Emc IP Holding Company LLC

Dell Products Inc.

Dragan Savic

Adrian Michaud

Michael Robillard

Stephen J Todd

Said Tabet

Assaf Natanzon

Junping Zhao

William Price Dawkins

John S Harwood

Robert W Beauchamp

James W Espy

Amnon Izhar

Patrick J Weiler

Jimmy Doyle Pike

Roy E Clark

Shyamkumar Thiyagarajan Iyer

Kurtis John Bowman

Kelly G Kanellakis

Systems and methods are provided for optimizing GPU memory allocation for high-performance applications such as deep learning (DL) computing. For example, a DL task is executed using GPU resources (GPU device and GPU memory) to process a DL model having functional layers that are processed in a predefined sequence. A current functional layer of the DL model is invoked and processed using the GPU device. In response to the invoking, a data compression operation is performed to compress data of a previous functional layer of the DL model, and store the compressed data in the GPU memory. Responsive to the invoking, compressed data of a next functional layer of the DL model is accessed from the GPU memory and a data decompression operation is performed to decompress the compressed data for subsequent processing of the next functional layer of the DL model by the GPU device.

Optimization of graphics processing unit memory for deep learning computing

A physical event to be modeled is selected. A profile for the physical event is generated based on an event type of the physical event. Data is obtained from a plurality of data sources, wherein the obtained data comprises data relevant to the physical event that is collected by the plurality of data sources, and further wherein at least a portion of the obtained data comprises one or more of spatial and temporal references associated with the collection of the data. A digital representation of the physical event is generated based on at least a portion of the obtained data and the generated profile. The digital representation is utilized to analyze one or more other physical events associated with the modeled physical event.

Multi-dimensional event model generation

Techniques are provided for implementing a parameter server within a networking infrastructure of a computing system to reduce the communication bandwidth and latency for performing communication synchronization operations of the parameter server. For example, a method includes executing a distributed deep learning (DL) model training process to train model parameters of a DL model using a plurality of worker nodes executing on one or more server nodes of a computing system, and executing a parameter server within a networking infrastructure of the computing system to aggregate local model parameters computed by the plurality of worker nodes and to distribute aggregated model parameters to the plurality of worker nodes using the networking infrastructure of the computing system.

Implementing parameter server in networking infrastructure for high-performance computing

A system includes a host processor, a volatile memory device coupled to the host processor, and at least a first persistent memory device coupled to the host processor. The host processor is configured to execute one or more applications. The volatile memory device and the first persistent memory device are in respective distinct fault domains of the system, and at least one of a plurality of data objects generated by a given one of the applications is accessible from multiple distinct storage locations in respective ones of the distinct fault domains. For example, the host processor and the volatile memory device may be in a first one of the distinct fault domains and the first persistent memory device may be in a second one of the distinct fault domains. The data object remains accessible in one of the fault domains responsive to a failure in another of the fault domains.

System comprising non-volatile memory device and one or more persistent memory devices in respective fault domains

A virtual machine migration system that includes an initiator host that provides a virtual machine and receives an instruction to migrate the virtual machine to the target host. The initiator host moves, while the virtual machine is operating on the initiator host, virtual machine data for the virtual machine from a local memory area in a local memory system to a memory fabric memory area in a memory fabric memory system. A first portion of the virtual machine data in the local memory area and a second portion of the virtual machine data in the memory fabric memory area are accessible to the virtual machine and the initiator host during the movement of the virtual machine data. The initiator host stops, in response to all the virtual machine data being moved to the memory fabric memory area, operation of the virtual machine on the initiator host.

Virtual machine migration system

Techniques are disclosed for decentralized data management using a geographic location-based consensus protocol in a network of computing resources such as, by way of example, a highly distributed system. For example, at a given consensus node of a consensus network comprising a plurality of consensus nodes configured to participate in a consensus protocol wherein at least a portion of the consensus nodes are mobile, a list is obtained of at least a subset of the plurality of consensus nodes that are predicted to be currently available to participate in the consensus protocol based on geographic location information. A message comprising a transaction to be validated is sent from the given consensus node to the subset of the plurality of consensus nodes in the obtained list. Techniques are also disclosed for adjusting a data protection policy based on the number of computing nodes, some of which are mobile, available to participate.

Decentralized data management via geographic location-based consensus protocol

A system includes a host processor and at least one storage device coupled to the host processor. The host processor is configured to execute instructions of an instruction set, the instruction set comprising a first move instruction for moving data identified by at least one operand of the first move instruction into each of multiple distinct storage locations. The host processor, in executing the first move instruction, is configured to store the data in a first one of the storage locations identified by one or more additional operands of the first move instruction, and to store the data in a second one of the storage locations identified based at least in part on the first storage location. The instruction set in some embodiments further comprises a second move instruction for moving the data from the multiple distinct storage locations to another storage location.

Host processor configured with instruction set comprising resilient data move instructions

Systems, methods, and articles of manufacture comprising processor-readable storage media are provided for implementing server architectures having dedicated systems for processing infrastructure-related workloads. For example, a computing system includes a server node. The server node includes a first processor, a second processor, and a shared memory system. The first processor is configured to execute data computing functions of an application. The second processor is configured to execute input/output (I/O) functions for the application in parallel with the data computing functions of the application executed by the first processor. The shared memory system is configured to enable exchange of messages and data between the first and second processors.

Server architecture having dedicated compute resources for processing infrastructure-related workloads

In a system environment comprising a plurality of computing resources, wherein at least a portion of the computing resources are mobile, a method manages a transfer of one or more portions of a data set between at least a subset of the plurality of computing resources in accordance with a data distribution process. The data distribution process comprises computing one or more probability values to estimate whether or not a given mobile computing resource that is seeking at least a portion of the data set will be in a vicinity of at least one other computing resource that currently has or can obtain the portion of the data set, and based on the computation step, causing a transfer of the portion of the data set to the given mobile computing resource over a communication link locally established between the two computing resources when in the vicinity of one another.

Data distribution management in mobile network

In a system environment comprising a plurality of computing resources, wherein at least a portion of the computing resources are mobile, a method maintains a decentralized messaging network of interconnected messaging nodes and a decentralized data network of interconnected data nodes. Each of the plurality of computing resources is associated with a given messaging node and a given data node. The method manages transfer of a data set between the plurality of computing resources in association with the decentralized messaging network and the decentralized data network. Managing transfer of the data set comprises inserting a policy file into the decentralized data network specifying one or more policies for managing the transfer of the data set and inserting a message into the decentralized messaging network instructing implementation of the one or more policies.

Decentralized data management across highly distributed systems

A system, computer program product, and computer-executable method for managing flash devices within a data storage environment utilized by an application of one or more applications, wherein the application accesses the managed flash devices through a pool of flash storage provided by the data storage, the system, computer program product, and computer-executable method comprising receiving a data I/O from the application, analyzing the data I/O directed toward the pool of flash storage in relation to the flash devices, and managing the flash devices based on data I/Os directed toward the pool of flash storage by the application.

Disaggregation of flash management

A first endpoint comprises a fabric attach point for attachment to a memory fabric, a first media controller, and a first non-volatile memory media. The memory fabric comprises a reliability zone comprising the first endpoint and at least a second endpoint. The first media controller is configured to receive, from at least one processor coupled to the first endpoint via the at least one fabric attach point, a memory fabric store command to store an object in the reliability zone. The first media controller is further configured to store the object in the first non-volatile memory media, to receive from the second endpoint a message indicating that the same object has been stored by the second endpoint, and to send to the at least one processor a single acknowledgement indicating that the at least one object has been stored in both the first and second endpoints of the reliability zone.

Memory fabric with reliability zone comprising two or more fabric attached memory endpoints

Systems and methods are provided to optimize checkpoint operations for deep learning (DL) model training tasks. For example, a distributed DL model training process is executed to train a DL model using multiple accelerator devices residing on one or more server nodes, and a checkpoint operation is performed to generate and store a checkpoint of an intermediate DL model. A checkpoint operation includes compressing a checkpoint of an intermediate DL model stored in memory of a given accelerator device to generate a compressed checkpoint, and scheduling a time to perform a memory copy operation to transfer a copy of the compressed checkpoint from the memory of the given accelerator device to a host system memory. The scheduling is performed based on information regarding bandwidth usage of a communication link to be utilized to transfer the compressed checkpoint to perform the memory copy operation, wherein the memory copy operation is performed at the scheduled time.

Optimization of checkpoint operations for deep learning computing

An ad-hoc computation system is formed from one or more clusters of idle mobile computing resources to execute an application program within a given time period. The forming step further comprises: (i) determining at least a subset of idle mobile computing resources from the one or more clusters of idle mobile computing resources that are available, or likely to be available, to execute the application program within the given time period, and that collectively comprise computing resource capabilities sufficient to execute the application program within the given time period; and (ii) distributing a workload associated with the execution of the application program to the subset of idle mobile computing resources. The workload associated with the application program is executed via the subset of idle mobile computing resources forming the ad-hoc computation system.

Ad-hoc computation system formed in mobile network

Systems and methods are provided for generating and managing ad-hoc mobile computing networks. For example, a method includes discovering, by a first mobile compute node, an existence of a second mobile compute node within a geographic location monitored by the first mobile compute node, and exchanging data between the first and second mobile compute nodes to negotiate conditions for forming a cluster of a mobile ad-hoc network. The conditions include, for example, a target purpose for forming the cluster, criteria for compute node membership within the cluster, and designation of one of the first and second mobile compute nodes as a master compute node for the cluster. The cluster including the first and second mobile compute nodes is then formed based on the negotiated conditions.

Growing community of inventors

Brookline, MA, United States of America

Dragan Savic