International Business Machines Corporation

Globalfoundries Inc.

Sameh Sherif Sharkawi

Charles J Archer

Nysal Jan K A

William P LePera

Austen William Lauria

Robert H Bell, Jr

Venkat Rajeev Indukuru

Donald R DeSota

Rajendra D Panda

Scott Miller

Nysal Jan Ka

Jason Frederick Cantin

Luigi Brochard

Sameer Kumar

Bryan Savoye Rosenburg

Joseph H Robichaux

An embodiment includes mapping, responsive to receiving a request for a container image from a container host, the requested container image to a first computer memory on a registry server. The embodiment also includes exposing a window storing the mapped container image to the container host using a collective window-creation call with the container host. The embodiment also includes processing a Remote Direct Memory Access (RDMA) data transfer request to select a lock type for the window during the RDMA data transfer. The embodiment also includes imposing the selected lock type on the window during the RDMA data transfer. The embodiment also includes releasing the selected lock type from the window upon detecting completion of the RDMA data transfer.

Remote direct memory access for container-enabled networks

On a first compute resource, execution of a first task is triggered, execution of a portion of the first task being conditioned on a second task executing on a second compute resource. A state indicator of the second task is monitored, the state indicator indicating whether or not the second task is currently executing on the second compute resource. Responsive to the state indicator indicating that the second task is not currently executing, execution of the portion of the first task is suspended. A change in the state indicator is determined to have occurred. Responsive to the determining, received connection information for the second task is forwarded to the first task. Execution of the portion of the first task is re-triggered on the first compute resource.

Parallel task initialization on dynamic compute resources

A computer-implemented method and system for inter-processor communications fault handling in high performance computing networks. The method includes detecting that an InfiniBand (IB) queue pair has transitioned into an error state based on an unsuccessful completion status that relates to unsuccessful delivery of a message from an initiator endpoint at a first server device to at least one target endpoint at a second server device. The initiator and target endpoints are associated with at least one application under execution. An embodiment includes inferring, when the unsuccessful completion status is indicated as flushed, that the message was in a send queue of the IB queue pair when the IB queue pair transitioned into the error state. An embodiment includes establishing an IB Direct Connect queue pair connection between the target and initiator endpoints. An embodiment includes re-queueing the message in the IB queue pair for dispatch to the target endpoint.

Inter-processor communications fault handling in high performance computing networks

A computer implemented method uses memory coherence to enhance latency and bandwidth performance, the method including receiving, by a host, a call from an application. The method also includes, determining that the call includes a device allocation command, wherein the device allocation command is configured to allocate a set of data on a graphical processing unit. The method further includes intercepting the call. The method includes, initiating an alternate data allocation command; and returning the alternate data allocation command to the application. Further aspects of the present disclosure are directed to systems and computer program products containing functionality consistent with the method described above.

Utilizing memory coherency to improve bandwidth performance

Technology for determining whether an inter-process type message has been successfully sent from a first process to a second process running on a single computer with a single processor(s) set. A variable (for example, a bit value) is used to indicate whether the inter-process message has been communicated between the processes. A timer and a predetermined timeout threshold are used to determine if the inter-process message has been pending for too long without being successfully communicated.

Detection of shared memory faults in a computing job

An embodiment includes sending, via a queue pair (QP) at a first endpoint, a message to a second endpoint. The embodiment also includes detecting an error state of the QP caused by a failure at a third endpoint that automatically halts messages via the QP. The embodiment includes determining that communication between the first and second endpoints via the QP is viable, and placing messages to the second endpoint in a separate queue from messages to an unreachable endpoint. The embodiment also includes re-establishing communications between the first and second endpoints via the QP, and polling the second endpoint for an indication of a delivered message. Any messages indicated as having already been delivered are removed from the queue before re-starting communications with the second endpoint.

Error state message management

Techniques are disclosed for efficient handling of messages in computing systems that include tag matching capable hardware. A message management module provides for handling message events including application receives and channel notifications such that hardware tag matching can continuously run in hardware channels, such as network adapters. When the message event is an application receive the message management module adds the application receive to a tracking queue and determines if the application receive can be posted to a hardware channel capable of tag matching. When the message event is a channel notification, the message management module determines a message action using the message tracking queue and the information in the channel notification.

Messaging software management for network devices supporting hardware tag matching

Performing a rooted-v collective operation by an operational group of compute nodes in a parallel computer includes: upon encountering a rooted-v collection operation during execution, identifying, by a root node of an operational group of compute nodes, a count to use for the selection of a collective algorithm for effecting the rooted-v collective operation; broadcasting, by the root node to the other computer nodes in the operational group, an active message, wherein the active message includes the identified count to use for the selection of the collective algorithm; and selecting, by all the compute nodes of the operational group based on the identified count, a same collective algorithm to effect the rooted-v collective operation; and executing the rooted-v collective operation by all compute nodes of the operational group using the selected algorithm.

Performing a rooted-v collective operation by an operational group of compute nodes in a parallel computer

Executing an all-to-allv operation on a parallel computer that includes a plurality of compute nodes, including: packing, by each task in an operational group of tasks, vectored contribution data from vectored storage in an all-to-allv contribution data buffer into an all-to-all contribution data buffer, wherein two or more entries in the all-to-allv contribution data buffer are different in size and each entry in the all-to-all contribution data buffer is identical in size; executing with the contribution data as stored in the all-to-all contribution data buffer an all-to-all collective operation by the operational group of tasks; and unpacking, by each task in the operational group of tasks, received contribution data from the all-to-all contribution data buffer into the vectored storage in an all-to-allv contribution data buffer.

Executing an all-to-allv operation on a parallel computer that includes a plurality of compute nodes

Adjusting environmental variables in an adaptive parameter adjustment runtime environment, including: executing a parallel program by the adaptive parameter adjustment runtime environment, including beginning operations with a set of default global parameter values; maintaining a list of configurable parameters; changing a parameter value for a parameter in the list of configurable parameters; determining whether an effect of changing the parameter value is positive, negative, or neutral; responsive to determining that the effect of changing the parameter value is positive, changing the parameter value for the parameter; responsive to determining that the effect of changing the parameter value is negative, changing the parameter value for the parameter to a previous value; and responsive to determining that the effect of changing the parameter value is neutral, performing a list management operation on the list of configurable parameters.

Adjusting environmental variables in an adaptive parameter adjustment runtime environment

Constructing a logical tree topology in a parallel computer that includes compute nodes, where each compute node includes a hardware acceleration unit and executes an identical number of tasks and the tasks of each node have a rank, includes: creating hardware acceleration groups, with each hardware acceleration group including one task from each node, where the one task from each node has the same rank; assigning one task of a root compute node as a global root of the logical tree topology; assigning tasks of the root compute node other than the global root as local children of the global root; and assigning each of the global root and local children of the root compute node as a root of a subtree of tasks, wherein each subtree comprises the tasks of a hardware acceleration group.

Constructing a logical tree topology in a parallel computer

In a parallel computer, performing a rooted-v collective operation by an operational group of compute nodes includes: identifying, in source code by a collective algorithm selection optimizing module, a gather operation followed by a rooted-v collective operation; replacing, by the collective algorithm selection optimizing module, the gather operation with an allgather operation; executing, by the compute nodes, the allgather operation; selecting, by each compute node in dependence upon results of the allgather operation, an algorithm for effecting the rooted-v collective operation; and executing, by each compute node, the rooted-v collective operation with the selected algorithm.

Growing community of inventors

Cypress, TX, United States of America

Sameh Sherif Sharkawi