Abstract: Compute nodes perform allreduce operations that swap processes at nodes. A first allreduce operation generates a first result and uses a first process from a first compute node, a second process from a second compute node, and zeros from other compute nodes. The first compute node replaces the first process with the first result. A second allreduce operation generates a second result and uses the first result from the first compute node, the second process from the second compute node, and zeros from others. The second compute node replaces the second process with the second result, which is the first process. A third allreduce operation generates a third result and uses the first result from first compute node, the second result from the second compute node, and zeros from others. The first compute node replaces the first result with the third result, which is the second process.

Abstract: An apparatus, program product and method checks for nodal faults in a group of nodes comprising a center node and all adjacent nodes. The center node concurrently communicates with the immediately adjacent nodes in three dimensions. The communications are analyzed to determine a presence of a faulty node or connection.

Abstract: A massively parallel computer system contains an inter-nodal communications network of node-to-node links. An automated routing strategy routes packets through one or more intermediate nodes of the network to reach a final destination. The default routing strategy is altered responsive to detection of overutilization of a particular path of one or more links, and at least some traffic is re-routed by distributing the traffic among multiple paths (which may include the default path). An alternative path may require a greater number of link traversals to reach the destination node.

Abstract: A massively parallel computer system contains an inter-nodal communications network of node-to-node links. Each node implements a respective routing strategy for routing data through the network, the routing strategies not necessarily being the same in every node. The routing strategies implemented in the nodes are dynamically adjusted during application execution to shift network workload as required. Preferably, adjustment of routing policies in selective nodes is performed at synchronization points. The network may be dynamically monitored, and routing strategies adjusted according to detected network conditions.

Abstract: An apparatus and program product check for nodal faults in a row of nodes by causing each node in the row to concurrently communicate with its adjacent neighbor nodes in the row. The communications are analyzed to determine a presence of a faulty node or connection.

Abstract: A method and apparatus for a template based parallel checkpoint save for a massively parallel super computer system using a parallel variation of the rsync protocol, and network broadcast. In preferred embodiments, the checkpoint data for each node is compared to a template checkpoint file that resides in the storage and that was previously produced. Embodiments herein greatly decrease the amount of data that must be transmitted and stored for faster checkpointing and increased efficiency of the computer system. Embodiments are directed to a parallel computer system with nodes arranged in a cluster with a high speed interconnect that can perform broadcast communication. The checkpoint contains a set of actual small data blocks with their corresponding checksums from all nodes in the system. The data blocks may be compressed using conventional non-lossy data compression algorithms to further reduce the overall checkpoint size.

Abstract: An apparatus and program product logically divide a group of nodes and causes node pairs comprising a node from each section to communicate. Results from the communications may be analyzed to determine performance characteristics, such as bandwidth and proper connectivity.

Abstract: A method and apparatus provide data security on a parallel computer system using virtual private networks. An access setup mechanism sets up access control data in the nodes that describes which virtual networks are protected and what applications have access to the protected private networks. When an application accesses data on a protected virtual network, a network access mechanism determines the data is protected and intercepts the data access. The network access mechanism in the kernel may also execute a rule depending on the kind of access that was attempted to the virtual network. Authorized access to the private networks can be made via a system call to the access control mechanism in the kernel. The access control mechanism enforces policy decisions on which data can be distributed through the system via an access control list or other security policies.

Abstract: Embodiments of the invention provide a method for deploying and running an application on a massively parallel computer system, while minimizing the costs associated with latency, bandwidth, and limited memory resources. The executable code of a program may be divided into multiple code fragments and distributed to different compute nodes of a parallel computing system. During program execution, one compute node may fetch code fragments from other compute nodes as necessary.

Abstract: A method determines a nodal fault along the boundary, or face, of a computing cell. Nodes on adjacent cell boundaries communicate with each other, and the communications are analyzed to determine if a node or connection is faulty.

Abstract: An apparatus, program product and method checks for nodal faults in a group of nodes comprising a center node and all adjacent nodes. The center node concurrently communicates with the immediately adjacent nodes in three dimensions. The communications are analyzed to determine a presence of a faulty node or connection.

Abstract: Embodiments off the invention provide a mechanism for process migration on a massively parallel computer system. In particular, embodiments of the invention may be used to update process state data for a migrated compute node, such as MPI (or other communication library) state data, across a full collection of compute nodes present in a given parallel system executing a parallel task. Migrating a process form one compute node to another may be useful to address a variety of sub-optimal operating conditions. For example, one or more processes may be migrated to cure network congestion resulting from a poorly mapped task or when a compute node is predicted to experience a hardware failure.

Abstract: An apparatus and program product determine a nodal fault along the boundary, or face, of a computing cell. Nodes on adjacent cell boundaries communicate with each other, and the communications are analyzed to determine if a node or connection is faulty.

Abstract: An apparatus and method retrieves a database record from an in-memory database of a parallel computer system using a unique key. The parallel computer system performs a simultaneous search on each node of the computer system using the unique key and then utilizes a global combining network to combine the results from the searches of each node to efficiently and quickly search the entire database.

Abstract: An apparatus and method retrieves a database record from an in-memory database of a parallel computer system using a non-unique key. The parallel computer system performs a simultaneous search on each node of the computer system using the non-unique key and then utilizes a global combining network to combine the local results from the searches of each node to efficiently and quickly search the entire database.

Abstract: A method and apparatus for a template based parallel checkpoint save for a massively parallel super computer system using a parallel variation of the rsync protocol, and network broadcast. In preferred embodiments, the checkpoint data for each node is compared to a template checkpoint file that resides in the storage and that was previously produced. Embodiments herein greatly decrease the amount of data that must be transmitted and stored for faster checkpointing and increased efficiency of the computer system. Embodiments are directed to a parallel computer system with nodes arranged in a cluster with a high speed interconnect that can perform broadcast communication. The checkpoint contains a set of actual small data blocks with their corresponding checksums from all nodes in the system. The data blocks may be compressed using conventional non-lossy data compression algorithms to further reduce the overall checkpoint size.

Abstract: A method and apparatus for a template based parallel checkpoint save for a massively parallel super computer system using a parallel variation of the rsync protocol, and network broadcast. In preferred embodiments, the checkpoint data for each node is compared to a template checkpoint file that resides in the storage and that was previously produced. Embodiments herein greatly decrease the amount of data that must be transmitted and stored for faster checkpointing and increased efficiency of the computer system. Embodiments are directed to a parallel computer system with nodes arranged in a cluster with a high speed interconnect that can perform broadcast communication. The checkpoint contains a set of actual small data blocks with their corresponding checksums from all nodes in the system. The data blocks may be compressed using conventional non-lossy data compression algorithms to further reduce the overall checkpoint size.

Abstract: An apparatus and program product check for nodal faults in a row of nodes by causing each node in the row to concurrently communicate with its adjacent neighbor nodes in the row. The communications are analyzed to determine a presence of a faulty node or connection.

Abstract: An apparatus, program product and method check for nodal faults in a row of nodes by causing each node in the row to concurrently communicate with its adjacent neighbor nodes in the row. The communications are analyzed to determine a presence of a faulty node or connection.

Abstract: An apparatus and program product logically divide a group of nodes and causes node pairs comprising a node from each section to communicate. Results from the communications may be analyzed to determine performance characteristics, such as bandwidth and proper connectivity.