Optimizing resource management in supercomputers with SLURM
Dig into this introduction to the Simple Linux Utility for Resource Management
Supercomputers are a classic example of an arms race. While the growing performance of modern supercomputers expands into new problem domains, these massive systems are providing platforms for solving new problems. Supercomputers are a source of national and corporate pride, as companies and nations work to improve LINPACK scores. Figure 1 illustrates the past five years of the supercomputing arms race, with the IBM Sequoia supercomputer the current projected leader in 2012. As shown, IBM Roadrunner was the first supercomputer to break the sustained petaflops barrier (and IBM Blue Gene®/L held the top spot from 2004 through 2008).
Figure 1. Supercomputer performance: 2008-2012
Early supercomputers were designed to model nuclear weapons. Today, their application is much more diverse, tackling massive computational problems in the fields of climate research, molecular modeling, large-scale physical simulations, and even brute force code breaking.
1964 to today
The first supercomputer is generally considered the Control Data Corporation (CDC) 6600, released in 1964 (designed by Seymour Cray). The 6600 filled four cabinets with hardware, a Freon cooling system, and a single CPU capable of 3 million floating-point operations per second (FLOPS). Although not lacking in aesthetics, its cabinets were visibly filled with colored wires tying its peripheral unit processors to the single CPU to keep it as busy as possible.
Fast-forward to today, and the current supercomputer leader is the Japanese Kei computer (built by Fujitsu). This system focuses on brute strength compute capacity, using more than 88,000 SPARC64 processors spread across 864 cabinets. The Kei supercomputer has the distinction of breaking the 10-petaflop barrier. Similar to the CDC 6600, the Kei uses water cooling in addition to air cooling.
What is a supercomputer?
A supercomputer is not about any particular architecture but simply a design that is on the bleeding edge of computational performance. Today, this means a system that operates in the performance range of petaflops (or quadrillions of FLOPS) as measured by the LINPACK benchmark.
Regardless of how the supercomputer achieves those FLOPS, a low-level goal of any supercomputer architecture is to optimally keep the compute resources busy when there's work to do. Similar to the CDC 6600 peripheral processors, which existed to keep its single CPU busy, modern supercomputers need the same fundamental capability. Let's look at one such implementation of compute node resource management, called the Simple Linux® Utility for Resource Management (SLURM).
SLURM in a nutshell
SLURM is a highly scalable and fault-tolerant cluster manager and job scheduling system for large clusters of compute nodes. SLURM maintains a queue of pending work and manages the overall utilization of resources by this work. It also manages the available compute nodes in an exclusive or nonexclusive fashion (as a function of the resource need). Finally, SLURM distributes jobs to a set of allocated nodes to perform the work in addition to monitoring the parallel jobs to their completion.
Under the covers, SLURM is a robust cluster manager (focusing on need over feature richness) that is highly portable, scalable to large clusters of nodes, fault tolerant, and most importantly, open source. SLURM began as open source resource manager, developed by several companies (including the Lawrence Livermore National Laboratory) in a collaborative effort. Today, SLURM is a leading resource manager on many of the most powerful supercomputers.
SLURM implements a fairly traditional cluster management architecture (see
Figure 2). At the top is a redundant pair of
cluster controllers (though redundancy is optional). These cluster
controllers serve as the managers of the compute cluster and implement a
management daemon called
slurmctld daemon provides monitoring of compute
resources, but most importantly, it maps incoming jobs (work) to
underlying compute resources.
Each compute node implements a daemon called
daemon manages the node on which it executes, including monitoring the
tasks that are running on the node, accepting work from the controller,
and mapping that work to tasks on the cores within the node. The
slurmd daemon also can stop tasks for
executing, if requested from the controller.
Figure 2. High-level view of the SLURM architecture
Other daemons exist within the architecture—for example, to implement secure authentication. But a cluster is more than just a random collection of nodes, as some of these nodes can be logically related at points in time for parallel computation.
A set of nodes can be collected into a logical group called a partition, which commonly includes a queue of incoming work. Partitions can be configured with constraints about which users can use it or the job size of time limit the partition supports. A further refinement of a partition is the mapping of a set of nodes within a partition to a user for a period of time for work, which is called a job. Within a job is one or more job steps, which are sets of tasks executing on that subset of nodes.
Figure 3 shows this hierarchy, which further illustrates the SLURM partitioning of resources. Note that this partitioning includes awareness of resource proximity to ensure low-latency communication among cooperating nodes.
Figure 3. Resource partitioning in SLURM
How you install SLURM ultimately depends on your particular Linux environment, but the process can be as simple as using a package manager. SLURM has been fully packaged, making it easy to install and configure. For my favorite distro, Ubuntu, I use the Advanced Packaging Tool (APT) to install the SLURM package and all of its dependencies:
$ sudo apt-get install slurm-llnl
This operation consumes less than 40MB of space and includes not just SLURM but each dependency, basic plug-ins, and other necessary packages.
Before starting SLURM, you must configure it for your particular environment. To create my configuration file, I used the online SLURM configurator, which generates the configuration file for me based on form data. Note that the file needs to be massaged in the end to remove options that were no longer supported. Listing 1 shows my resulting configuration file (stored at /etc/slurm-llnl/slurm.conf).
Listing 1. SLURM configuration file for a single-node cluster
# slurm.conf file generated by configurator.html. # Put this file on all nodes of your cluster. # See the slurm.conf man page for more information. # ControlMachine=mtj-VirtualBox # AuthType=auth/none CacheGroups=0 CryptoType=crypto/openssl MpiDefault=none ProctrackType=proctrack/pgid ReturnToService=1 SlurmctldPidFile=/var/run/slurmctld.pid SlurmctldPort=6817 SlurmdPidFile=/var/run/slurmd.pid SlurmdPort=6818 SlurmdSpoolDir=/tmp/slurmd SlurmUser=slurm StateSaveLocation=/tmp SwitchType=switch/none TaskPlugin=task/none # # TIMERS InactiveLimit=0 KillWait=30 MinJobAge=300 SlurmctldTimeout=120 SlurmdTimeout=300 Waittime=0 # # SCHEDULING FastSchedule=1 SchedulerType=sched/backfill SchedulerPort=7321 SelectType=select/linear # # LOGGING AND ACCOUNTING AccountingStorageType=accounting_storage/none ClusterName=cluster JobCompType=jobcomp/none JobCredentialPrivateKey = /usr/local/etc/slurm.key JobCredentialPublicCertificate = /usr/local/etc/slurm.cert JobAcctGatherFrequency=30 JobAcctGatherType=jobacct_gather/none SlurmctldDebug=3 SlurmdDebug=3 # # COMPUTE NODES NodeName=mtj-VirtualBox State=UNKNOWN PartitionName=debug Nodes=mtj-VirtualBox Default=YES MaxTime=INFINITE State=UP
Note that in a real cluster,
refer to a range of nodes, such as
snode[0-8191], to indicate 8192 unique nodes
snode8191) in the cluster.
The final step is to create a set of job credential keys for my site. I
chose to use
openssl for my credential keys
(which are referenced in the configuration file in Listing 1 as
JobCredential*. I simply use
openssl to generate these credentials, as
in Listing 2.
Listing 2. Creating credentials for SLURM
$ sudo openssl genrsa -out /usr/local/etc/slurm.key 1024 Generating RSA private key, 1024 bit long modulus .................++++++ ............................................++++++ e is 65537 (0x10001) $ sudo openssl rsa -in /usr/local/etc/slurm.key -pubout -out /usr/local/etc/slurm.cert writing RSA key
With these steps complete, I have everything I need to tell SLURM about my configuration. I can now start SLURM and interact with it.
To start SLURM, you simply use the management script defined in
/etc/init.d/slurm. This script accepts
startclean (to ignore all previously saved
states). Starting SLURM with this method causes the
slurmctld daemon to begin (as well as the
slurmd daemon on your node, in this simple
$ sudo /etc/init.d/slurm-llnl start
To validate that SLURM is now running, use the
sinfo command. The
sinfo command returns information about the
SLURM nodes and partitions (in this case, a single node makes up your
cluster), as in Listing 3.
Listing 3. Using the sinfo command to view the cluster
$ sinfo PARTITION AVAIL TIMELIMIT NODES STATE NODELIST debug* up infinite 1 idle mtj-VirtualBox $
More SLURM commands
You can get more information about your SLURM cluster from the variety of
commands available in SLURM. In Starting SLURM, you saw the
sinfo command used to learn about your cluster.
You can get more information with the
command, which allows you to view detailed information about the various
aspects of the cluster (as in Listing 4, the
partition and the node).
Listing 4. Getting more detailed information on a cluster with scontrol
$ scontrol show partition PartitionName=debug AllocNodes=ALL AllowGroups=ALL Default=YES DefaultTime=NONE DisableRootJobs=NO Hidden=NO MaxNodes=UNLIMITED MaxTime=UNLIMITED MinNodes=1 Nodes=mtj-VirtualBox Priority=1 RootOnly=NO Shared=NO PreemptMode=OFF State=UP TotalCPUs=1 TotalNodes=1 $ scontrol show node mtj-VirtualBox NodeName=mtj-VirtualBox Arch=i686 CoresPerSocket=1 CPUAlloc=0 CPUErr=0 CPUTot=1 Features=(null) Gres=(null) OS=Linux RealMemory=1 Sockets=1 State=IDLE ThreadsPerCore=1 TmpDisk=0 Weight=1 BootTime=2012-03-07T14:59:01 SlurmdStartTime=2012-04-17T11:10:43 Reason=(null)
To test your simple SLURM cluster, you can use the
srun command. The
srun command allocates a compute resource and
launches a task for your job. Note that you can do this separately (through
as well. As shown in Listing 5, you submit a simple
shell command as your job to demonstrate
and then submit a
sleep command (with an
argument) and demonstrate the use of the
command to show the jobs that exist in the cluster.
Listing 5. Submitting jobs to the cluster and checking queue status
$ srun -l hostname 0: mtj-VirtualBox $ srun -l sleep 5 &  24127 $ squeue JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 15 debug sleep mtj R 0:03 1 mtj-VirtualBox $ + Done srun -l sleep 5 $
Note in Listing 5 that the job you submit to the cluster can be a simple Linux command, a shell script file, or a proper executable.
As a final example, look at how to stop a job. In this case, you start a
longer-running job and use
squeue to identify
its ID. Then, you use the job ID with the
scancel command to terminate that job step (see
Listing 6. Terminating a job step
$ srun -l sleep 60 &  24262 $ squeue JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 16 debug sleep mtj R 0:03 1 mtj-VirtualBox $ scancel 16 srun: Force Terminated job 16 $ srun: Job step aborted: Waiting up to 2 seconds for job step to finish. 0: slurmd[mtj-VirtualBox]: error: *** STEP 16.0 CANCELLED AT 2012-04-17T12:08:08 *** srun: error: mtj-VirtualBox: task 0: Terminated + Exit 15 srun -l sleep 60 $
Finally, you can stop your cluster using the same
slurm-llnl script, as in Listing 7.
Listing 7. Stopping the SLURM cluster
$ sudo /etc/init.d/slurm-llnl stop * Stopping slurm central management daemon slurmctld [ OK ] * Stopping slurm compute node daemon slurmd [ OK ] slurmd is stopped $
Unlike Apache Hadoop, SLURM has no concept of a distributed file system. As
such, it requires a bit more processing to distribute data to the nodes
for a given computation. SLURM includes a command called
sbcast, which is used to transfer a file to all
nodes allocated for a SLURM job. It is possible—and potentially
more efficient—to use a parallel or distributed file system across
the nodes of a SLURM cluster, thus not requiring
sbcast to distribute data to process.
In this simple demonstration of SLURM, you used a subset of the available
commands and an even smaller subset of the available options to these
commands (for example, see the options available in the
srun command). Even with the minimal number of
commands available, SLURM implements a capable and efficient cluster
SLURM isn't a static resource manager but rather a highly dynamic one that can incorporate new behaviors. SLURM implements a plug-in application programming interface (API) that permits run time libraries to be loaded dynamically at run time. This API has been used to develop a variety of new behaviors, including an interconnect fabric, authentication, and scheduling. Plug-in interfaces support a variety of other capabilities, such as job accounting, cryptographic functions, Message Passing Interface (MPI), process tracking, and resource selection. All this allows SLURM to easily support different cluster architectures and implementations. Check out the SLURM Programmer's Guide in Related topics for details.
What's ahead for SLURM
In 2011, SLURM was updated with a variety of new features, including partial support for the IBM Blue Gene/Q supercomputer and the Cray XT and XE computers. Support for Linux control groups (cgroups) was also added, which provides greater control over Linux process containers.
In 2012, Blue Gene/Q support will be fully implemented, along with improved resource selection as a function of job need and resource capabilities (for example, node feature AMD). A new tool is planned to report scheduling statistics, and in the near future, a web-based administration tool. Another future plan for SLURM is in the context of cloud bursting, which involves allocating resources in a cloud provider and migrating overflow work from a local cluster to the cloud (also running SLURM daemons). This model can be quite useful and supports the idea of elasticity within certain supercomputer workloads.
Finally, the SLURM developers are considering the use of power and thermal data to more effectively distribute work in a cluster—for example, placing jobs that will consume high power (and thus generate more heat) in areas of the cluster that are more likely to shed heat.
This short introduction to SLURM illustrates the simplicity of this open source resource manager. Although modern supercomputers are beyond the price range of most people, SLURM provides the basis for a scalable cluster manager that can turn commodity servers into a high-performing cluster. Further, SLURM's architecture makes it easy to tailor the resource manager for any supercomputer (or commodity cluster) architecture. This is likely why it's the leading cluster manager in the supercomputer space.
- SLURM: A Highly Scalable Resource Manager has a useful website administered at the Lawrence Livermore National Laboratory. This site provides an overview of SLURM, a great list of publications and presentations, documentation, and a useful FAQ. You can also find professional support by the developers of SLURM at SchedMD.
- A Brief History of Supercomputing from Gordon Bell of Microsoft presents an interesting (though slightly dated) history of supercomputing, from the CDC 6600 to Beowulf clusters.
- The LINPACK Benchmark: Past, Present, and Future (Jack Dongarra, Piotr Luszczek, and Anttoine Petitet) provides a useful introduction to LINPACK history and internals. It also demonstrates LINPACK over a set of common hardware (AMD Athlon, Intel® Pentium®, and so on).
- SLURM Elastic Computing implements the capability called cloud bursting, which allows a cluster to grow and shrink as a function of the needs placed in it. In particular, growth of the cluster uses public cloud resources to cost-effectively grow the capacity of the cluster.
- The Top500 project maintains a ranking of the world's fastest supercomputers, along with details on current and historical systems. Top500 releases its updated list twice per year. Not surprisingly, the last 10 positions in the Top500 all ran a form of Linux.
- IBM supercomputers have a long and distinguished history. Check out this reference from Wikipedia for a list and details of some of the supercomputers developed by IBM.
- The SLURM Programmer's Guide provides a readable introduction to build plug-ins and generally get to know the layout of the SLURM source and documentation.
- The SLURM Configuration Tool provides a simple way to generate a slurm.conf file from a simple form of input. After entering the required set of information, the website emits a slurm.conf that is ready to drop into your cluster.
- In the developerWorks Linux zone, find hundreds of how-to articles and tutorials, as well as downloads, discussion forums, and a wealth of other resources for Linux developers and administrators.
- Evaluate IBM products in the way that suits you best: Download a product trial, try a product online, use a product in a cloud environment, or spend a few hours in the SOA Sandbox learning how to implement Service Oriented Architecture efficiently.
- Follow Tim on Twitter. You can also follow developerWorks on Twitter, or subscribe to a feed of Linux tweets on developerWorks.