Cray Research, Inc.

Silicon Graphics International Corporation

Hewlett Packard Enterprise Development LP

Silicon Graphics, Incorporated

Cray Incorporated

Eric Carl Fromm

Steven M Oberlin

Randal Steven Passint

Steven Lee Scott

Joseph George Tietz

Michael L Anderson

Kevin M Knecht

Gregory Michael Thorson

Rodney A Ruesch

Richard E Kessler

Lonnie R Heidtke

Frank R Dropps

Steven J Dean

Mark S Birrittella

Thomas Edward McGee

Michael S Woodacre

Kirill Malkin

Michael E Malewicki

Douglas M Pase

Brian Justin Johnson

Robert J Lutz

Chris Dickson

Robert W Cutler

Richard G Finstad

Janet M Eberhart

Thomas A MacDonald

Mark F Sauder

Dale R Purdy

John F Steiner

Harro Zimmermann

Gary W Elsesser

This patent application relates generally to a predictive arbitration circuit for use in arbitrating access by a number of data streams to a shared resource managed by a destination (arbiter), where each data stream is associated with a number of sources competing for the shared resource, and the destination provides access to the shared resource based on the number of sources competing for the shared resource rather than just on the number of data streams. Among other things, this approach can more fairly distribute access to the shared resource among the competing sources.

Predictive arbitration circuit

This patent application relates generally to an age-based arbitration circuit for use in arbitrating access by a number of data streams to a shared resource managed by a destination (arbiter), in which age-based determinations are performed at the input sources of the data streams in order to designate certain packets as high-priority packets based on packet ages, and the destination expedites processing of the high-priority packets. Among other things, this approach offloads the age-based determinations from the destination, where they otherwise can cause delays in processing packets.

Age-based arbitration circuit

A node controller for a first processor socket group may include a node memory storing a coherence directory and logic. Logic may cause the node controller to: receive a memory operation request directly from a second processor socket group, follow a coherence protocol based on the memory operation request and the coherence directory and directly access a socket group memory of the first processor socket group based on the request.

Node controller direct socket group memory access

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

A high-performance-computer system includes a statistics accumulation apparatus configured to efficiently accumulate system performance data from a variety of system components, and periodically write such data to processor local memory for efficient subsequent software processing of the thus acquired data, thereby reducing the system hardware and software overhead needed for collection of such data as compared to prior art systems.

Enhanced performance monitoring method and apparatus

An adaptive router anticipates possible future congestion and enables selection of an alternative route before the congestion occurs, thereby avoiding the congestion. The adaptive router may use a primary route until it predicts congestion will occur. The adaptive router measures packet traffic volume, such as flit volume, on a primary network interface to anticipate the congestion. The adaptive router maintains a trailing sum of the number of flits handled by the primary network interface over a trailing time period. If the sum exceeds a threshold value, the adaptive router assumes the route will become congested, and the adaptive router enables considering routing future packets, or at least the current packet, over possible secondary routes.

Bandwidth on-demand adaptive routing

Processors in a compute node offload transactional memory accesses addressing shared memory to a transactional memory agent. The transactional memory agent typically resides near the processors in a particular compute node. The transactional memory agent acts as a proxy for those processors. A first benefit of the invention includes decoupling the processor from the direct effects of remote system failures. Other benefits of the invention includes freeing the processor from having to be aware of transactional memory semantics, and allowing the processor to address a memory space larger than the processor's native hardware addressing capabilities. The invention also enables computer system transactional capabilities to scale well beyond the transactional capabilities of those found computer systems today.

Transactional memory proxy

Processor clock signals are generated for each processor in a HPC system, such that all the processor clock signals are of the same frequency. Furthermore, as part of a startup (boot) procedure, a process sets all time stamp counters (TSCs) of the processors, such they indicate identical times. Each blade of the HPC system recovers a recovered clock signal from a synchronous communication network, to which the blade is coupled. The blade generates a processor clock from the recovered clock signal and provides the processor clock to processor(s) on the blade. Each chassis is coupled to a second, system-wide, synchronous communication network, and each chassis synchronizes its chassis synchronous communication network with the system-wide synchronous communication system. Thus, all the processor clock signals are generated with the same frequency.

Global synchronous clock circuit and method for blade processors

Processing transaction requests in a shared memory multi-processor computer network is described. A transaction request is received at a servicing agent from a requesting agent. The transaction request includes a request priority associated with a transaction urgency generated by the requesting agent. The servicing agent provides an assigned priority to the transaction request based on the request priority, and then compares the assigned priority to an existing service level at the servicing agent to determine whether to complete or reject the transaction request. A reply message from the servicing agent to the requesting agent is generated to indicate whether the transaction request was completed or rejected, and to provide reply fairness state data for rejected transaction requests.

Non-saturating fairness protocol and method for NACKing systems

A method includes specifying a first set of interconnected devices associated with a first leaf cell in Verilog syntax, and specifying a second set of interconnected devices associated with a second leaf cell in Verilog syntax. A connection between the first leaf cell and the second leaf cell is also specified in Verilog syntax. This specifies a circuit. The functionality of the logic can be tested by running a logic simulation on the circuit without converting to Verilog syntax. The Verilog syntax, associated with the circuit, can be converted directly from Verilog syntax to a SPICE netlist. The SPICE netlist can be used to simulate the timing and other parameters of the circuit. The Verilog syntax can be used to verify the circuit. Also included are a computer readable medium including an instruction set for the above method, and a data structure necessary to carry out the above method.

Transistor level verilog

A method of performing remote address translation in a multiprocessor system includes determining a connection descriptor and a virtual address at a local node, accessing a local connection table at the local node using the connection descriptor to produce a system node identifier for a remote node and a remote address space number, communicating the virtual address and remote address space number to the remote node, and translating the virtual address to a physical address at the remote node (qualified by the remote address space number). A user process running at the local node provides the connection descriptor and virtual address. The translation is performed by matching the virtual address and remote address space number with an entry of a translation-lookaside buffer (TLB) at the remote node. Performing the translation at the remote node reduces the amount of translation information needed at the local node for remote memory accesses. The method supports communication within a scalable multiprocessor, and across the machine boundaries in a cluster.

Remote address translation in a multiprocessor system

A bus interface circuit and method for reliable data capture in the presence of bus-master changeovers and/or for synchronizing received data to an internal clock signal, wherein the received data includes a strobe. Since the strobe may have a delay that is unknown (due to varying distances from the driver, clock jitter, and/or other causes), it is important to re-synchronize to the internal clock, and to do so with the smallest delay possible. This synchronization is provided in a way that also eliminates potential problems due to bus-master changeover, and in a way that minimizes time-critical signal generation. One aspect provides a method and/or apparatus for reliable data capture. The method includes: providing an N-stage latch including a first stage latch and a second stage latch, wherein N is two or larger; loading every Nth word of a data stream into the first stage latch using a first signal based on a strobe passed in the data stream; loading every N+1st word of the data stream into the second stage latch using a second signal based on the strobe passed in the data stream; unloading every Nth word from the first stage latch using a third signal based on an internal bus clock; and unloading every N+1st word from the second stage latch using a fourth signal based on the internal bus clock. In some embodiments, the first signal and the second signal are further based on a first stage selector and on a data_ready signal passed in the data stream.

Method and circuit for reliable data capture in the presence of bus-master changeovers

A method and apparatus for accessing memory-mapped registers that are distributed across a large integrated circuit. Some embodiments provide a method for accessing memory-mapped registers that are distributed across a first integrated circuit, the first integrated circuit including a plurality of logic subset modules, wherein each of the plurality of logic subset modules includes one or more memory-mapped registers. This method includes receiving a memory-mapped register access request into the first integrated circuit, serially transmitting, through each of the plurality of logic subset modules, a first plurality of data packets based on the memory-mapped register access request, wherein the first plurality of data packets includes an address specification for a memory-mapped register associated with a first one of the logic subset modules, and within the first logic subset module, accessing the memory-mapped register associated with the first logic subset module. Another aspect of the present invention provides an MMR circuit for accessing memory-mapped registers that are distributed across a first integrated circuit chip, the first integrated circuit chip including a plurality of logic subset modules.

Method and apparatus for accessing MMR registers distributed across a large asic

A method and apparatus for deallocating memory in a multi-processor, shared memory system. In one aspect, a node in the system has a node controller that contains sequencing logic. The sequencing logic receives a command across a network. The sequencing logic translates the received command into a Purge Translation Cache (PTC) instruction and sends the PTC instruction across a bus to a processor. The processor contains bus control logic that receives the PTC instruction and purges a virtual address specified in the PTC instruction from the processor's translation lookaside buffer. By purging the virtual address, the memory is deallocated.

Distribution of address-translation-purge requests to multiple processors

A method for extracting a PE number and offset from an array index by recursive centrifuging. According to one aspect of the present invention, a processing element number is assigned to each processing element, a local memory address is assigned to each memory location and a linearized index is assigned to each array element in a multidimensional array. The processing element number of the processing element in which a particular array element is stored is computed as a function of a linearized index associated with the array element and a mask word determined from the distribution specification associated with the array. The mask word is generated from the distribution specification and applied to a linearized index associated with a particular array element to obtain processing element number bits and local offset bits. The processing element number bits and local offset bits are then accumulated to create the processing element number and local offset for the memory location associated with the array element.

Growing community of inventors

Eau Claire, WI, United States of America

Eric Carl Fromm