<!--#include virtual="header.txt"-->

<h1>SLURM User and Administrator Guide for Cray systems</h1>

<b>NOTE: As of January 2009, the SLURM interface to Cray systems is incomplete.</b>

<h2>User Guide</h2>

<p>This document describes the unique features of SLURM on
Cray computers.
You should be familiar with the SLURM's mode of operation on Linux clusters 
before studying the relatively few differences in Cray system 
operation described in this document.</p>

<p>SLURM's primary mode of operation is designed for use on clusters with
nodes configured in a one-dimensional space. 
Minor changes were required for the <i>smap</i> and <i>sview</i> tools 
to map nodes in a three-dimensional space. 
Some changes are also desirable to optimize job placement in three-dimensional 
space.</p>

<p>SLURM has added an interface to Cray's Application Level Placement Scheduler
(ALPS). The ALPS <i>aprun</i> command must used for task launch rather than SLURM's 
<i>srun</i> command. You should create a resource reservation using SLURM's
<i>salloc</i> or <i>sbatch</i> command and execute <i>aprun</i> from within
that allocation. <//p>

<h2>Administrator Guide</h2>

<h3>Cray/ALPS configuration</h3>

<p>Node names must have a three-digit suffix describing their 
zero-origin position in the X-, Y- and Z-dimension respectively (e.g. 
"tux000" for X=0, Y=0, Z=0; "tux123" for X=1, Y=2, Z=3). 
Rectangular prisms of nodes can be specified in SLURM commands and
configuration files using the system name prefix with the end-points 
enclosed in square brackets and separated by an "x". 
For example "tux[620x731]" is used to represent the eight nodes in a 
block with endpoints at "tux620" and "tux731" (tux620, tux621, tux630, 
tux631, tux720, tux721, tux730, tux731).
<b>NOTE:</b> We anticipate that Cray will provide node coordinate
information via the ALPS interface in the future, which may result
in a more flexible node naming convention.</p>

<p>In ALPS, configure each node to be scheduled using SLURM as type
BATCH.</p>

<h3>SLURM configuration</h3>

<p>Four variables must be defined in the <i>config.h</i> file: 
<i>APBASIL_LOC</i> (location of the <i>apbasil</i> command), 
<i>HAVE_FRONT_END</i>, <i>HAVE_CRAY_XT</i> and <i>HAVE_3D</i>.
The <i>apbasil</i> command should automatically be found. 
If that is not the case, please notify us of its location on your system
and we will add that to the search paths tested at configure time.
The other variable definitions can be initiated in several different 
ways depending upon how SLURM is being built.
<ol>
<li>Execute the <i>configure</i> command with the option 
<i>--enable-cray-xt</i> <b>OR</b></li>
<li>Execute the <i>rpmbuild</i> command with the option 
<i>--with cray_xt</i> <b>OR</b></li>
<li>Add <i>%with_cray_xt 1</i> to your <i>~/.rpmmacros</i> file.</li>
</ol></p>

<p>One <i>slurmd</i> will be used to run all of the batch jobs on
the system. It is from here that users will execute <i>aprun</i>
commands to launch tasks.
This is specified in the <i>slurm.conf</i> file by using the
<i>NodeName</i> field to identify the compute nodes and both the
<i>NodeAddr</i> and <i>NodeHostname</i> fields to identify the 
computer when <i>slurmd</i> runs (normally some sort of front-end node)
as seen in the examples below.</p>

<p>Next you need to select from two options for the resource selection 
plugin (the <i>SelectType</i> option in SLURM's <i>slurm.conf</i> configuration
file):
<ol>
<li><b>select/cons_res</b> - Performs a best-fit algorithm based upon a 
one-dimensional space to allocate whole nodes, sockets, or cores to jobs
based upon other configuration parameters.</li>
<li><b>select/linear</b> - Performs a best-fit algorithm based upon a 
one-dimensional space to allocate whole nodes to jobs.</li>
</ol>

<p>In order for <i>select/cons_res</i> or <i>select/linear</i> to 
allocate resources physically nearby in three-dimensional space, the 
nodes be specified in SLURM's <i>slurm.conf</i> configuration file in 
such a fashion that those nearby in <i>slurm.conf</i> (one-dimensional
space) are also nearby in the physical three-dimensional space. 
If the definition of the nodes in SLURM's <i>slurm.conf</i> configuration 
file are listed on one line (e.g. <i>NodeName=tux[000x333]</i>),
SLURM will automatically perform that conversion using a 
<a href="http://en.wikipedia.org/wiki/Hilbert_curve">Hilbert curve</a>.
Otherwise you may construct your own node name ordering and list them
one node per line in <i>slurm.conf</i>.
Note that each node must be listed exactly once and consecutive
nodes should be nearby in three-dimensional space. 
Also note that each node must be defined individually rather than using 
a hostlist expression in order to preserve the ordering (there is no 
problem using a hostlist expression in the partition specification after
the nodes have already been defined).
The open source code used by SLURM to generate the Hilbert curve is 
included in the distribution at <i>contribs/skilling.c</i> in the event
that you wish to experiment with it to generate your own node ordering.
Two examples of SLURM configuration files are shown below:</p>

<pre>
# slurm.conf for Cray XT system of size 4x4x4
# Parameters removed here
SelectType=select/linear
NodeName=DEFAULT Procs=8 RealMemory=2048 State=Unknown
NodeName=tux[000x333] NodeAddr=front_end NodeHostname=front_end
PartitionName=debug Nodes=tux[000x333] Default=Yes State=UP
</pre>

<pre>
# slurm.conf for Cray XT system of size 4x4x4
# Parameters removed here
SelectType=select/linear
NodeName=DEFAULT Procs=8 RealMemory=2048 State=Unknown
NodeName=tux000 NodeAddr=front_end NodeHostname=front_end
NodeName=tux100 NodeAddr=front_end NodeHostname=front_end
NodeName=tux110 NodeAddr=front_end NodeHostname=front_end
NodeName=tux010 NodeAddr=front_end NodeHostname=front_end
NodeName=tux011 NodeAddr=front_end NodeHostname=front_end
NodeName=tux111 NodeAddr=front_end NodeHostname=front_end
NodeName=tux101 NodeAddr=front_end NodeHostname=front_end
NodeName=tux001 NodeAddr=front_end NodeHostname=front_end
PartitionName=debug Nodes=tux[000x111] Default=Yes State=UP
</pre>

<p>In both of the examples above, the node names output by the
<i>scontrol show nodes</i> will be ordered as defined (sequentially 
along the Hilbert curve or per the ordering in the <i>slurm.conf</i> file)
rather than in numeric order (e.g. "tux001" follows "tux101" rather 
than "tux000"). 
SLURM partitions should contain nodes which are defined sequentially
by that ordering for optimal performance.</p>

<p class="footer"><a href="#top">top</a></p>

<p style="text-align:center;">Last modified 9 January 2009</p></td>

<!--#include virtual="footer.txt"-->