Samsung Electronics Co., Ltd.

Luminous Computing, Inc.

Intel Corporation

Anahita Shayesteh

Ramdas P Kachare

Vijay Balakrishnan

Zvi Guz

Oscar Prem Pinto

Son Truong Pham

Stephen Garry Fischer

Xuebin Yao

Fred Worley

David Cureton Baker

Donovan Popps

Ari Novack

Bryan Cope

Mark Baur

Benjamin Wiley Melton

Mani Azimi

Sina Hassani

Changho Choi

Ioannis T Schoinas

Akhilesh Kumar

Naveen Cherukuri

Jason Martineau

Jingpei Yang

Mitchell A Nahmias

Dennis W Brzezinski

Timothy Clayton Bisson

Aniruddha Vaidya

Dong Hyuk Woo

Yasser Zaghloul

Dave Baker

Saikat Saharoy

Salvatore Arcuri

Vishal Kasliwal

Conventional high performance computer connections are electron-based systems, which require the memory packages to be as close as mechanically possible to the computation engine. Low power and high bandwidth long distance communication, e.g. photonic or electronic, links can drastically change the architecture of high-performance computers by eliminating the bottlenecks in communication. A computer system comprises: a plurality of memory aggregation devices configured to retrieve data from and store data in a plurality of random access memory modules forming a unified contiguous memory address space disaggregated from a processing unit; one or more computational devices configured for simultaneously launching a plurality of data signals including memory read and/or write requests for the data to the plurality of memory aggregation devices; and a plurality of communication links coupling each of the plurality of memory aggregation devices to each of the one or more computational devices for transferring the data therebetween.

Computer architecture with disaggregated memory and high-bandwidth communication interconnects

A computer system comprises one or more processing units and a plurality of memory locations interconnected by a plurality of data links, each processing unit comprising a data request engine and a processor. A method for accessing data in memory of the computer system, includes: receiving a data access job at a first data request engine specifying data to be accessed; and generating a plurality of memory access requests to request the plurality of data partitions needed to fulfill the data access job. The plurality of memory access requests are then delivered to the plurality of memory locations, and the plurality of data partitions are copied or moved from the plurality of memory locations to the first processing unit via the data links. The data access job specifies a plurality of data partitions of a source data structure distributed over the plurality of memory locations of the computer system. One purpose of this computer system is to accelerate the accessing of the rows or vectors in an embedding table for machine learning.

Computer system and method for data access

According to one general aspect, a device may include a host interface circuit configured to communicate with a host device via a data protocol that employs data messages. The device may include a storage element configured to store data in response to a data message. The host interface circuit may be configured to detect when a tunneling command is embedded within the data message; extract a tunneled message address information from the data message; retrieve, via the tunneled message address information, a tunneled message stored in a memory of the host device; and route the tunneled message to an on-board processor and/or data processing logic. The on-board processor and/or data processing logic may be configured to execute one or more instructions in response to the tunneled message.

Systems and methods for message tunneling

A field programmable gate array (FPGA) may: identify a continuous match of atoms between the search sequence and the reference sequence; divide the search sequence into a left portion of the search sequence that includes atoms before the continuous match of atoms in the search sequence, a center portion of the search sequence that includes the continuous match of atoms in the search sequence, and a right portion of the search sequence that includes atoms after the continuous match of atoms in the search sequence; match the left portion of the search sequence with the reference sequence; and match the right portion of the search sequence with the reference sequence.

Parallelizable sequence alignment systems and methods

Conventional high performance computer connections are electron-based systems, which require the memory packages to be as close as mechanically possible to the computation engine. Low power and high bandwidth communication, e.g. photonic, links can drastically change the architecture of high-performance computers by eliminating the bottlenecks in communication. A computer system comprises: a plurality of memory aggregation devices configured to retrieve data from and store data in a plurality of random access memory modules forming a unified contiguous memory address space disaggregated from a processing unit; a plurality of computational devices configured for simultaneously launching a plurality of data signals including memory read and/or write requests for the data to the plurality of memory aggregation devices; and a plurality of communication links coupling each of the plurality of memory aggregation devices to each of the plurality of computational devices for transferring the data therebetween.

Conventional high performance computer connections are electron-based systems, which require the memory packages to be as close as mechanically possible to the computation engine. Low power and high bandwidth communication, e.g. photonic, links can drastically change the architecture of high-performance computers by eliminating the bottlenecks in communication. A computer system comprises: a plurality of memory aggregation devices configured to retrieve data from and store data in a plurality of random access memory modules forming a unified contiguous memory address space disaggregated from a central processing unit; a plurality of computational devices configured for simultaneously launching a plurality of data signals including memory read and/or write requests for the data to the plurality of memory aggregation devices; and a plurality of communication links coupling each of the plurality of memory aggregation devices to each of the plurality of computational devices for transferring the data therebetween.

A system includes a plurality of storage processing accelerators (SPAs), at least one SPA of the plurality of SPAs including a plurality of programmable processors or storage processing engines (SPEs), the plurality of SPEs including n SPEs (n is a natural number greater than zero), where 1st to (n&#x2212;1) SPEs of the n SPEs are configured to provide an output of the SPE to a next SPE of the n SPEs in a pipeline to be used as an input of the next SPE; and an acceleration platform manager (APM) connected to the plurality of the SPAs and the plurality of SPEs, and configured to control data processing in the plurality of SPAs and the plurality of SPEs.

System and method for accelerated data processing in SSDS

A solid-state drive (SSD) includes: a plurality of data blocks; a plurality of flash channels and a plurality of ways to access the plurality of data blocks; and an SSD controller that configures a block size of the plurality of data blocks. A data file is stored in the SSD with one or more key-values pairs, and each key-value pair has a block identifier as a key and a block data as a value. A size of the data file is equal to the block size or a multiple of the block size.

System and method for leveraging key-value storage to efficiently store data and metadata in a distributed file system

A method includes: receiving, at an acceleration platform manager (APM) from an application service manager (ASM), application function processing information; allocating, by the APM, a first storage processing accelerator (SPA) from a plurality of SPAs, wherein at least one SPA of the plurality of SPAs comprises a plurality of programmable processors or storage processing engines (SPEs), the plurality of SPEs comprising n SPEs, enabling the plurality of SPEs in the first SPA, wherein once enabled, the at least one SPE of the plurality of SPEs in the first SPA is configured to process data based on the application function processing information; determining, by the APM, if data processing is completed by the at least one SPE of the plurality of SPEs in the first SPA; and sending, by the APM, a result of the data processing by the SPEs of the first SPA, to the ASM.

System and method for accelerated data processing in SSDs

Accelerated data processing in SSDs comprises SPAs an APM and host processor whereby the SPAs has multiple of SPEs

According to one general aspect, an apparatus may include a storage memory to store a plurality of key-value pairs. The apparatus may include at least one snapshot counter configured to store an operation number associated with a respective snapshot of the plurality of key-value pairs. The apparatus may include a snapshot data structure configured to identify, for at least one key-value pair, which, if any, snapshot(s) the respective key-value pair is associated with.

Implementing snapshot and other functionality in KVSSD through garbage collection and FTL

A host machine is disclosed. The host machine may include a host processor, a memory, an operating system running on the host processor, and an application running under the operating system on the host processor. The host machine may also include a Peripheral Component Interconnect Express (PCIe) tunnel to a Non-Volatile Memory Express (NVMe) Solid State Drive (SSD) and an RPC capture module which may capture the RPC from the application and deliver a result of the RPC to the application as though from the host processor, where the NVMe SSD may execute the RPC to generate the result.

Method for NVMe SSD based storage service using RPC and gRPC tunneling over PCIe +

A method of selecting among a plurality of I/O streams through which data is to be written to a multi-streaming flash storage device is presented. According to an example embodiment, the method comprises: assigning write sequences of similar length to the same I/O streams; receiving instructions for a write operation, the instructions including a starting logical block address (LBA) and a number of blocks of data to be written; determining whether the write operation is part of an existing write sequence; identifying an I/O stream associated with an existing write sequence; and providing a stream ID of the identified I/O stream to the multi-streaming flash storage device.

Growing community of inventors

Los Altos, CA, United States of America

Anahita Shayesteh