Abstract: A target task ensures complete delivery of a global shared memory (GSM) message from an originating task to the target task. The target task's HFI receives a first of multiple GSM packets generated from a single GSM message sent from the originating task. The HFI logic assigns a sequence number and corresponding tuple to track receipt of the complete GSM message. The sequence number is unique relative to other sequence numbers assigned to GSM messages that have not been completely received from the initiating task. The HFI updates a count value within the tuple, which comprises the sequence number and the count value for the first GSM packet and for each subsequent GSM packet received for the GSM message. The HFI determines when receipt of the GSM message is complete by comparing the count value with a count total retrieved from the packet header.

Abstract: A system is provided for implementing a multi-tiered full-graph interconnect architecture. In order to implement a multi-tiered full-graph interconnect architecture, a plurality of processors are coupled to one another to create a plurality of processor books. The plurality of processor books are coupled together to create a plurality of supernodes. Then, the plurality of supernodes are coupled together to create the multi-tiered full-graph interconnect architecture. Data is then transmitted from one processor to another within the multi-tiered full-graph interconnect architecture based on an addressing scheme that specifies at least a supernode and a processor book associated with a target processor to which the data is to be transmitted.

Abstract: Mechanisms for providing hardware based dynamic load balancing of message passing interface (MPI) tasks are provided. Mechanisms for adjusting the balance of processing workloads of the processors executing tasks of an MPI job are provided so as to minimize wait periods for waiting for all of the processors to call a synchronization operation. Each processor has an associated hardware implemented MPI load balancing controller. The MPI load balancing controller maintains a history that provides a profile of the tasks with regard to their calls to synchronization operations. From this information, it can be determined which processors should have their processing loads lightened and which processors are able to handle additional processing loads without significantly negatively affecting the overall operation of the parallel execution system. As a result, operations may be performed to shift workloads from the slowest processor to one or more of the faster processors.

Abstract: According to at least one embodiment, a method of data processing in a multiprocessor data processing system includes a requesting processing unit initiating an interconnect operation including a memory access request that indicates an acceptability of a variable amount of data to service the interconnect request for data. In response to snooping the memory access request on an interconnect, a snooper selects an amount of data to supply to the requesting processing unit and transmits the selected amount of data to the requesting processing unit. The requesting processing unit receives the selected amount of data and utilizes at least some of the selected amount of data to service a processor request.

Abstract: A method of data processing in a cache memory includes caching a plurality of cache lines of data in a corresponding plurality of entries in a cache array, where each of the plurality of cache lines includes multiple data granules. For each of the plurality of cache entries, a plurality of line coherency state fields indicates an associated coherency state applicable to two or more data granules. For at least a particular cache line among the plurality of cache lines, a granule coherency state field indicates a coherency state for a particular granule of the multiple data granules in the particular cache line, where the coherency state field indicated by the granule coherency state field differs from that indicated for the particular cache line by its line coherency state field.

Abstract: A mechanism is provided for packet coalescing in virtual channels of a data processing system. A first processor bundles original data into a data packet to be transmitted to a destination processor, the original data comprising payload data and overhead data. The first processor transmits the data packet to a second processor along a path to the destination processor. The second processor determines if the second processor has additional payload data destined for the same destination processor. Responsive to the second processor having the additional payload data, the second processor unbundles the data packet, adds the additional payload data to the payload data, and rebundles the payload data along with the additional payload data and the overhead data into a rebundled data packet. Then the second processor transmits the rebundled data packet to at least one other processor along the path to the destination processor.

Abstract: Mechanisms for modifying an operation of one or more processors executing message passing interface (MPI) tasks are provided. Mechanisms for adjusting the balance of processing work loads of the processors are provided so as to minimize wait periods for waiting for all of the processors to call a synchronization operation. Each processor has an associated hardware implemented MPI load balancing controller. The MPI Load balancing controller maintains a history that provides a profile of the tasks with regard to their calls to synchronization operations. From this information, it can be determined which processors should have their processing loads lightened and which processors are able to handle additional processing loads without significantly negatively affecting the overall operation of the parallel execution system. As a result, operations may be performed to shift workloads from the slowest processor to one or more of the faster processors.

Abstract: Mechanisms for performing dynamic request routing based on broadcast depth queue information are provided. Each processor chip in the system may use a synchronized heartbeat signal it generates to provide queue depth information to each of the other processor chips in the system. The queue depth information identifies a number of requests or amount of data in each of the queues of a processor chip that originated the heartbeat signal. The queue depth information from each of the processor chips in the system may be used by the processor chips in determining optimal routing paths for data from a source processor chip to a destination processor chip. As a result, the congestion of data for processing at each of the processor chips along each possible routing path may be taken into account when selecting to which processor chip to forward data.

Abstract: 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.

Abstract: According to a method of data processing in a multiprocessor data processing system, in response to a processor request to read a target granule of a target cache line of data containing multiple granules, a processing unit originates on an interconnect of the multiprocessor data processing system a partial read request that requests permission to read only the target granule of the target cache line. In response to a combined response to the partial read request indicating success, the combined response representing a system-wide response to the partial read request, the processing unit receives the target granule of the target cache line, supplies the target granule to a requesting processor core, and updates a coherency state of the target granule while retaining a coherency state of at least one other granule of the target cache line.

Abstract: A mechanism is provided for transmitting data from a first processor of a data processing system to a second processor of the data processing system. In one or more switches, a set of virtual channels is created, the one or more switches comprising, for each processor, a corresponding switch in the one or more switches. The data is transmitted from the first processor to the second processor through a path comprising a subset of processors of a set of processors in the data processing system. In each processor of the subset of processors, the data is stored in a virtual channel of a corresponding switch before transmitting the data to a next processor. The virtual channel of the corresponding switch in which the data is stored corresponds to a position of the processor in the path through which the data is transmitted.

Abstract: A distributed data processing system executes multiple tasks within a parallel job, including a first local task on a local node and at least one task executing on a remote node, with a remote memory having real address (RA) locations mapped to one or more of the source effective addresses (EA) and destination EA of a data move operation initiated by a task executing on the local node. On initiation of the data move operation, remote asynchronous data move (RADM) logic identifies that the operation moves data to/from a first EA that is memory mapped to an RA of the remote memory. The local processor/RADM logic initiates a RADM operation that moves a copy of the data directly from/to the first remote memory by completing the RADM operation using the network interface cards (NICs) of the source and destination processing nodes, determined by accessing a data center for the node IDs of remote memory.

Abstract: A mechanism is provided for collective acceleration unit tree flow control forms a logical tree (sub-network) among those processors and transfers “collective” packets on this tree. The system supports many collective trees, and each collective acceleration unit (CAU) includes resources to support a subset of the trees. Each CAU has limited buffer space, and the connection between two CAUs is not completely reliable. Therefore, to address the challenge of collective packets traversing on the tree without colliding with each other for buffer space and guaranteeing the end-to-end packet delivery, each CAU in the system effectively flow controls the packets, detects packet loss, and retransmits lost packets.

Abstract: A mechanism for performing collective operations. In software executing on a parent processor in a first processor book, a number of other processors are determined in a same or different processor book of the data processing system that is needed to execute the collective operation, thereby establishing a plurality of processors comprising the parent processor and the other processors. In software executing on the parent processor, the plurality of processors are logically arranged as a plurality of nodes in a hierarchical structure. The collective operation is transmitted to the plurality of processors based on the hierarchical structure. In hardware of the parent processor, results are received from the execution of the collective operation from the other processors, a final result is generated of the collective operation based on the received results, and the final result is output.

Abstract: A method of data processing in a processing unit supported by a memory hierarchy includes the processing unit performing a plurality of memory accesses to the memory hierarchy. The plurality of memory accesses includes one or more memory accesses targeting a full cache line of data. The processing unit monitors utilization of data accessed by the plurality of memory accesses, and based upon the utilization of the data, dynamically alters a memory access mode of operation so that a subsequent storage-modifying memory access targets less than a full cache line of data.

Abstract: A mechanism is provided for performing collective operations. In hardware of a parent processor in a first processor book, a number of other processors are determined in a same or different processor book of the data processing system that is needed to execute the collective operation, thereby establishing a plurality of processors comprising the parent processor and the other processors. In hardware of the parent processor, the plurality of processors are logically arranged as a plurality of nodes in a hierarchical structure. The collective operation is transmitted to the plurality of processors based on the hierarchical structure. In hardware of the parent processor, results are received from the execution of the collective operation from the other processors, a final result is generated of the collective operation based on the received results, and the final result is output.

Abstract: Mechanisms for providing a cluster-wide system clock in a multi-tiered full graph (MTFG) interconnect architecture are provided. Heartbeat signals transmitted by each of the processor chips in the computing cluster are synchronized. Internal system clock signals are generated in each of the processor chips based on the synchronized heartbeat signals. As a result, the internal system clock signals of each of the processor chips are synchronized since the heartbeat signals, that are the basis for the internal system clock signals, are synchronized. Mechanisms are provided for performing such synchronization using direct couplings of processor chips within the same processor book, different processor books in the same supernode, and different processor books in different supernodes of the MTFG interconnect architecture.

Abstract: A mechanism is provided for routing information through the data processing system. Data is received at a source processor within a set of processors that is to be transmitted to a destination processor, where the data includes address information. A first determination is performed as to whether the destination processor is within a same processor book as the source processor based on the address information. A second determination is performed as to whether the destination processor is within a same supernode as the source processor based on the address information if the destination processor is not within the same processor book. A routing path is identified for the data based on results of the first determination, the second determination, and one or more routing table data structures. The data is then transmitted from the source processor along the identified routing path toward the destination processor.

Abstract: A method and a data processing system for completing checkpoint processing of a distributed job with local tasks communicating with other remote tasks via a host fabric interface (HFI) and assigned HFI window. Each HFI window has a send count and a receive count, which tracks GSM messages that are sent from and received at the HFI window. When a checkpoint is initiated by a master task, each local task forwards the send count and the receive count to the master task. The master task sums the respective counts and then compares the totals to each other. When the send count total is equal to the receive count total, the tasks are permitted to continue processing. However, when the send count total is not equal to the receive count total, the master task notifies each task of the job to rollback to a previous checkpoint or kill the job execution.

Abstract: A host fabric interface (HFI) enables debugging of global shared memory (GSM) operations received at a local node from a network fabric. The local node has a memory management unit (MMU), which provides an effective address to real address (EA-to-RA) translation table that is utilized by the HFI to evaluate when EAs of GSM operations/data from a received GSM packet is memory-mapped to RAs of the local memory. The HFI retrieves the EA associated with a GSM operation/data within a received GSM packet. The HFI forwards the EA to the MMU, which determines when the EA is mapped to RAs within the local memory for the local task. The HFI processing logic enables processing of the GSM packet only when the EA of the GSM operation/data within the GSM packet is an EA that has a local RA translation. Non-matching EAs result in an error condition that requires debugging.