Abstract: The present disclosure provides devices and methods relating to remote direct memory access (RDMA). In one implementation, a target device of the RDMA operation is configured to receive a packet including a first destination address and a destination key, obtain one or more offset values, and obtain a second destination address based on the first destination address, the destination key, and the one or more offset values. Further, the target device is configured to initiate the RDMA operation on a memory based on the second destination address.

Abstract: A sequencer and method for batching execution of artificial intelligence (AI) jobs. A method includes receiving, by a plurality of AI servers, a plurality of AI jobs from a plurality of clients connected to the plurality of AI servers over a network, determining a first group of AI jobs from the plurality of AI jobs that are candidates for batching, batching the first group of AI jobs based on least one service parameter, and sending, for execution, the batch of first group AI jobs to a plurality of first compute engines reside in different AI servers of the plurality of AI servers.

Abstract: Apparatuses, systems, program products, and method are disclosed for machine learning abstraction. An apparatus includes an objective module configured to receive an objective to be analyzed using machine learning. An apparatus includes a grouping module configured to select a logical grouping of one or more machine learning pipelines to analyze a received objective. An apparatus includes an adjustment module configured to dynamically adjust one or more machine learning settings for a logical grouping of one or more machine learning pipelines based on feedback generated in response to analyzing a received objective.

Abstract: A method and server for batching execution of artificial inelegance (AI) jobs are provided. The method includes receiving, by an AI server, a plurality of AI jobs from a plurality of clients connected to an AI appliance over a network; for each AI job of the plurality of AI jobs: deriving at least one attribute of the received AI job; determining based on the at least one AI job attribute and at least one batching parameter if the received AI job is a candidate for batching; aggregating the received AI job into a created batch when the received AI job is determined to be a candidate for batching; continuing aggregating newly received AI jobs determined to be candidates for batching in the created batch until at least one service parameter is met; and sending the batch of AI jobs to a compute engine dedicated to executing the batch.

Abstract: A requesting device includes a memory, a controller, and a communication interface. The memory is configured to store a plurality of work elements in one or more requesting queues. Each work element indicates a requestor, a responder, and an operation. The controller is configured to retrieve at least a first work element of the plurality of work elements from the memory, generate a first hint message that includes an indication of at least one operation of the first work element, and transmit the first hint message to a first responding device over the communication interface. The first responding device corresponds to a first responder of the first work element. The controller is further configured to transmit a first request relating to the at least one operation of the first work element to the first responding device over the communication interface. The first request indicates the at least one operation.

Abstract: In response to a bootup or a reset of an input/output (I/O) device providing virtualized hardware resources for use by virtual function (VF) drivers of virtual machines (VM) and a load of a physical function (PF) driver, a global configuration status register (CSR) is set by the PF driver and/or the I/O device, to a mapping-value defining a certain mapping between base address register (BAR) roles supported by virtual functions exposed by the I/O device and I/O device BARs of the VFs. In response to the setting, the VF drivers for the VFs corresponding to the PF driver are loaded. Each respective VF driver obtains the mapping-value of the global CSR and maps, according to the certain mapping, BARs of the VF driver designated for the BAR roles, to the corresponding I/O device BARs of the virtual functions mapped to the BAR roles.

Abstract: An automated machine learning (“ML”) method may include training a first machine learning model using a first machine learning algorithm and a training data set; validating the first machine learning model using a validation data set, wherein validating the first machine learning model comprises generating an error data set; training a second machine learning model to predict a suitability of the first machine learning model for analyzing an inference data set, wherein the second machine learning model is trained using a second machine learning algorithm and the error data set; and triggering a remedial action associated with the first or second machine learning model in response to a predicted suitability of the first machine learning model for analyzing the inference data set not satisfying a suitability threshold.

Abstract: A method and system for communicating artificial intelligence (AI) tasks for a server chaining are presented. The method includes establishing a first connection between an AI client and a first AI server; encapsulating a request to process an AI task in at least one request data frame compliant with a communication protocol; and transporting the at least one request data frame over a network using a transport protocol over the first connection to the first AI server, wherein the first AI server spans the AI task over at least one second AI server, wherein the transport protocol provisions transport characteristics of the AI task and the transport protocol is different than the communication protocol, wherein AI task includes processing of a single compute graph thereby allow spanning the processing of the compute graph over one more AI servers.

Abstract: A method and apparatus are described. The method comprises receiving a data packet comprising data to be written into the memory of a computing system and address data comprising an address in a set of addresses of a first address space of the computing system, identifying a subset of the set of addresses of the first address space with a subset of addresses in a second address space associated with the memory of the computing system, determining an address from a further subset of addresses in the second address space, writing the data to the region of the memory associated with the determined address and updating an address translation table on the computing system on the basis of the determined address.

Abstract: A sending device is configured to generate a first message that includes a first indication of a first operation type; transmit the first message to a receiving device over the communications interface; generate a second message that includes a second indication of a second operation type; determine whether the second operation type is associated with the first operation type; determine, in response to determining that the second operation type is associated with the first operation type, that the local pacing timer has exceeded a timer duration since transmitting the first message; and transmit, in response to determining that the local pacing timer has exceeded the timer duration since transmitting the first message, the second message to the receiving device over the communications interface.

Abstract: The present disclosure provides devices and methods relating to remote direct memory access (RDMA). In one implementation, a target device of the RDMA operation is configured to receive a packet including a first destination address and a destination key, obtain one or more offset values, and obtain a second destination address based on the first destination address, the destination key, and the one or more offset values. Further, the target device is configured to initiate the RDMA operation on a memory based on the second destination address.

Abstract: A system and method for communicating artificial intelligence (AI) tasks between AI resources are provided. The method comprises establishing a connection between a first AI resource and a second AI resource; encapsulating a request to process an AI task in at least one request data frame compliant with a communication protocol, wherein the at least one request data frame is encapsulated at the first AI resource; and transporting the at least one request data frame over a network using a transport protocol to the second AI resource, wherein the transport protocol provisions the transport characteristics of the AI task, and wherein the transport protocol is different than the communication protocol.

Abstract: In order to reduce remote direct memory access (RDMA) requests drop in RDMA systems, a requesting device transmits a message that includes a prefetch operation to a responding device. The prefetch operation indicates a memory area to be loaded by the responding device to a memory of the responding device before receiving a new RDMA request or a RDMA command.

Abstract: A virtual function (VF) driver is provided, which receives an indication of base address registers (BARs) exposed by an I/O device providing virtualized hardware resources for use by the VF driver. The VF driver performs an internal probe and analysis to determine a target adaptation of at least one of the BARs exposed by the I/O device. A physical function (PF) driver receives a request message from the VF driver to adapt at least one of the BARs exposed by the I/O device. The BARs exposed by the I/O device are adapted by the PF driver according to the request message. A confirmation message is provided to the VF driver indicating the adaptation to the BARs exposed by the I/O device. The VF driver accesses regions of memory allocated to the adapted BARs exposed by the I/O device.

Abstract: Described herein are systems and methods for executing efficiently, in real-time, a plurality of machine learning processes. In one embodiment, a computing platform with multiple compute elements receives multiple data streams, each such stream associated with its own respective machine learning process. Each machine learning process is operative to use its data stream as input to train, in real-time, a respective mathematical model. Each of the processes has peaks and dips in processing demands. The system re-allocates, in real-time, compute elements from the processes with lower processing demands to processes with higher processing demands, thereby handling all of the multiple processes on-the-fly, preventing peak demands from causing the system to stall, and reducing overall the computational resources required by the system.

Abstract: Apparatuses, systems, program products, and methods are disclosed for interpretability-based machine learning adjustment during production. An apparatus includes a first results module that is configured to receive a first set of inference results of a first machine learning algorithm during inference of a production data set. An apparatus includes a second results module that is configured to receive a second set of inference results of a second machine learning algorithm during inference of a production data set. An apparatus includes an action module that is configured to trigger one or more actions that are related to a first machine learning algorithm in response to a comparison of first and second sets of inference results not satisfying explainability criteria.

Abstract: Described herein are systems and methods for executing efficiently, in real-time, a plurality of machine learning processes. In one embodiment, a computing platform with multiple compute elements receives multiple data streams, each such stream associated with its own respective machine learning process. Each machine learning process is operative to use its data stream as input to train, in real-time, a respective mathematical model. Each of the processes has peaks and dips in processing demands. The system re-allocates, in real-time, compute elements from the processes with lower processing demands to processes with higher processing demands, thereby handling all of the multiple processes on-the-fly, preventing peak demands from causing the system to stall, and reducing overall the computational resources required by the system.

Abstract: Apparatuses, systems, program products, and methods are disclosed for determining validity of machine learning algorithms for datasets. An apparatus includes a primary training module that is configured to train a first machine learning model for a first machine learning algorithm. An apparatus includes a primary validation module that is configured to validate a first machine learning model to generate an error data set. An apparatus includes a secondary training module that is configured to train a second machine learning model for a second machine learning algorithm using an error data set. A second machine learning algorithm may be configured to predict a suitability of a first machine learning model for analyzing an inference data set. An apparatus includes an action module that is configured to trigger an action in response to a predicted suitability of the first machine learning model not satisfying a predetermined suitability threshold.

Abstract: Apparatuses, systems, program products, and method are disclosed for detecting suitability of machine learning models for datasets. An apparatus includes a training evaluation module configured to calculate a first statistical data signature for a training data set of a machine learning system using one or more predefined statistical algorithms. An apparatus includes an inference evaluation module configured to calculate a second statistical data signature for an inference data set of a machine learning system using one or more predefined statistical algorithms. An apparatus includes a score module configured to calculate a suitability score describing the suitability of a training data set to an inference data set as a function of a first and a second statistical data signature. An apparatus includes an action module configured to perform an action related to a machine learning system in response to a suitability score satisfying an unsuitability threshold.

Abstract: Apparatuses, systems, program products, and method are disclosed for machine learning abstraction. An apparatus includes an objective module configured to receive an objective to be analyzed using machine learning. An apparatus includes a grouping module configured to select a logical grouping of one or more machine learning pipelines to analyze a received objective. An apparatus includes an adjustment module configured to dynamically adjust one or more machine learning settings for a logical grouping of one or more machine learning pipelines based on feedback generated in response to analyzing a received objective.