Cray Research, Inc.

Silicon Graphics, Incorporated

Hewlett Packard Enterprise Development LP

Silicon Graphics International Corporation

Sgi International, Inc.

Randal Steven Passint

Greg Thorson

Steven M Oberlin

Gregory Michael Thorson

Eric Carl Fromm

Donglai Dai

Michael Brian Galles

Martin M Deneroff

Mark S Birrittella

Mark Ronald Sikkink

Richard E Kessler

Steven Lee Scott

William A Huffman

Michael S Woodacre

Joseph Martin Placek

John B Carter

Russell Leonard Nicol

Lixin Zhang

Steven J Dean

Thomas Edward McGee

Kirill Malkin

Michael E Malewicki

Liqun Cheng

Douglas M Pase

Nan Ma

Vernon W Swanson

Janet M Eberhart

Thomas A MacDonald

Zhen Fang

Timothy Stremcha

Gary W Elsesser

Various embodiments improve the operation of computers by providing methods of transmitting data with low latency and high bandwidth. Data may be transmitted in a packet composed of data flits, the data flits having at least two different formats configured to implement different communication protocols. In some embodiments, a given flit may be transmitted using two different modulation methods, with a first part of the flit transmitted using a first modulation method, such as a binary method, and a second part of the flit using a higher-order modulation method.

Method and system for high speed data links

A high performance computing (HPC) system has an architecture that separates data paths used by compute nodes exchanging computational data from the data paths used by compute nodes to obtain computational work units and save completed computations. The system enables an improved method of saving checkpoint data, and an improved method of using an analysis of the saved data to assign particular computational work units to particular compute nodes. The system includes a compute fabric and compute nodes that cooperatively perform a computation by mutual communication using the compute fabric. The system also includes a local data fabric that is coupled to the compute nodes, a memory, and a data node. The data node is configured to retrieve data for the computation from an external bulk data storage, and to store its work units in the memory for access by the compute nodes.

Workload management system and process

An apparatus and method exchange data between two nodes of a high performance computing (HPC) system using a data communication link. The apparatus has one or more processing cores, RDMA engines, cache coherence engines, and multiplexers. The multiplexers may be programmed by a user application, for example through an API, to selectively couple either the RDMA engines, cache coherence engines, or a mix of these to the data communication link. Bulk data transfer to the nodes of the HPC system may be performed using paged RDMA during initialization. Then, during computation proper, random access to remote data may be performed using a coherence protocol (e.g. MESI) that operates on much smaller cache lines.

Caching network fabric for high performance computing

A system and method for interconnecting a plurality of processing element nodes within a scalable multiprocessor system is provided. Each processing element node includes at least one processor and memory. A scalable interconnect network includes physical communication links interconnecting the processing element nodes in a cluster. A first set of routers in the scalable interconnect network route messages between the plurality of processing element nodes. One or more metarouters in the scalable interconnect network route messages between the first set of routers so that each one of the routers in a first cluster is connected to all other clusters through one or more metarouters.

Network topology for a scalable multiprocessor system

A method for maintaining data coherency in a shared-memory computer system having a plurality of nodes divides the local memory of a given node into one or more blocks and stores a data record for each block indicating a plurality of node groups and a selection of the node groups. Each selected node group represents a number of nodes, and selected node groups represent at least one node that has requested access to the block. In response to receiving an access request from a requesting node that may or may not be in a selected node group, the method and system update the data record to indicate the correct selection. If the requesting node is not in any node group, the data record is adjusted to have new node groups, one of which represents the requesting node.

Data coherence method and apparatus for multi-node computer system

A method and apparatus for maintaining data coherency in a computer system having a plurality of nodes forms a directory by grouping the nodes into a plurality of hierarchical groups of two or more levels. The method and apparatus also 1) set the directory to have data relating to a first set of groups within a first level, and 2) determine if a requesting node requesting data is a member of one of the first set of groups. The directory then is set to have data relating to a second group of nodes if the requesting node is determined not to be a member of the first set of groups within the first level. The second group of nodes is in a higher level than the first level.

A method and apparatus for managing X4 or larger types of memory first receives a data word to be stored in the memory, and then generates a check datum, which is a function of the data word and a set of encode data. After storing the data word in memory, the method and apparatus use the check datum and the data word to generate a syndrome. The method and apparatus then determine if the data word in the memory is correct as a function of the syndrome.

Fault tolerant memory system

A method and apparatus for controlling access by a set of accessing nodes to memory of a home node (in a multimode computer system) determines that each node in the set of nodes has accessed the memory, and forwards a completion message to each node in the set of nodes after it is determined that each node has accessed the memory. The completion message has data indicating that each node in the set of nodes has accessed the memory of the home node.

Node synchronization for multi-processor computer systems

An apparatus and method of controlling data sharing in a shared memory computer system transfers control of a cache coherency directory (entry) to a node having control of the data. Specifically, the plurality of nodes includes a home node and a second node. The home node has given data in a cache line in its memory, and also has a directory identifying the state of the cache line. The method and apparatus thus detect a request for ownership of the cache line from the second node, and enable the second node to control the directory after receipt of the request.

Apparatus and method of controlling data sharing on a shared memory computer system

The present invention provides an interface and method for synchronous data transfer between domains clocked at different frequencies. The interface includes a first latch for receiving data from a first domain clocked at one frequency when the first latch is selected and a second latch for receiving data from the first domain when the second latch is selected. A third latch is provided for transferring data from either the first latch or the second latch to the second domain when the second domain is clocked.

Interface for synchronous data transfer between domains clocked at different frequencies

Scalable hypercube multiprocessor network for massive parallel processing

A method and system for controlling an access to a first memory arrangement and a second memory arrangement. The method and system are adapted for controlling access to the first memory arrangement and to the second memory arrangement. A token is passed from a device associated with the first memory arrangement if the access to at least one portion of the first memory arrangement is completed, and the access to the portion of the memory arrangement is disabled. Then, upon a receipt of the token, the access to at least one portion of the second memory arrangement is enabled.

Method and system for controlling data access between at least two memory arrangements

In a multiprocessor system having a plurality of nodes connected to a network, wherein communication between the plurality of nodes is in the form of packets, a system and method of aging packets. A packet having an age value is built and transmitted through the network. The age value is increased at predetermined intervals, wherein increasing includes determining a current age of the packet and changing the interval as a function of the current age. A method of avoiding livelock and a method of preaging response packets is also described.

Age-based network arbitration system and method

Growing community of inventors

Chippewa Falls, WI, United States of America

Randal Steven Passint