•OS Jitter mitigation through co-scheduling capability: POE provides services for periodically adjusting the dispatch priority of a user’s task between set boundaries, giving the tasks improved execution priority. As an alternative to the priority adjustment coscheduler, POE provides a run queue-based coscheduler for reducing operating system Jitter (OS Jitter). OS Jitter is operating system interference that is caused by the scheduling of daemon processes and the handling of asynchronous events, such as interrupts. This interference can have adverse effects on parallel applications running on large-scale systems.

•Affinity support of processes to CPU and memory: IBM PE Runtime Edition provides the capability to control the placement of the tasks of a parallel job using affinity to process memory or CPU placement during its execution. As a result, applications might see improved performance if the processor, the memory it uses, and the I/O adapter it connects to are in close proximity based on the affinity of the tasks to memory or CPUs it uses.

•User-controlled workflow/subjob support: IBM PE Runtime Edition provides support for launching and managing multiple parallel dynamic jobs (subjobs) using a single scheduler or resource management allocation of cluster resources.

•Lightweight core file support: Lightweight core file is designed to save CPU time, network bandwidth, and disk space that is required to generate standard core files. The lightweight format also provides the capability to easily examine the state of all threads in a parallel program at the time the event that caused the core file occurred.

•Serial and parallel job launch: Allows POE to be used not only for launching traditional MPI or other message passing programs, but also as a distributed shell to quickly obtain information about all nodes in the cluster, such as disk space, currently running jobs, and so on.

•Support running SPMD and MPMD programs: Single Process, Multiple Data (SPMD), Multiple Process, and Multiple Data (MPMD) gives users the flexibility to run the same program on all nodes (SPMD, most common) or different programs on each node (MPMD). The MPMD function is useful for master/worker programs where the master program coordinates and synchronizes the execution of all worker tasks, where neither program can run without the other.

•Resource management: POE supports running without a separate scheduler or resource manager. If a scheduler or resource manager is not used or not available, POE can manage the node and adapter resources itself. The network resource tables is also loaded by POE to establish the communication mechanism needed for message-passing programs.

•Flow control

•Support for large messages

•Support for generic non-contiguous messages

•Non-blocking calls

•Interrupt and polling modes

•Efficient exploitation of interconnect functions

•Even monitoring support (to simulate blocking calls, for example) for various types od completion events

•Support for MPI, OpenMP, OpenACC, and UPC parallel programming models

•Support for a wide range of batch systems and runtime systems, including IBM LoadLeveler, IBM Parallel Environment, Open MPI, and MPICH2

•Provide access to hardware performance counters for performing low-level analysis of an application, including analyzing cache usage and floating-point performance.

•Profile and trace an MPI application for analyzing MPI communication patterns and performance problems.

•Profile an OpenMP application for analyzing OpenMP performance problems and to help you determine if an OpenMP application properly structures its processing for best performance.

•Profile application I/O for analyzing an application’s I/O patterns and whether you can improve the application’s I/O performance.

•Profile an application’s execution for identifying hotspots in the application and for locating relationships between functions in your application to help you better understand the application’s performance.