Skip to main content

Cornelis Technical Documentation

3.2.1. Fabric Design Prerequisites

It is important that the design and installation of the hardware be planned carefully prior to the installation and setup of the fabric. The design plan must include the following information:

  • Identification of servers that will function as the administration or management Nodes.

    • Server memory requirements based on the software being used:

      • The OPX Software, including the Fabric Manager, requires at least 2.5 GB of physical memory for each Fabric Manager instance.

        • When managing a cluster of 10000 nodes or more, 5 GB of memory per Fabric Manager instance is required.

        • For very large fabrics, with more than 16000 nodes, 15 GB of memory per Fabric Manager instance is required.

      • When running multiple Fabric Manager instances on a single management node, the physical memory requirements should be multiplied by the number of Fabric Manager instances.

    • Swap disk space allowance should follow recommendations for the given version of Linux. Swap disk space should be twice the size of the physical memory on the server running the Fabric Manager. Minimum swap disk space should be at least 5 GB + 34 GB (for PM short-term history).

    • Cornelis recommends, but does not require, the following:

      • 32 GB or more of ECC memory (ideally fully populating DIMM channels on processor to ensure maximum memory bandwidth)

      • 256 GB or more of storage (ideally RAIDed for resiliency)

  • Plan the cabling of the fabric and create a cable planning spreadsheet using the sample .xlsx files installed into /usr/share/opa/samples/ on the management node. After a plan is established, opaxlattopology or opaxlttopology_cust may be used to convert the spreadsheet into a topology XML file that can be used during fabric verification, as well as being potentially used by the Fabric Manager for topology verification at runtime.

    Note

    The use of cable planning spreadsheets and the resulting topology XML files is highly recommended as an effective and efficient way to ensure the cluster is assembled and installed as intended.

  • Plan the naming conventions for hosts and switches in the fabric. Cornelis recommends that all switches and hosts be given unique names. Having unique names simplifies operations that are performed using host and switch names.

Error-Correcting Code - This technique detects and fixes errors in data transmission and storage.

Dual In-Line Memory Module - This type of computer memory uses random access memory mounted on a printed circuit board to connect to a motherboard, often called a stick

Redundant Array of Independent Disks - This is a data storage virtualization technology that combines multiple physical disk drive components into one or more logical units for the purposes of data redundancy, performance improvement, or both.