Intel® Cluster Ready Architecture Specification 1.3
The Intel® Cluster Ready Architecture Specification has been updated to version 1.3. This replaces the previous architecture specification, version 1.2, published on March 24, 2011. The new architecture specification maintains the same goals and overall design as its predecessor, but updates many areas to reflect changes to technology and software requirements.
One significant change is a new section covering requirements for add-in coprocessors, accelerators, and GPGPUs. General-purpose (GP) coprocessor devices, as they are called generically in the specification, operate independently and in parallel with the host-platform CPU. These new requirements for software and operation are needed to maintain the Intel Cluster Ready promise.
To reflect increased memory requirements for many applications, the minimum memory requirement for all compute nodes has been changed to 8 gibibytes (GiB) (8 × 230 bytes). In addition, compute node persistent storage must be at least 20 GiB and the head node storage is 200 GiB.
The minimum Linux* kernel is now 2.6.32, which is required to support many hew hardware technologies. Exceptions for Red Hat* Enterprise Linux (RHEL) and clones, which are branched from older kernels, are included. Red Hat Enterprise Linux 5.6 or later will meet new requirements.
As part of the specification development, ISV applications are regularly checked for Linux operating system requirements. Using an automated tool, each registered application is analyzed to determine which system libraries and binaries are called. This data is then used to define library and binary requirements in the architecture specification. The latest analysis revealed that several libraries and one binary are used by applications and are readily available in all relevant Linux distributions.
- libcap.so.2.
- libcursesw.so.5
- libpanelw.so.5.
- libcrypto.so.6
- libnuma.so.1
- libuuid.so.1
- libGLU.so.1
- xauth
Conversely, the requirements for libselinux.so.1 and libsepol.so.1 were removed because these libraries were not used and were not available on some Linux distributions.
The libnss_ldap.so.2 was added to match requirements for other authentication libraries.
Intel Runtime Libraries were updated to support development of applications for new technologies available as of 2012. The updates include the following or later versions of these runtime components:
- Intel® C++ Compiler 2013.1
- Intel® Fortran Compiler 2013.1
- Intel® Math Kernel Library 11.01
- Intel® MPI Library 4.1.0
- Intel® Threading Building Blocks 4.1.1
All of the Intel® software runtime components are provided for program members. They can be downloaded from Intel® Registration Center under Intel Cluster Ready. Contact Intel Cluster Ready support to obtain a new serial number, if needed.
Required versions of environments, scripting languages, and components have been updated. The following components or newer versions must be installed on the cluster:
- GNU C Compiler (gcc) runtimes 4.1
- OpenJDK version 6 runtimes and browser plug-in. This also reflects the transfer of the Java environment from Sun Microsystem to Oracle and an open source project.
- Perl 5.8.8
- Python 2.4.3
- Tcl 8.4.13
- X11 R7.1
- NFS 4.0
Installed MPI libraries must now conform to the MPI Standard 2.0 or later.
In order to properly test conformity, the architecture specification requires the installation of Intel Cluster Checker 1.2 or later.
Intel® Cluster Checker 2.1
The new Intel® Cluster Checker 2.1 tool provides the means to automatically evaluate a cluster for functionality, performance, and compliance to the Intel® Cluster Ready architecture specification. Although compliance against the new architecture specification 1.3 is the default, Intel Cluster Checker 2.1 can optionally check against version 1.2. The compliance-only check can be executed separately from the certification check.
The most significant addition to Intel Cluster Checker 2.1 is the ability to check Intel® Xeon Phi™ coprocessors in both native and offload modes. This includes tests in symmetric mode.
Previously available in version 1.8, the <generic> test module has returned. With this configurable test module, users can define their own cluster-wide tests. The ability to run Intel Cluster Checker as root is also included again.
The following benchmark components have been updated:
- The HPL benchmark is the same version used in the Intel® Composer Studio XE 2013.3.163.
- The Triad STREAM benchmark is now version 5.10.
Updates to the next version of Intel Cluster Checker are facilitated with a new and improved uninstaller. This is packaged in the intel-clck-pset RPM package
Intel Reference Implementation
Integrators looking for a ready-to-run solution that incorporates these updated components, Intel® Xeon Phi™ coprocessors, InfiniBand* fabric, can use the latest reference implementation. The installed cluster will support both offload and native MIC execution models, including InfiniBand communication with the Intel® Coprocessor Communication Link (CCL).
This recipe utilizes current open source software components to build a stateless cluster solution. Based on Warewulf systems management suite 3.4 and using the CentOS 6.3 Linux* OS distribution, the recipe will provide a complete Intel® Cluster Ready certified solution. The hardware configuration uses a dedicated head node and 4 to 8 compute nodes with:
- Intel® Server System R2300GZ series
- Intel® Xeon® processors E5-2600
- Intel Xeon Phi coprocessors
- Mellanox* Connect-3* host channel adapters
A similar recipe for using Intel® True Scale fabric with Intel Xeon Phi coprocessors is in development.
The recipe can be downloaded from the Intel Cluster Ready website at www.intel.com/go/cluster. It should be used as written to provide a certified solution, but it can also be used as a template for other solutions.