International Business Machines Corporation

Intel Corporation

Pavlos Michael Vranas

Alan G Gara

Philip Heidelberger

Dong Chen

Mark Edwin Giampapa

Matthias Augustin Blumrich

Dirk Hoenicke

Burkhard D Steinmacher-Burow

Martin Ohmacht

Paul William Coteus

Todd Edward Takken

Valentina Salapura

Burkhard Steinmacher-Burow

Dong Chen Chen

Gyan Vatsaraja Bhanot

Sameer Kumar

Gerard Vincent Kopcsay

Jeffrey John Parker

Wayne Melvin Barrett

Gheorghe Almasi

Thomas A Liebsch

Rory D Jackson

Ben Jesse Nathanson

Phillip Heidelberger

Sarabjeet Singh

Matthew A Blumrich

A system and method for enabling high-speed, low-latency global collective communications among interconnected processing nodes. The global collective network optimally enables collective reduction operations to be performed during parallel algorithm operations executing in a computer structure having a plurality of the interconnected processing nodes. Router devices are included that interconnect the nodes of the network via links to facilitate performance of low-latency global processing operations at nodes of the virtual network and class structures. The global collective network may be configured to provide global barrier and interrupt functionality in asynchronous or synchronized manner. When implemented in a massively-parallel supercomputing structure, the global collective network is physically and logically partitionable according to needs of a processing algorithm.

Collective network for computer structures

An error-recovery method to enable error-free message transfer between nodes of a computer network. A first node of the network sends a packet to a second node of the network over a link between the nodes, and the first node keeps a copy of the packet on a sending end of the link until the first node receives acknowledgment from the second node that the packet was received without error. The second node tests the packet to determine if the packet is error free. If the packet is not error free, the second node sets a flag to mark the packet as corrupt. The second node returns acknowledgement to the first node specifying whether the packet was received with or without error. When the packet is received with error, the link is returned to a known state and the packet is sent again to the second node.

Error recovery to enable error-free message transfer between nodes of a computer network

An apparatus and method for tracking coherence event signals transmitted in a multiprocessor system. The apparatus comprises a coherence logic unit, each unit having a plurality of queue structures with each queue structure associated with a respective sender of event signals transmitted in the system. A timing circuit associated with a queue structure controls enqueuing and dequeuing of received coherence event signals, and, a counter tracks a number of coherence event signals remaining enqueued in the queue structure and dequeued since receipt of a timestamp signal. A counter mechanism generates an output signal indicating that all of the coherence event signals present in the queue structure at the time of receipt of the timestamp signal have been dequeued. In one embodiment, the timestamp signal is asserted at the start of a memory synchronization operation and, the output signal indicates that all coherence events present when the timestamp signal was asserted have completed. This signal can then be used as part of the completion condition for the memory synchronization operation.

Method and apparatus for efficiently tracking queue entries relative to a timestamp

A method and apparatus for supporting cache coherency in a multiprocessor computing environment having multiple processing units, each processing unit having one or more local cache memories associated and operatively connected therewith. The method comprises providing a snoop filter device associated with each processing unit, each snoop filter device having a plurality of dedicated input ports for receiving snoop requests from dedicated memory writing sources in the multiprocessor computing environment. Each snoop filter device includes a plurality of parallel operating port snoop filters in correspondence with the plurality of dedicated input ports, each port snoop filter implementing one or more parallel operating sub-filter elements that are adapted to concurrently filter snoop requests received from respective dedicated memory writing sources and forward a subset of those requests to its associated processing unit.

Snoop filter for filtering snoop requests

A system and method for enabling high-speed, low-latency global collective communications among interconnected processing nodes. The global collective network optimally enables collective reduction operations to be performed during parallel algorithm operations executing in a computer structure having a plurality of the interconnected processing nodes. Router devices are included that interconnect the nodes of the network via links to facilitate performance of low-latency global processing operations at nodes of the virtual network. The global collective network may be configured to provide global barrier and interrupt functionality in asynchronous or synchronized manner. When implemented in a massively-parallel supercomputing structure, the global collective network is physically and logically partitionable according to needs of a processing algorithm.

Method and system for hardware packet pacing using a direct memory access controller in a parallel computer which, in one aspect, keeps track of a total number of bytes put on the network as a result of a remote get operation, using a hardware token counter.

Hardware packet pacing using a DMA in a parallel computer

A computing apparatus for reducing the amount of processing in a network computing system which includes a network system device of a receiving node for receiving electronic messages comprising data. The electronic messages are transmitted from a sending node. The network system device determines when more data of a specific electronic message is being transmitted. A memory device stores the electronic message data and communicating with the network system device. A memory subsystem communicates with the memory device. The memory subsystem stores a portion of the electronic message when more data of the specific message will be received, and the buffer combines the portion with later received data and moves the data to the memory device for accessible storage.

Extended write combining using a write continuation hint flag

A system and method for supporting cache coherency in a computing environment having multiple processing units, each unit having an associated cache memory system operatively coupled therewith. The system includes a plurality of interconnected snoop filter units, each snoop filter unit corresponding to and in communication with a respective processing unit, with each snoop filter unit comprising a plurality of devices for receiving asynchronous snoop requests from respective memory writing sources in the computing environment; and a point-to-point interconnect comprising communication links for directly connecting memory writing sources to corresponding receiving devices; and, a plurality of parallel operating filter devices coupled in one-to-one correspondence with each receiving device for processing snoop requests received thereat and one of forwarding requests or preventing forwarding of requests to its associated processing unit. Each of the plurality of parallel operating filter devices comprises parallel operating sub-filter elements, each simultaneously receiving an identical snoop request and implementing one or more different snoop filter algorithms for determining those snoop requests for data that are determined not cached locally at the associated processing unit and preventing forwarding of those requests to the processor unit. In this manner, a number of snoop requests forwarded to a processing unit is reduced thereby increasing performance of the computing environment.

Snoop filtering system in a multiprocessor system

The present in invention is directed to a method, system and program storage device for efficiently implementing a multidimensional Fast Fourier Transform (FFT) of a multidimensional array comprising a plurality of elements initially distributed in a multi-node computer system comprising a plurality of nodes in communication over a network, comprising: distributing the plurality of elements of the array in a first dimension across the plurality of nodes of the computer system over the network to facilitate a first one-dimensional FFT; performing the first one-dimensional FFT on the elements of the array distributed at each node in the first dimension; re-distributing the one-dimensional FFT-transformed elements at each node in a second dimension via 'all-to-all' distribution in random order across other nodes of the computer system over the network; and performing a second one-dimensional FFT on elements of the array re-distributed at each node in the second dimension, wherein the random order facilitates efficient utilization of the network thereby efficiently implementing the multidimensional FFT. The 'all-to-all' re-distribution of array elements is further efficiently implemented in applications other than the multidimensional FFT on the distributed-memory parallel supercomputer.

Efficient implementation of multidimensional fast fourier transform on a distributed-memory parallel multi-node computer

A system and method for enabling high-speed, low-latency global collective communications among interconnected processing nodes. The global collective network optimally enables collective reduction operations to be performed during parallel algorithm operations executing in a computer structure having a plurality of the interconnected processing nodes. Router devices ate included that interconnect the nodes of the network via links to facilitate performance of low-latency global processing operations at nodes of the virtual network and class structures. The global collective network may be configured to provide global barrier and interrupt functionality in asynchronous or synchronized manner. When implemented in a massively-parallel supercomputing structure, the global collective network is physically and logically partitionable according to needs of a processing algorithm.

A data capture technique for high speed signaling to allow for optimal sampling of an asynchronous data stream. This technique allows for extremely high data rates and does not require that a clock be sent with the data as is done in source synchronous systems. The present invention also provides a hardware mechanism for automatically adjusting transmission delays for optimal two-bit simultaneous bi-directional (SiBiDi) signaling.

Data capture technique for high speed signaling

A low latency memory system access is provided in association with a weakly-ordered multiprocessor system. Bach processor in the multiprocessor shares resources, and each shared resource has an associated lock within a locking device that provides support for synchronization between the multiple processors in the multiprocessor and the orderly sharing of the resources. A processor only has permission to access a resource when it owns the lock associated with that resource, and an attempt by a processor to own a lock requires only a single load operation, rather than a traditional atomic load followed by store, such that the processor only performs a read operation and the hardware locking device performs a subsequent write operation rather than the processor. A simple prefetching for non-contiguous data structures is also disclosed. A memory line is redefined so that in addition to the normal physical memory data, every line includes a pointer that is large enough to point to any other line in the memory, wherein the pointers to determine which memory line to prefetch rather than some other predictive algorithm. This enables hardware to effectively prefetch memory access patterns that are non-contiguous, but repetitive.

Low latency memory access and synchronization

A parallel computer system is constructed as a network of interconnected compute nodes to operate a global message-passing application for performing communications across the network. Each of the compute nodes includes one or more individual processors with memories which run local instances of the global message-passing application operating at each compute node to carry out local processing operations independent of processing operations carried out at other compute nodes. Each compute node also includes a DMA engine constructed to interact with the application via Injection FIFO Metadata describing multiple Injection FIFOs where each Injection FIFO may containing an arbitrary number of message descriptors in order to process messages with a fixed processing overhead irrespective of the number of message descriptors included in the Injection FIFO.

DMA engine for repeating communication patterns

A method for passing remote messages in a parallel computer system formed as a network of interconnected compute nodes includes that a first compute node (A) sends a single remote message to a remote second compute node (B) in order to control the remote second compute node (B) to send at least one remote message. The method includes various steps including controlling a DMA engine at first compute node (A) to prepare the single remote message to include a first message descriptor and at least one remote message descriptor for controlling the remote second compute node (B) to send at least one remote message, including putting the first message descriptor into an injection FIFO at the first compute node (A) and sending the single remote message and the at least one remote message descriptor to the second compute node (B).

Growing community of inventors

Bedford Hills, NY, United States of America

Pavlos Michael Vranas