Intel Corporation

Sayantan Sur

Keith D Underwood

Todd M Rimmer

Robert J Woodruff

James Dinan

Mario Flajslik

Mark Sean Hefty

William R Magro

Andrew Friedley

Maria Garzaran

Nusrat Islam

Ajaya V Durg

Ravindra Babu Ganapathi

Gengbin Zheng

Rajesh M Sankaran

Andrew J Herdrich

David M Puffer

Pratik M Marolia

Rahul Pal

Mark Debbage

Bruce G Warren

Arlin R Davis

Robert Zak

Charles A Giefer

Timo Schneider

Shiow-wen Wendy Cheng

Nayan Amrutlal Suthar

Michael Heinz

Anupama Kurpad

Akhil Langer

Travis Hamilton

Shiow-Wen Cheng

In an embodiment, at least one interface mechanism may be provided. The mechanism may permit, at least in part, at least one process allocate, at least in part, and/or configure, at least in part, at least one network-associated object. Such allocation and/or configuration, at least in part, may be in accordance with at least one parameter set that may correspond, at least in part, to at least one query issued by the at least one process via the mechanism. Many modifications are possible without departing from this embodiment.

Discovery, configuration, and control of fabric interfaces

Algorithms for optimizing small message collectives with hardware supported triggered operations and associated methods, apparatus, and systems. The algorithms are implemented in a distributed compute environment comprising a plurality of ranks including a root, a plurality of intermediate nodes, and a plurality of leaf nodes, where each of the plurality of ranks comprising a compute platform having a communication interface including embedded logic for implementing the algorithms. Collectives are employed to transfer data between parent ranks and child ranks. In connection with the collectives, control messages are sent from children of a collective to the parent of the collective informing the parent that the children of the collective have free buffers ready to receive data. The parent employs a counter to determine that a control message has been received from each of its children indicating each child has a free buffer prior to sending data to the children in the collective.

Algorithms for optimizing small message collectives with hardware supported triggered operations

Methods and apparatus for an accelerator controller hub (ACH). The ACH may be a stand-alone component or integrated on-die or on package in an accelerator such as a GPU. The ACH may include a host device link (HDL) interface, one or more Peripheral Component Interconnect Express (PCIe) interfaces, one or more high performance accelerator link (HPAL) interfaces, and a router, operatively coupled to each of the HDL interface, the one or more PCIe interfaces, and the one or more HPAL interfaces. The HDL interface is configured to be coupled to a host CPU via an HDL link and the one or more HPAL interfaces are configured to be coupled to one or more HPALs that are used to access high performance accelerator fabrics (HPAFs) such as NVlink fabrics and CCIX (Cache Coherent Interconnect for Accelerators) fabrics. Platforms including ACHs or accelerators with integrated ACHs support RDMA transfers using RDMA semantics to enable transfers between accelerator memory on initiators and targets without CPU involvement.

Accelerator controller hub

Examples described herein relate to a first graphics processing unit (GPU) with at least one integrated communications system, wherein the at least one integrated communications system is to apply a reliability protocol to communicate with a second at least one integrated communications system associated with a second GPU to copy data from a first memory region to a second memory region and wherein the first memory region is associated with the first GPU and the second memory region is associated with the second GPU.

Direct memory writes by network interface of a graphics processing unit

Examples described herein relate to receiving, at a network interface, an allocation of a first group of one or more buffers to store data to be processed by a Message Passing Interface (MPI) and based on a received packet including an indicator that permits the network interface to select a buffer for the received packet and store the received packet in the selected buffer, the network interface storing a portion of the received packet in a buffer of the first group of the one or more buffers. The indicator can permit the network interface to select a buffer for the received packet and store the received packet in the selected buffer irrespective of a tag and sender associated with the received packet. In some examples, based on a received packet including an indicator that does not permit storage of the received packet in a buffer irrespective of a tag and source associated with the second received packet, the network interface is to store a portion of the second received packet in a buffer of the second group of one or more buffers, wherein the buffer of the second group of one or more buffers corresponds to a tag and source associated with the second received packet.

Buffer allocation for parallel processing of data by message passing interface (MPI)

An embodiment of a semiconductor package apparatus may include technology to embed one or more trigger operations in one or more messages related to collective operations for a neural network, and issue the one or more messages related to the collective operations to a hardware-based message scheduler in a desired order of execution. Other embodiments are disclosed and claimed.

Triggered operations to improve allreduce overlap

Examples include a method of managing storage for triggered operations. The method includes receiving a request to allocate a triggered operation; if there is a free triggered operation, allocating the free triggered operation; if there is no free triggered operation, recovering one or more fired triggered operations, freeing one or more of the recovered triggered operations, and allocating one of the freed triggered operations; configuring the allocated triggered operation; and storing the configured triggered operation in a cache on an input/output (I/O) device for subsequent asynchronous execution of the configured triggered operation.

Triggered operations for collective communication

At least one mechanism to permit, at least in part, allocation and/or configuration, at least in part, of at least one network-associated object

Methods, software, and systems for improved data transfer operations using overlapped rendezvous memory registration. Techniques are disclosed for transferring data between a first process operating as a sender and a second process operating as a receiver. The sender sends a PUT request message to the receiver including payload data stored in a send buffer and first and second match indicia. The first match indicia is used to determine whether the PUT request is expected or unexpected. If the PUT request is unexpected, an RMA GET operation is performed using the second matching indicia to pull data from the send buffer and write the data to a memory region in the user space of the process associated with the receiver. If the PUT request message is expected, the data payload with the PUT request is written to a receive buffer on the receiver determined using the first match indicia.

Overlapped rendezvous memory registration

Technologies for fine-grained completion tracking of memory buffer accesses include a compute device. The compute device is to establish multiple counter pairs for a memory buffer. Each counter pair includes a locally managed offset and a completion counter. The compute device is also to receive a request from a remote compute device to access the memory buffer, assign one of the counter pairs to the request, advance the locally managed offset of the assigned counter pair by the amount of data to be read or written, and advance the completion counter of the assigned counter pair as the data is read from or written to the memory buffer. Other embodiments are also described and claimed.

Technologies for fine-grained completion tracking of memory buffer accesses

Technologies for offloaded management of communication are disclosed. In order to manage communication with information that may be available to applications in a compute device, the compute device may offload communication management to a host fabric interface using a credit management system. A credit limit is established, and each message to be sent is added to a queue with a corresponding number of credits required to send the message. The host fabric interface of the compute device may send out messages as credits become available and decrease the number of available credits based on the number of credits required to send a particular message. When an acknowledgement of receipt of a message is received, the number of credits required to send the corresponding message may be added back to an available credit pool.

Technologies for offloaded management of communication

Examples include a computing system having an input/output (I/O) device including a plurality of counters, each counter operating as one of a completion counter and a trigger counter, a processing device; and a memory device. The memory device stores instructions that, in response to execution by the processing device, cause the processing device to represent a plurality of triggered operations of collective communication for high-performance computing executable by the I/O device as a directed acyclic graph stored in the memory device, with triggered operations represented as vertices of the directed acyclic graph and dependencies between triggered operations represented as edges of the directed acyclic graph; traverse the directed acyclic graph using a first process to identify and mark vertices that can share a completion counter; and traverse the directed acyclic graph using a second process to assign a completion counter and a trigger counter for each vertex.

Minimizing usage of hardware counters in triggered operations for collective communication

Techniques are disclosed to throttle bandwidth imbalanced data transfers. In some examples, an example computer-implemented method may include splitting a payload of a data transfer operation over a network fabric into multiple chunk get operations, starting the execution of a threshold number of the chunk get operations, and scheduling the remaining chunk get operations for subsequent execution. The method may also include executing a scheduled chunk get operation in response determining a completion of an executing chunk get operation. In some embodiments, the chunk get operations may be implemented as triggered operations.

Throttling for bandwidth imbalanced data transfers

In one embodiment, an apparatus includes: a plurality of queues having a plurality of first entries to store receive information for a process; a master queue having a plurality of second entries to store wild card receive information, where redundant information of the plurality of second entries is to be included in a plurality of redundant entries of the plurality of queues; and a control circuit to match an incoming receive operation within one of the plurality of queues. Other embodiments are described and claimed.

Method, apparatus and system for performing matching operations in a computing system

An embodiment includes a low-latency mechanism for performing a checkpoint on a distributed application. More specifically, an embodiment of the invention includes processing a first application on a compute node, which is included in a cluster, to produce first computed data and then storing the first computed data in volatile memory included locally in the compute node; halting the processing of the first application, based on an initiated checkpoint, and storing first state data corresponding to the halted first application in the volatile memory; storing the first state information and the first computed data in non-volatile memory included locally in the compute node; and resuming processing of the halted first application and then continuing the processing the first application to produce second computed data while simultaneously pulling the first state information and the first computed data from the non-volatile memory to an input/output (IO) node.

Growing community of inventors

Portland, OR, United States of America

Sayantan Sur