International Business Machines Corporation

Hanhong Xue

Chulho Kim

Robert Sidney Blackmore

Tsai-Yang Jea

Ravi Kumar Arimilli

Balaram Sinharoy

Lakshminarayana Baba Arimilli

Ramakrishnan Rajamony

Hung Q Thai

Rajeev Sivaram

Deryck X Hong

Zhi Zhang

Rama Krishna Govindaraju

Jun He

Uman Chan

Bernard Charles Drerup

Paul Frank Lecocq

William P LePera

Zenon J Piatek

Abhinav Vishnu

William John Starke

Ronald Nick Kalla

Piyush Chaudhary

Ilie G Tanase

Richard R Treumann

Gary J Mincher

Serban C Maerean

Xin Chen

Wen Chong Chen

Wiliam P LePera

Charles J Archer

Brian E Smith

Michael A Blocksome

Joseph D Ratterman

Jody B Joyner

Peter H Hochschild

Jay Robert Herring

Su-Hsuan Huang

Richard J Coppinger

Scott D Epter

Amy Xin Chen

Lakshminarayana B Armilli

William P Lepera

Lakshimarayana B Arimilli

A method, apparatus and program product simulate a high performance computing (HPC) application environment by creating a cluster of virtual nodes in one or more operating system instances executing on one or more physical computing node, thereby enabling a plurality of parallel tasks from an HPC application to be executed on the cluster of virtual nodes.

Simulation of high performance computing (HPC) application environment using virtual nodes

A method, apparatus and program product utilize an empirical approach to determine the locations of one or more IO adapters in an HPC environment. Performance tests may be run using a plurality of candidate mappings that map IO adapters to various locations in the HPC environment, and based upon the results of such testing, speculative adapter affinity information may be generated that assigns one or more IO adapters to one or more locations to optimize adapter affinity performance for subsequently-executed tasks.

Empirical determination of adapter affinity in high performance computing (HPC) environment

A message flow controller limits a process from passing a new message in a reliable message passing layer from a source node to at least one destination node while a total number of in-flight messages for the process meets a first level limit. The message flow controller limits the new message from passing from the source node to a particular destination node from among a plurality of destination nodes while a total number of in-flight messages to the particular destination node meets a second level limit. Responsive to the total number of in-flight messages to the particular destination node not meeting the second level limit, the message flow controller only sends a new packet from among at least one packet for the new message to the particular destination node while a total number of in-flight packets for the new message is less than a third level limit.

Flow control for reliable message passing

Load balancing of adapters on a multi-adapter node of a communications environment. A task executing on the node selects an adapter resource unit to be used as its primary port for communications. The selection is based on the task's identifier, and facilitates a balancing of the load among the adapter resource units. Using the task's identifier, an index is generated that is used to select a particular adapter resource unit from a list of adapter resource units assigned to the task. The generation of the index is efficient and predictable.

Load balancing of adapters on a multi-adapter node

A parallel computer executes a number of tasks, each task includes a number of endpoints and the endpoints are configured to support collective operations. In such a parallel computer, establishing a group of endpoints receiving a user specification of a set of endpoints included in a global collection of endpoints, where the user specification defines the set in accordance with a predefined virtual representation of the endpoints, the predefined virtual representation is a data structure setting forth an organization of tasks and endpoints included in the global collection of endpoints and the user specification defines the set of endpoints without a user specification of a particular endpoint; and defining a group of endpoints in dependence upon the predefined virtual representation of the endpoints and the user specification.

Establishing a group of endpoints in a parallel computer

Methods, apparatuses, and computer program products for responding to a timeout of a message in a parallel computer are provided. The parallel computer includes a plurality of compute nodes operatively coupled for data communications over one or more data communications networks. Each compute node includes one or more tasks. Embodiments include a first task of a first node sending a message to a second task on a second node. Embodiments also include the first task sending to the second node a command via a parallel operating environment (POE) in response to a timeout of the message. The command instructs the second node to perform a timeout motivated operation.

Responding to a timeout of a message in a parallel computer

A job may be divided into multiple tasks that may execute in parallel on one or more compute nodes. The tasks executing on the same compute node may be coordinated using barrier synchronization. However, to perform barrier synchronization, the tasks use (or attach) to a barrier synchronization register which establishes a common checkpoint for each of the tasks. A leader task may use a shared memory region to publish to follower tasks the location of the barrier synchronization register&#x2014;i.e., a barrier synchronization register ID. The follower tasks may then monitor the shared memory to determine the barrier synchronization register ID. The leader task may also use a count to ensure all the tasks attach to the BSR. This advantageously avoids any task-to-task communication which may reduce overhead and improve performance.

Preparing parallel tasks to use a synchronization register

Acknowledging incoming messages, including: determining, by an acknowledgement dispatching module, whether an incoming message has been received in an active message queue; responsive to determining that the incoming message has been received in the active message queue, resetting, by the acknowledgement dispatching module, an acknowledgment iteration counter; incrementing, by the acknowledgement dispatching module, the acknowledgment iteration counter; determining, by the acknowledgement dispatching module, whether the acknowledgment iteration counter has reached a predetermined threshold; and responsive to determining that the acknowledgment iteration counter has reached the predetermined threshold, processing, by the acknowledgement dispatching module, all messages in the active message queue.

Acknowledging incoming messages

A mechanism is provided in a collective acceleration unit for performing a collective operation to distribute or collect data among a plurality of participant nodes. The mechanism receives an input collective packet for a collective operation from a neighbor node within a collective tree. The input collective packet comprises a tree identifier and an input data field and wherein the collective tree comprises a plurality of sub trees. The mechanism maps the tree identifier to an index within the collective acceleration unit. The index identifies a portion of resources within the collective acceleration unit and is associated with a set of neighbor nodes in a given sub tree within the collective tree. For each neighbor node the collective acceleration unit stores destination information. The collective acceleration unit performs an operation on the input data field using the portion of resources to effect the collective operation.

Collective acceleration unit tree structure

Growing community of inventors

Poughkeepsie, NY, United States of America

Hanhong Xue