Large Cluster Administration Guide
This document contains Slurm administrator information specifically for clusters containing 1,024 nodes or more. Large system currently managed by Slurm include Tianhe-2 (at the National University of Defense Technology in China with 16,000 compute nodes and 3.1 million cores) and Sequoia (IBM Bluegene/Q at Lawrence Livermore National Laboratory with 98,304 compute nodes and 1.6 million cores). Slurm operation on systems orders of magnitude larger has been validated using emulation. Getting optimal performance at that scale does require some tuning and this document should help you off to a good start. A working knowledge of Slurm should be considered a prerequisite for this material.
Times below are for execution of an MPI program printing "Hello world" and exiting and includes the time for processing output. Your performance may vary due to differences hardware, software, and configuration.
- 1,966,080 tasks on 122,880 compute nodes of a BlueGene/Q: 322 seconds
- 30,000 tasks on 15,000 compute nodes of a Linux cluster: 30 seconds
Three system configuration parameters must be set to support a large number of open files and TCP connections with large bursts of messages. Changes can be made using the /etc/rc.d/rc.local or /etc/sysctl.conf script to preserve changes after reboot. In either case, you can write values directly into these files (e.g. "echo 32832 > /proc/sys/fs/file-max").
- /proc/sys/fs/file-max: The maximum number of concurrently open files. We recommend a limit of at least 32,832.
- /proc/sys/net/ipv4/tcp_max_syn_backlog: Maximum number of remembered connection requests, which are still did not receive an acknowledgment from connecting client. The default value is 1024 for systems with more than 128Mb of memory, and 128 for low memory machines. If server suffers of overload, try to increase this number.
- /proc/sys/net/core/somaxconn: Limit of socket listen() backlog, known in userspace as SOMAXCONN. Defaults to 128. The value should be raised substantially to support bursts of request. For example, to support a burst of 1024 requests, set somaxconn to 1024.
The transmit queue length (txqueuelen) may also need to be modified
using the ifconfig command. A value of 4096 has been found to work well for one
site with a very large cluster
There is a newly introduced limit in SLES 12 SP2 (used on Cray systems with CLE 6.0UP04, to be released mid-2017). The version of systemd shipped with SLES 12 SP2 contains support for the PIDs cgroup controller. Under the new systemd version, each init script or systemd service is limited to 512 threads/processes by default. This could cause issues for the slurmctld and slurmd daemons on large clusters or systems with a high job throughput rate. To increase the limit beyond the default:
- If using a systemd service file: Add TasksMax=N to the [Service] section. N can be a specific number, or special value infinity.
- If using an init script: Create the file
/etc/systemd/system/<init script name>.service.d/override.conf
with these contents:
Note: Earlier versions of systemd that don't support the PIDs cgroup controller simply ignore the TasksMax setting.
The ulimit values in effect for the slurmctld daemon should be set quite high for memory size, open file count and stack size.
Node Selection Plugin (SelectType)
While allocating individual processors within a node is great for smaller clusters, the overhead of keeping track of the individual processors and memory within each node adds significant overhead. For best scalability, allocate whole nodes using select/linear and avoid select/cons_res.
Job Accounting Gather Plugin (JobAcctGatherType)
Job accounting relies upon the slurmstepd daemon on each compute node periodically sampling data. This data collection will take compute cycles away from the application inducing what is known as system noise. For large parallel applications, this system noise can detract for application scalability. For optimal application performance, disabling job accounting is best (jobacct_gather/none). Consider use of job completion records (JobCompType) for accounting purposes as this entails far less overhead. If job accounting is required, configure the sampling interval to a relatively large size (e.g. JobAcctGatherFrequency=300). Some experimentation may also be required to deal with collisions on data transmission.
While Slurm can track the amount of memory and disk space actually found on each compute node and use it for scheduling purposes, this entails extra overhead. Optimize performance by specifying the expected configuration using the available parameters (RealMemory, CPUs, and TmpDisk). If the node is found to contain less resources than configured, it will be marked DOWN and not used. Also set the FastSchedule parameter. While Slurm can easily handle a heterogeneous cluster, configuring the nodes using the minimal number of lines in slurm.conf will both make for easier administration and better performance.
The EioTimeout configuration parameter controls how long the srun command will wait for the slurmstepd to close the TCP/IP connection used to relay data between the user application and srun when the user application terminates. The default value is 60 seconds. Larger systems and/or slower networks may need a higher value.
If a high throughput of jobs is anticipated (i.e. large numbers of jobs with brief execution times) then configure MinJobAge to the smallest interval practical for your environment. MinJobAge specifies the minimum number of seconds that a terminated job will be retained by Slurm's control daemon before purging. After this time, information about terminated jobs will only be available through accounting records.
The configuration parameter SlurmdTimeout determines the interval at which slurmctld routinely communicates with slurmd. Communications occur at half the SlurmdTimeout value. The purpose of this is to determine when a compute node fails and thus should not be allocated work. Longer intervals decrease system noise on compute nodes (we do synchronize these requests across the cluster, but there will be some impact upon applications). For really large clusters, SlurmdTimeout values of 120 seconds or more are reasonable.
If MPICH-2 is used, the srun command will manage the key-pairs used to bootstrap the application. Depending upon the processor speed and architecture, the communication of key-pair information may require extra time. This can be done by setting an environment variable PMI_TIME before executing srun to launch the tasks. The default value of PMI_TIME is 500 and this is the number of microseconds alloted to transmit each key-pair. We have executed up to 16,000 tasks with a value of PMI_TIME=4000.
The individual slurmd daemons on compute nodes will initiate messages to the slurmctld daemon only when they start up or when the epilog completes for a job. When a job allocated a large number of nodes completes, it can cause a very large number of messages to be sent by the slurmd daemons on these nodes to the slurmctld daemon all at the same time. In order to spread this message traffic out over time and avoid message loss, The EpilogMsgTime parameter may be used. Note that even if messages are lost, they will be retransmitted, but this will result in a delay for reallocating resources to new jobs.
Slurm uses hierarchical communications between the slurmd daemons in order to increase parallelism and improve performance. The TreeWidth configuration parameter controls the fanout of messages. The default value is 50, meaning each slurmd daemon can communicate with up to 50 other slurmd daemons and over 2500 nodes can be contacted with two message hops. The default value will work well for most clusters. Optimal system performance can typically be achieved if TreeWidth is set to the square root of the number of nodes in the cluster for systems having no more than 2500 nodes or the cube root for larger systems.
The srun command automatically increases its open file limit to the hard limit in order to process all of the standard input and output connections to the launched tasks. It is recommended that you set the open file hard limit to 8192 across the cluster.
Last modified 14 June 2018