Planning Volume 2, Control Workstation and Software Environment

Planning for LoadLeveler

LoadLeveler is an IBM software product that provides workload management of both interactive and batch processing on your RS/6000 SP system or RS/6000 workstations. The LoadLeveler software lets you build, submit, and process both serial and parallel jobs. It allows multiple user space tasks per node, enabling applications that use the message passing interface (MPI) protocol to exploit SMP nodes or servers and realize significant performance improvements. Applications using the LAPI User Space API or Parallel Environment MPI 2.4 and LoadLeveler 2.1, or later versions of the software, can start up to four user space tasks per node with the SP Switch and up to 16 with the SP Switch2. These tasks can be from the same or from different parallel jobs. This allows parallel applications to exploit the symmetric multiprocessors on the node or SP-attached server without restructuring or recompiling. |A user space MPI or LAPI job can consist of up to 1024 tasks on SP |Switch systems and up to 4096 tasks on SP Switch2 system (up to 2048 tasks for |the MPI/IP library).

LoadLeveler is an integral piece of the total System Management solution on the RS/6000 SP system. The latest version of LoadLeveler is included with your new SP order. You choose whether to use it. LoadLeveler can take advantage of features provided in PSSP, such as event management, performance monitoring, and SP switch management. LoadLeveler will also interoperate with other schedulers to support batch job processing on other hardware platforms. These schedulers can include Network Queueing System (NQS) and the IBM Network Queuing System/MVS (NQS/MVS).

Compatibility

|The latest release is LoadLeveler 3.1. It is available |for AIX 4.3.3. For compatibility information see LoadLeveler in Chapter 11, Planning for migration.

|For the most up-to-date migration instructions, see the README file |distributed with LoadLeveler 3.1.

Planning for a highly available LoadLeveler cluster

LoadLeveler provides features within the product for automatic recovery in the event of failure of the central manager in the batch configuration and of the domain name server running Interactive Network Dispatcher in the interactive configuration. Additionally, the availability of individual compute nodes and file systems in the LoadLeveler cluster can be enhanced by using the High Availability Cluster Multi-Processing (HACMP) product as well as the High Availability Control Workstation (HACWS) optional feature of PSSP. For details on how to configure LoadLeveler for high availability, see the ITSO Redbook Implementing High Availability on the RS/6000 SP.

Planning your LoadLeveler configuration

In general, planning the LoadLeveler installation for workload management requires making the following configuration decisions. You must decide what is suitable to your environment. Be sure to address the following:

• Consider ticket and credentials lifetimes.
• Decide what ticket and credential lifetimes need to be set with respect to PE and LoadLeveler. The lifetime needs to be long enough to allow for the expected longest running job to complete.
• Consider using DCE server replication when large numbers of tasks are to be used in parallel jobs.
• The DCE user id is used if it is different from the AIX user id. IBM suggests you use DCE integrated login.
• Select a node to serve as central manager and one or more alternate central managers. The central manager can be any node in the cluster. In selecting one, consider the current workload and network access. Note that no new work can be performed while the central manager is down, and no queries can be made about any of the running jobs without the central manager.
• Determine which nodes will be scheduling nodes, execution nodes, submit-only nodes, and public submit nodes.
• Determine where to locate home and local directories. For maximum performance, keep the log, spool, and execute directories in a local file system.
• Determine if LoadLeveler daemons should communicate over the switch. It may not be desirable in your environment to have the daemons communicate over the switch. You need to evaluate the network traffic in your system to determine if LoadLeveler IP communications over the switch is desirable.
• Determine if HACMP is necessary to provide failover capability of individual compute nodes or the switch. If using LoadLeveler in conjunction with HACMP, decide which nodes will be grouped together for backup purposes. (HACMP can only provide capability for up to eight nodes.) Each backup node needs to know which set of seven nodes it will back up. This relationship is defined in the form of HACMP resource groups.
• Determine if your SP workload includes parallel jobs and if they will involve the SP Switch. If so, you will need to perform additional configuration activities. See the LoadLeveler publication for details.

Other planning considerations:

1. LoadLeveler requires a common name space for the entire LoadLeveler cluster. To run jobs on any machine in the LoadLeveler cluster, you must have the same uid (system ID number for a user) and gid (system ID number for a group) on every machine in the cluster. If you do not have a user ID on one machine, your jobs will not run on that machine.
2. LoadLeveler works in conjunction with the NFS or AFS file systems. Allowing users to share file systems to obtain a single, network-wide image, is one way to make managing LoadLeveler easier.
3. Some nodes in the LoadLeveler cluster might have special software installed that you might need to run your jobs successfully. You should configure LoadLeveler to distinguish those nodes from other nodes using, for example, job classes.