ plugins of 30 different varieties currently available
○ Network topology: 3D-torus, tree, etc
○ MPI: OpenMPI, PMI2, PMIx
○ Process tracking: cgroup, linuxproc, pgid, ipmi, etc.
Slurm Kernel (65% of code)

Authentication MPI Job Submit Topology Accounting Storage
Plugin Plugin Plugin Plugin Plugin
MUNGE PMI2 Lua Tree MySQL
www.schedmd.com
Plugin Design
● Plugins typically loaded when the daemon or command starts

and persist indefinitely
● Provide a level of indirection to a configurable underlying
function
slurmctld slurmdbd
Write job completion Write job completion
accounting record accounting record
accounting_storage/slurmdbd accounting_storage/mysql
Send to SlurmDBD Write to MySQL database

http://www.schedmd.com
Slurm Entities
● Jobs: Resource allocation requests

● Job steps: Set of (typically parallel) tasks
○ Typically an MPI, UPC and/or multi-threaded application program
○ Allocated resources from the job's allocation
○ A job can contain multiple job steps which can execute sequentially or
concurrently
○ Use cases with thousands of job steps are common
○ Lighter weight than jobs
● Partitions: Job queues with limits and access controls
● QoS: Limits and polices Copyright 2018 SchedMD LLC
www.schedmd.com
Slurm Entities Example
● Users submit jobs to partitions (queue):
Priority ordered queue of jobs
Partition debug
Job 1
Job 2
Job 3

www.schedmd.com
● Jobs are allocated resources:
Node: tux123
#!/bin/bash Socket: 0
#SBATCH -n6
.. Core 0
Partition debug Core 1
Core 2
Job 1
Core 3
Job 2
Core 4
Job 3
Core 5

www.schedmd.com
● Jobs spawn steps, which are allocated resources from within

the job's allocation: Node:
#!/bin/bash tux123
srun -n4 --exclusive a.out & Socket: 0
srun -n2 --exclusive a.out & Core 0
wait
Partition debug Core 1
Step 0 Core 2
Job 1
Core 3
Job 2
Step 1 Core 4
Job 3
Core 5

www.schedmd.com
Node State Information
● Baseboards, Sockets, Cores, Threads

● CPUs (Core or thread count depending upon configuration)
● Memory size
● Generic resources (with names and counts)
● Features (arbitrary string, e.g. OS version or CPU type)
● State (e.g. drain, down, etc.)
○ Reason, time and user ID
■ e.g. “Bad PDU [operator@12:40:10T12/20/2013]”
www.schedmd.com
Queue/Partition State Information
● Associated with specific set of nodes

○ Nodes can be in more than one partition
● Job size and time limits (e.g. small size and time limits for some
partition and larger limits for others)
● Access control list (by account, Quality Of Service, or Linux group)
● Preemption rules
● State information (e.g. up, down, drain, etc.)
● Over-subscription and gang scheduling rules

www.schedmd.com
Job State Information
● ID (a number)
● Name
● Time limit (minimum and/or maximum)
● Size specification (minimum and/or maximum; nodes, CPUs, sockets, cores, and/or
threads)
● Specific node names to include or exclude in allocation
● Node features required in allocation
● Dependency
● Account name
● Quality Of Service (QOS)
● State (Pending, Running, Suspended, Canceled, Failed, etc.)

www.schedmd.com
Job States
Submission Pending
Configuring
(node booting)
Resizing Running Suspended
Completing
Canceled Completed Failed TimeOut NodeFail

(zero exit code) (non-zero exit code) (time limit reached)
Preempted SpecialExit
www.schedmd.com
Step State Information
● ID (a number): <jobid>.<stepid>
● Name
● Time limit (maximum)
● Size specification (minimum and/or maximum; nodes, CPUs,
sockets, cores, and/or threads)
● Specific node names to include or exclude in allocation
● Node features required in allocation
www.schedmd.com
Summary
● Job is submitted to a Slurm queue/partition
● Job is allocated resources (cores, memory, etc.)
● Job steps execute applications using the job's resources
● Job and associated resources are monitored:

– By plugins: JobAcctGather, Cgroup, NHC, etc.
www.schedmd.com
Daemons
● slurmctld – Central controller (typically one per cluster)

○ Monitors state of resources
○ Manages job queues
○ Allocates resources
● slurmdbd – Database daemon (typically one per enterprise)

○ Collects accounting information
○ Manages accounting configuration (e.g. limits, fair-share, etc.)
■ Pushes to controller(s)

www.schedmd.com
Daemons
● slurmd – Compute node daemon

○ Typically one per compute node
○ Launches and manages slurmstepd (see below)
○ Small and very light-weight
○ Quiescent after launch except for optional accounting
○ Supports hierarchical communications with configurable fanout
● slurmstepd – Job step shepherd

○ Launched for batch job and each job step
○ Launches user application tasks
○ Manages accounting, application I/O, signals, etc.

www.schedmd.com
Cluster Architecture - Typical Linux Cluster
Slurmctld Slurmdbd
(backup) (backup)
Slurm
Slurmctld Slurmdbd
user MySQL
(master) (master)
tools
Accounting and configuration

records
Slurmd daemons on compute nodes
(Note hierarchical communications with configurable fanout)
www.schedmd.com
Typical Enterprise Architecture
Slurm administration
tools
Slurm Accounting data

(cluster 1)
Slurm Slurmdbd
Jobs &
user status
tools Slurm
User and account MySQL
(cluster N)
limits and
preferences
User account and limit info

Job and step accounting info

www.schedmd.com
Configuration
●
General conf
●
Plugin activation ●
Slurmdbd params.
●
Scheduler ●
Archive/Purge
parameters
●
Storage options
●
Node definition
●
Partition definition

www.schedmd.com
Configuration
●
Other files configure specific plugin options
– burst_buffer.conf, acct_gather.conf, knl.conf...

www.schedmd.com
Outline

● Managing jobs
● Accounting

www.schedmd.com
Commands: General Information
● Man pages available for all commands, daemons and

configuration files
● --help option prints brief description of all options
● --usage option prints a list of the options
● Commands can be run on any node in the cluster
● Any failure results in a non-zero exit code
● APIs make new tool development easy
○ Man pages available for all APIs
www.schedmd.com
Commands: General Information
● Almost all options have two formats

○ A single letter option (e.g. “-p debug” for partition “debug”)
○ A verbose option (e.g. “--partition=debug”)
● Almost all commands support verbose logging with “-v” option
○ Use more v’s for more verbosity, e.g. -vvvv
● Many environment variables can be used to establish site-
specific and/or user-specific defaults
○ For example “SQUEUE_STATES=all” for the squeue command to
display jobs in any state, including COMPLETED or CANCELLED
www.schedmd.com
Slurm Commands: Job/step Allocation
● sbatch – Submit script for later execution (batch mode)
● salloc – Create job allocation and start a shell to use it

(interactive mode)
● srun – Create a job allocation (if needed) and launch a job

step (typically an MPI job)
● sattach – Connect stdin/out/err for an existing job step

www.schedmd.com
Job allocation options
● The job allocation commands (salloc, sbatch, and srun)

accept almost identical options
● There are a handful of options that only apply to a subset of

these commands (e.g. batch job requeue options, cpu binding)

www.schedmd.com
sbatch Example - Options in Batch Script
Submit a batch job

> sbatch my_work2.bash
Submitted batch job 44005
> cat my_work2.bash
#!/bin/bash
#SLURM --ntasks=128
#SLURM --cpus-per-task=2
#SLURM --mem-per-cpu=20
#SLURM --time=60
…

www.schedmd.com
sbatch Example - Job Dependencies
Submit sequence of three batch jobs

> sbatch --ntasks=1 --time=10 pre_process.bash
> sbatch --ntasks=128 --time=60 --dependency=afterok:45001 do_work.bash

> sbatch --ntasks=1 --time=30 --dependency=afterok:45002 post_process.bash

> sbatch --begin=17:00:00 -n1 -t10 night.bash

Options used
--ntasks or -n Number of tasks and by default the number of CPUs
--time or -t Wall time limit (minutes in our example)
--dependency or -d Job dependency
--begin Delay job initiation at least
Copyright 2018 until LLC
SchedMD this date and time
www.schedmd.com
srun Example - Multiple Serial Job Steps
> salloc -n128 bash

Granted job allocation 10090
> srun step.1
> srun step.2
> srun step.3
> exit

www.schedmd.com
Outline
● Managing jobs
● Accounting

www.schedmd.com
Job Management Commands
● scancel – Signal / terminate job and steps
● squeue – Report job and job step status

www.schedmd.com
Why is my job not running?
● As soon as some reason is found why a job cannot be started, that is recorded in
the job's “reason” field and the scheduler moves on to the next job
Some common reasons why jobs are pending:
Priority Resources being reserved for higher priority job
Resources Required resources are in use
Dependency Job dependencies not yet satisfied
Reservation Waiting for advanced reservation
AssociationJobLimit User or account job limit reached

AssociationResourceLimit User or account resource limit reached
AssociationTimeLimit User or account time limit reached
QOSJobLimit Quality Of Service (QOS) job limit reached

QOSResourceLimit Quality Of Service (QOS) resource limit reached
QOSTimeLimit Quality Of Service (QOS) time limit reached
www.schedmd.com
Slurm System Information
● sinfo – Report system status (nodes, queues, etc.)

● sview – Report and/or update system, job, step, partition or
reservation status with topology (GTK-based GUI)
● smap – Report system, job or step status with topology (curses-
based GUI), less functionality than sview
● scontrol – Administrator tool to view and/or update system,
job, step, partition or reservation status

www.schedmd.com
sview Functionality
● Get details about specific jobs, steps, nodes, etc.

● Can be used for job control functions
○ Cancel job
○ Hold/release job, etc.
○ Support for job submission with relatively few options
● Can be used for administrative functions
○ Change configuration, etc.

www.schedmd.com
sview on Cray (3D torus)

www.schedmd.com
smap on Cray (3D torus)

www.schedmd.com
scontrol Command
● Designed primarily for system administrator use

● Shows all available fields, but no filtering, sorting or formatting options
● Many fields can be modified when pending or running
> scontrol show job 123
JobId=123 Name=tmp
UserId=jette(1000) GroupId=jette(1000)
Priority=4294901737 Nice=0 Account=(null) QOS=(null)
JobState=RUNNING Reason=None Dependency=(null)
RunTime=00:00:11 TimeLimit=00:30:00 TimeMin=N/A
> scontrol update JobId=123 TimeLimit=20

www.schedmd.com
Outline
● Managing jobs
● Accounting

www.schedmd.com
Database Use
● Accounting information written to a database plus

○ Information pushed out live to scheduler daemons
○ Quality of Service (QOS) definitions
○ Fair-share resource allocations
○ Many limits (max job count, max job size, etc)
○ Based upon hierarchical accounts
■ Limits by user AND by accounts

www.schedmd.com
Hierarchical Accounts Example
Root 100%
Division A 33.3% Division B 33.3% Division C 33.3%
Group Alpha 50% Group Beta 30% Group Gamma 20%
Ted Pat Bob

Pam
30% 25% 25%
20%
www.schedmd.com
Hierarchical Accounts
● All users are not created equal

○ Different shares of resources
○ Different measures of being over- or under-served
○ Different limits
● There are many limits available
○ Per job limits (e.g. MaxNodes)
○ Aggregate limits by user, accounts, or QOS (e.g. MaxJobs)
○ A single user may have different shares and limits in different accounts,
QOS, or partitions
www.schedmd.com
Accounting Commands
● sshare – Reports account hierarchy

● sprio – Report components of job priority
● sacctmgr – Get and set account data and limits
● sstat – Report accounting information of running job by
individual job and job step
● sacct – Report accounting information by individual job and
job step
● sreport – Report resources usage by cluster, partition, user,
account, etc.
www.schedmd.com
Summary
● You should now have a basic overview of Slurm concepts
● Documentation available at https://slurm.schedmd.com
● Questions?

www.schedmd.com

II Slurm Overview

Uploaded by

Copyright:

Available Formats

II Slurm Overview

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

II Slurm Overview

Uploaded by

Copyright:

Available Formats

Slurm Overview

Felip Moll - felip.moll@schedmd.com

Copyright 2018 SchedMD LLC

● Roles of resource manager and job scheduler

Copyright 2018 SchedMD LLC

● Roles of resource manager and job scheduler

Copyright 2018 SchedMD LLC

● The “glue” for a parallel computer to execute parallel jobs

● MPI or UPC would typically be used to manage

Copyright 2018 SchedMD LLC

● Allocate resources within a cluster Requires extensive knowledge about the

● Launch and otherwise manage jobs

● When there is more work than resources, the job scheduler

○ Supports complex scheduling algorithms

■ Optimized for network topology, fair share scheduling, advanced

○ Supports resource limits, QoS: by queue, user, group, etc.

Copyright 2018 SchedMD LLC

Many span both roles

● Roles of resource manager and job scheduler

Copyright 2018 SchedMD LLC

● Historically, Slurm was an acronym of:

Copyright 2018 SchedMD LLC

● Autoconf configuration engine adapts to environment

● Provides scheduling framework with general-purpose plugin

● Various system-specific plugins available (e.g. select/cray)

● Huge range of use cases:

● Dynamically linked objects loaded at run time based upon

Slurm Kernel (65% of code)

● Plugins typically loaded when the daemon or command starts

Send to SlurmDBD Write to MySQL database

● Jobs: Resource allocation requests

● Users submit jobs to partitions (queue):

Priority ordered queue of jobs

Copyright 2018 SchedMD LLC

● Jobs are allocated resources:

Copyright 2018 SchedMD LLC

● Jobs spawn steps, which are allocated resources from within

Copyright 2018 SchedMD LLC

● Baseboards, Sockets, Cores, Threads

● Associated with specific set of nodes

Copyright 2018 SchedMD LLC

Copyright 2018 SchedMD LLC

Resizing Running Suspended

Canceled Completed Failed TimeOut NodeFail

● Job is submitted to a Slurm queue/partition

● Job is allocated resources (cores, memory, etc.)

● Job steps execute applications using the job's resources

● Job and associated resources are monitored:

● slurmctld – Central controller (typically one per cluster)

● slurmdbd – Database daemon (typically one per enterprise)

Copyright 2018 SchedMD LLC

● slurmd – Compute node daemon

● slurmstepd – Job step shepherd

Copyright 2018 SchedMD LLC

Accounting and configuration

Slurm Accounting data

User account and limit info

Copyright 2018 SchedMD LLC

Copyright 2018 SchedMD LLC

Copyright 2018 SchedMD LLC