VLF Group compute cluster – Infiniband installation

Background notes

Background notes

These instructions are for Fedora Core 13. Infiniband is a relatively complicated network stack, and the people who maintain the drivers (the OpenFabrics Alliance) are slow to release drivers for newer linux versions. This effectively limits us to using either an ancient, explicitly supported distribution like RHEL, or a slightly less ancient, unsupported distribution like Fedora Core 13. Upgrading the Linux distribution on the Infiniband nodes (cluster001 through cluster016) should only be done if it is known the Infiniband drivers can be compiled and installed on that distribution. We need at a bare minimum the QIB driver, but there are some other requisite parts of the stack needed by the network and other utilities (e.g., IP over IB).

Pre-installation

Note: these steps are consolidated in the file /shared/clientconfig/infiniband_install/preinstall.sh.

Before installing, it is necessary to remove the existing drivers (which don’t work properly), set some compiler flags, and get rid of a spurious autoconf.h that breaks the build:


rm -f /lib/modules/`uname -r`/build/include/linux/autoconf.h
export CFLAGS="-lpthread -ldl"

mv /lib/modules/`uname -r`/kernel/drivers/infiniband/hw/ipath/ib_ipath.ko \
  /lib/modules/`uname -r`/kernel/drivers/infiniband/hw/ipath/ib_ipath.ko.disabled
rmmod ib_ipath
rmmod ib_core

Build (do this only once to generate the RPM files)

Untar the OFED package:


tar xvfz OFED-1.5.4.1.tgz

Run the installation script:


cd OFED-1.5.4.1
./install.pl

Choose option 2) Install OFED Software. On the next menu, choose 4) Customize.

Select yes on each of the following modules:


kernel-ib core mthca qib ipoib sdp srp rds kernel-ib-devel libibverbs
libibverbs-devel libibverbs-devel-static libibverbs-utils
libibverbs-debuginfo libipathverbs libipathverbs-devel
libipathverbs-debuginfo libibcm libibcm-devel libibcm-debuginfo
libibumad libibumad-devel libibumad-static libibumad-debuginfo
libibmad libibmad-devel libibmad-static libibmad-debuginfo ibsim
ibsim-debuginfo ibacm librdmacm librdmacm-utils librdmacm-devel
librdmacm-debuginfo libsdp libsdp-devel libsdp-debuginfo opensm
opensm-libs opensm-devel opensm-debuginfo opensm-static compat-dapl
compat-dapl-devel dapl dapl-devel dapl-devel-static dapl-utils
dapl-debuginfo perftest mstflint sdpnetstat srptools rds-tools
rds-devel ibutils infiniband-diags qperf qperf-debuginfo ofed-docs
infinipath-psm infinipath-psm-devel

Be sure to select no on these modules:


libmthca libmthca-devel-static libmthca-debuginfo libmlx4
libmlx4-devel libmlx4-debuginfo libcxgb3 libcxgb3-devel
libcxgb3-debuginfo libcxgb4 libcxgb4-devel libcxgb4-debuginfo libnes
libnes-devel-static libnes-debuginfo ofed-scripts mpi-selector
mvapich_gcc mvapich2_gcc openmpi_gcc mpitests_mvapich_gcc
mpitests_mvapich2_gcc mpitests_openmpi_gcc build32

Install (do this once per machine after building)

Be sure to use the -c ofed.conf option to re-use the options selected when building in the first place. This step will use the RPMs just built in the previous step (they’re in the RPMS/fedora-release-13-1/x86_64/ directory):


cd OFED-1.5.4.1
./install.pl -c ofed.conf

Post-install steps

Note: these steps are consolidated in the file /shared/clientconfig/infiniband_install/postinstall.sh.

Disable the network manager:


chkconfig network on
chkconfig NetworkManager off
/etc/init.d/NetworkManager stop

Add some flags to force the driver to try to negotiate the fastest data rate, or fail. This is preferable to coming up silently with the wrong infiniband bitrate:


echo "options ib_qib compat_ddr_negotiate=0 ibmtu=4" > /etc/modprobe.d/ib_qib.conf

Apply a ulimit fix so MPI works properly


echo "* soft memlock unlimited" >> /etc/security/limits.conf
echo "* hard memlock unlimited" >> /etc/security/limits.conf

Configure the network:


cd /shared/clientconfig/infiniband_install
cp ifcfg-ib0 /etc/sysconfig/network-scripts/ifcfg-ib0
cp ifcfg-eth0 /etc/sysconfig/network-scripts/ifcfg-eth0

Edit the network configuration files just copied in the step above. Change the IPADDR line to the host-specific IP address (you choose this).