1. Introduction

This chapter describes the LSF Parallel system and its architecture.

This chapter discusses the following topics:

What Is the LSF Parallel System?

How Does LSF Parallel Fit Into the LSF Batch System?

LSF Parallel Architecture

What Is the LSF Parallel System?

The LSF Parallel system is a fully supported commercial software system that supports the programming, testing, and execution of parallel applications in production environments.

The LSF Parallel system is fully integrated with the LSF Batch system, the de-facto industry standard resource management software product, to provide load sharing in a distributed system and batch scheduling for compute intensive jobs. The LSF Parallel system provides support for:

• Dynamic resource discovery and allocation (resource reservation) for parallel batch job execution
• Transparent invocation of the distributed job processes
• Full control of the distributed job processes to ensure no processes will become un-managed. This effectively reduces the possibility of one parallel job causing severe disruption to an organization's computer service.
• The standard MPI interface
• All major UNIX operating systems
• Full integration with the LSF Batch system providing heterogeneous resource-based batch job scheduling.

How Does LSF Parallel Fit Into the LSF Batch System?

The LSF Parallel system adopts a layered approach, shown in Figure 1, that is fully integrated with the LSF Batch system. In addition to the LSF Batch system resources, the following components make up the LSF Parallel system:

• The MPI Library
• The Parallel Application Manager (PAM).

Figure 1 Major Components of the LSF Parallel system

MPI Library

The Message Passing Interface (MPI) library is a message-passing library that must be linked to the parallel applications that are to be run in the LSF Batch system. The MPI library translates MPI message calls to messages for the machine-dependent layer and it interfaces the user application to PAM.

See Vendor MPI Implementations on page 41 for a description of vendor-specific MPI implementations.

PAM

The Parallel Application Manager (PAM) is the point of control for the LSF Parallel system. PAM is fully integrated with the LSF Batch system. PAM interfaces the user application with the LSF Batch system. For all parallel application processes (tasks), PAM:

• Maintains the communication connection map
• Monitors and forwards control signals
• Receives requests to add, delete, start, and connect tasks
• Monitors resource usage while the user application is running
• Enforces job-level resource limits
• Collects resource usage information and exit status upon termination
• Handles standard I/O.

LSF Batch System

The LSF Batch system is a sophisticated resource-based batch job scheduling system. It accepts user jobs and holds them in queues until suitable hosts are available and resource requirements are satisfied. Host selection is based on up-to-the-minute load information provided by the master Load Information Manager (LIM).

LSF Batch runs user jobs on batch server hosts. It has sophisticated controls for sharing hosts with interactive users; there is no need to set aside dedicated hosts for processing batch jobs.

See the LSF Batch User's Guide and the LSF Batch Administrator's Guide for a detailed description of the LSF Batch system.

LSF Parallel Architecture

The LSF Parallel system fully utilizes the resources of the LSF Batch System for resource selection and process invocation and control. The process of parallel batch job invocation and control is shown in Figure 2 and described in Table 1 on page 6. The LSF components shown in Figure 2 are described in Table 2 on page 6.

Figure 2 LSF Parallel Architecture

The LSF Parallel system also supports interactive parallel job submission. The process is similar to that shown in Figure 2 except the user request is submitted directly to PAM which makes a simple placement query to LIM. Job queuing and resource reservations are not supported in interactive mode.