This information is meant primarily for the Slurm developers.
System administrators should read the instructions at
https://slurm.schedmd.com/quickstart_admin.html
(also found in the file doc/html/quickstart_admin.shtml).
The "INSTALL" file contains generic Linux build instructions.
Simple build/install on Linux:
./configure --enable-debug \
--prefix=<install-dir> --sysconfdir=<config-dir>
make
make install
To build the files in the contribs directory:
make contrib
make install-contrib
(The RPMs are built by default)
If you make changes to any auxdir/* or Makefile.am file, then run
_snowflake_ (where there are recent versions of autoconf, automake
and libtool installed):
./autogen.sh
then check-in the new Makefile.am and Makefile.in files
Here is a step-by-step HOWTO for creating a new release of Slurm on a
Linux cluster (See BlueGene and AIX specific notes below for some differences).
0. Get current copies of Slurm and buildfarm
> git clone https://<user_name>@github.com/chaos/slurm.git
> svn co https://eris.llnl.gov/svn/chaos/private/buildfarm/trunk buildfarm
place the buildfarm directory in your search path
> export PATH=~/buildfarm:$PATH
1. Update NEWS and META files for the new release. In the META file,
the API, Major, Minor, Micro, Version, and Release fields must all
by up-to-date. **** DON'T UPDATE META UNTIL RIGHT BEFORE THE TAG ****
The Release field should always be 1 unless one of
the following is true
- Changes were made to the spec file, documentation, or example
files, but not to code.
- this is a prerelease (Release = 0.preX)
2. Tag the repository with the appropriate name for the new version.
Note the first three digits are the version number. For a proper release,
the last digit is "1" (except for a rebuild without code changes which
could be "2"). For pre-releases, the last digit should be "0" followed by
"pre#" or "rc#".
> git tag -a slurm-2-6-7-1 -m "create tag v2.6.7" OR
> git tag -a slurm-2-7-0-0pre5 -m "create tag v2.7.0-pre5"
> git push --tags
3. Use the rpm make target to create the new RPMs. This requires a .rpmmacros
(.rpmrc for newer versions of rpmbuild) file containing:
%_slurm_sysconfdir /etc/slurm
%_with_debug 1
%_with_sgijob 1
NOTE: build will make a tar-ball based upon ALL of the files in your current
local directory. If that includes scratch files, everyone will get those
files in the tar-ball. For that reason, it is a good idea to clone a clean
copy of the repository and build from that
> git clone https://<user_name>@github.com/chaos/slurm.git <local_dir>
Build using the following syntax:
> build --snapshot -s <local_dir> OR
> build --nosnapshot -s <local_dir>
--nosnapshot will name the tar-ball and RPMs based upon the META file
--snapshot will name the tar-ball and RPMs based upon the META file plus a
timestamp. Do this to make a tar-ball for a non-tagged release.
NOTE: <local_dir> should be a fully-qualified pathname
4. scp the files to schedmd.com in to ~/www/download/latest or
~/www/download/development. Move the older files to ~/www/download/archive,
login to schedmd.com, cd to ~/download, and execute "php process.php" to
update the web pages.
BlueGene build notes:
0. If on a bgp system and you want sview export these variables
> export CFLAGS="-I/opt/gnome/lib/gtk-2.0/include -I/opt/gnome/lib/glib-2.0/include $CFLAGS"
> export LIBS="-L/usr/X11R6/lib64 $LIBS"
> export CMD_LDFLAGS='-L/usr/X11R6/lib64'
> export PKG_CONFIG_PATH="/opt/gnome/lib64/pkgconfig/:$PKG_CONFIG_PATH"
1. Use the rpm make target to create the new RPMs. This requires a .rpmmacros
(.rpmrc for newer versions of rpmbuild) file containing:
%_prefix /usr
%_slurm_sysconfdir /etc/slurm
%_with_bluegene 1
%_without_pam 1
%_with_debug 1
Build on Service Node with using the following syntax
> rpmbuild -ta slurm-...bz2
The RPM files get written to the directory
/usr/src/packages/RPMS/ppc64
To build and run on AIX:
0. Get current copies of Slurm and buildfarm
> git clone https://<user_name>@github.com/chaos/slurm.git
> svn co https://eris.llnl.gov/svn/chaos/private/buildfarm/trunk buildfarm
put the buildfarm directory in your search path
> export PATH=~/buildfarm:$PATH
Put the buildfarm directory in your search path
Also, you will need several commands to appear FIRST in your PATH:
/usr/local/tools/gnu/aix_5_64_fed/bin/install
/usr/local/gnu/bin/tar
/usr/bin/gcc
I do this by making symlinks to those commands in the buildfarm directory,
then making the buildfarm directory the first one in my PATH.
Also, make certain that the "proctrack" rpm is installed.
1. Export some environment variables
> export OBJECT_MODE=32
> export PKG_CONFIG="/usr/bin/pkg-config"
2. Build with:
> ./configure --enable-debug --prefix=/opt/freeware \
--sysconfdir=/opt/freeware/etc/slurm \
--with-ssl=/opt/freeware --with-munge=/opt/freeware \
--with-proctrack=/opt/freeware
make
make uninstall # remove old shared libraries, aix caches them
make install
3. To build RPMs (NOTE: GNU tools early in PATH as described above in #0):
Create a .rpmmacros file specifying system specific files:
#
# RPM Macros for use with Slurm on AIX
# The system-wide macros for RPM are in /usr/lib/rpm/macros
# and this overrides a few of them
#
%_prefix /opt/freeware
%_slurm_sysconfdir %{_prefix}/etc/slurm
%_defaultdocdir %{_prefix}/doc
%_with_debug 1
%_with_aix 1
%with_ssl "--with-ssl=/opt/freeware"
%with_munge "--with-munge=/opt/freeware"
%with_proctrack "--with-proctrack=/opt/freeware"
Log in to the machine "uP". uP is currently the lowest-common-denominator
AIX machine.
NOTE: build will make a tar-ball based upon ALL of the files in your current
local directory. If that includes scratch files, everyone will get those
files in the tar-ball. For that reason, it is a good idea to clone a clean
copy of the repository and build from that
> git clone https://<user_name>@github.com/chaos/slurm.git <local_dir>
Build using the following syntax:
> export CC=/usr/bin/gcc
> build --snapshot -s <local_dir> OR
> build --nosnapshot -s <local_dir>
--nosnapshot will name the tar-ball and RPMs based upon the META file
--snapshot will name the tar-ball and RPMs based upon the META file plus a
timestamp. Do this to make a tar-ball for a non-tagged release.
4. Test POE after telling POE where to find Slurm's LoadLeveler wrapper.
> export MP_RMLIB=./slurm_ll_api.so
> export CHECKPOINT=yes
5. > poe hostname -rmpool debug
6. To debug, set SLURM_LL_API_DEBUG=3 before running poe - will create a file
/tmp/slurm.*
It can also be helpful to use poe options "-ilevel 6 -pmdlog yes"
There will be a log file create named /tmp/mplog.<jobid>.<taskid>
7. If you update proctrack, be sure to run "slibclean" to clear cached
version.
8. Remove the RPMs that we don't want:
rm -f slurm-perlapi*rpm slurm-torque*rpm
and install the other RPMs into /usr/admin/inst.images/slurm/aix5.3 on an
OCF AIX machine (pdev is a good choice).
Debian build notes:
Since Debian doesn't have PRMs, the rpmbuild program can not locate
dependencies, so build without them by patching the build program:
Index: build
===================================================================
--- build (revision 173)
+++ build (working copy)
@@ -798,6 +798,7 @@
$cmd .= " --define \"_tmppath $rpmdir/TMP\"";
$cmd .= " --define \"_topdir $rpmdir\"";
$cmd .= " --define \"build_bin_rpm 1\"";
+ $cmd .= " --nodeps";
if (defined $conf{rpm_dist}) {
my $dist = length $conf{rpm_dist} ? $conf{rpm_dist} : "%{nil}";
$cmd .= " --define \"dist $dist\"";
AIX/Federation switch window problems
To clean switch windows: ntblclean =w 8 -a sni0
To get switch window status: ntblstatus
BlueGene bglblock boot problem diagnosis
- Logon to the Service Node (bglsn, ubglsn)
- Execute /admin/bglscripts/fatalras
This will produce a list of failures including Rack and Midplane number
<date> R<rack> M<midplane> <failure details>
- Translate the Rack and Midplane to Slurm node id: smap -R r<rack><midplane>
- Drain only the bad Slurm node, return others to service using scontrol
Configuration file update procedures:
- cd /usr/bgl/dist/slurm (on bgli)
- co -l <filename>
- vi <filename>
- ci -u <filename>
- make install
- then run "dist_local slurm" on SN and FENs to update /etc/slurm
Some RPM commands:
rpm -qa | grep slurm (determine what is installed)
rpm -qpl slurm-1.1.9-1.rpm (check contents of an rpm)
rpm -e slurm-1.1.8-1 (erase an rpm)
rpm --upgrade slurm-1.1.9-1.rpm (replace existing rpm with new version)
rpm -i --ignoresize slurm-1.1.9-1.rpm (install a new rpm)
For main Slurm plugin installation on BGL service node:
rpm -i --force --nodeps --ignoresize slurm-1.1.9-1.rpm
rpm -U --force --nodeps --ignoresize slurm-1.1.9-1.rpm (upgrade option)
To clear a wedged job:
/bgl/startMMCSconsole
> delete bgljob ####
> free RMP###
Starting and stopping daemons on Linux:
/etc/init.d/slurm stop
/etc/init.d/slurm start
Patches:
- cd to the top level src directory
- Run the patch command with epilog_complete.patch as stdin:
patch -p[path_level_to_filter] [--dry-run] < epilog_complete.patch
To get the process and job IDs with proctrack/sgi_job:
- jstat -p
CVS and gnats:
Include "gnats:<id> e.g. "(gnats:123)" as part of cvs commit to
automatically record that update in gnats database. NOTE: Does
not change gnats bug state, but records source files associated
with the bug.
For memory leaks (for AIX use zerofault, zf; for linux use valgrind)
- Run configure with the option "--enable-memory-leak-debug" to completely
release allocated memory when the daemons exit
- valgrind --tool=memcheck --leak-check=yes --num-callers=8 --leak-resolution=med \
./slurmctld -Dc >valg.ctld.out 2>&1
- valgrind --tool=memcheck --leak-check=yes --num-callers=8 --leak-resolution=med \
./slurmd -Dc >valg.slurmd.out 2>&1 (Probably only one one node of cluster)
- valgrind --tool=memcheck --leak-check=yes --num-callers=8 --leak-resolution=med \
./slurmdbd -D >valg.dbd.out 2>&1
- Run the regression test. In the globals.local file include:
"set enable_memory_leak_debug 1"
- Shutdown the daemons using "scontrol shutdown"
- Examine the end of the log files for leaks. pthread_create() and dlopen()
have small memory leaks on some systems, which do not grow over time
- Functions in the plugins will not have their symbols preserved when the
plugin is unloaded and the function names will appear as "???" in the
valgrind backtrace after shutdown. Rebuilding the daemons without the
configure option of "--enable-memory-leak-debug" typically prevents the
plugin from being unloaded so the symbols will be properly reported. However
many memory leaks will be reported due to not unloading plugins. You will
need to match the call sequence from the first log (with
"--enable-memory-leak-debug") to the second log (without
"--enable-memory-leak-debug" and ignore memory leaks reported that are not
real leaks in the second log) to identify the full code path through plugins.
Job profiling:
- "export CFLAGS=-pg", then run "configure" and "make install" as usual.
- Run the slurm daemons through a stress test and exit normally
- Run "gprof [executable-file] >outfile"
Before new major release:
- Test on ia64, i386, x86_64, BGQ, NRT/POE, Cray
- Test on Elan and IB switches
- Test fail-over of slurmctld
- Test pam_slurm and pam_slurm_adopt.
- Test for memory leaks in slurmctld, slurmd and slurmdbd with various plugins
- Change API version number
- Run "make check" (requires "dejagnu" package)
- Test that the prolog and epilog run
- Run the test suite with SlurmUser NOT being self
- Test for errors reported by CLANG tool:
NOTE: Run "configure" with "--enable-developer" option so assert functions
take effect.
scan-build -k -v make >m.sb.out 2>&1
# and look for output in /tmp/scan-build-<DATE>