20 results about "Workload scheduling" patented technology

Thermal output aware workload placement is disclosed. When a scheduler receives a workload to be deployed in computing resources such as a pool of servers, thermal data that may include fan related data and power consumption data is used to select one of the servers for the workload. The thermal data is used to select the server that reduces or minimizes the thermal output of the server and / or the computing resources.

Methods, systems, and devices for managing performance of workloads by hardware components housed in a power supply free chassis of a rack system are disclosed. To manage the performance, a request may be obtained to perform a workload of the workloads. Based, at least in part, on a phase of a lifecycle of an inference model that must be used to perform the workload, workload requirements may be obtained for the workload. Using the workload requirements and information regarding power available to data processing systems of the power supply free chassis, a scheduling process may be performed to identify a data processing system of the data processing systems to perform the workload. The workload request may be forwarded to a power manager of the data processing system to attempt to complete performance of the workload to thereby provide desired computer implemented services.

The invention provides a large language model application workload scheduling method, system and device, and relates to the technical field of model load scheduling, and the method comprises the steps: modeling a composite large language model application into a directed acyclic graph comprising a conventional stage, an LLM stage and a dynamic stage; modeling execution correlation among stages in the directed acyclic graph through a Bayesian network, dynamically predicting duration distribution of uncompleted stages, and calibrating a duration estimated value of an LLM stage in combination with a real-time batch processing size of an LLM executor; the uncertainty reduction amount of each ready stage is quantitatively scheduled based on information entropy; an epsilon-greedy strategy is adopted, and a JCT priority queue and an uncertainty reduction priority queue are combined to allocate scheduling resources; and assigning the task to a corresponding executor for execution, and repeating the above process until all operations are completed. The technical problem that the scheduling technology in the prior art is difficult to effectively deal with the execution time uncertainty and the structure uncertainty of the composite LLM application is solved.

Voltage droop is mitigated within an integrated circuit (IC) device including a first processing device and scheduler circuitry. The first processing device executes operations of a first workload. The first processing device is associated with a first processing device voltage droop characteristic. The scheduler circuitry receives the first workload. The first workload is associated with a first workload voltage droop characteristic. Further, the scheduler circuitry assigns the first workload to the first processing device of the IC device based on the first workload voltage droop characteristic and the first processing device voltage droop characteristic.

Systems and methods are provided for using historic input power periodic data from a server in an IT data center to train a machine learning (ML) model to obtain forecasted power consumption data of the server for a future time period. Time windows of hotspots or coldspots are then identified in the forecasted power consumption data, hotspots being defined as areas or regions of over-utilization in a time series data, and coldspots being defined as areas or regions of under-utilization in a time series data. The hotspots and coldspots are identified by calculating an exponential mean average (EMA) of the forecasted power consumption data, taking points above the EMA as hotspots and points below the EMA as coldspots. The identified hotspots and coldspots can be used to schedule workloads for a server or a data center, to more efficiently plan existing workloads, or to introduce new workloads at more optimal time periods.

To provide a method for generating a task graph for workload scheduling based on a task graph specification program.SOLUTION: The method includes a task graph specification processor receiving 902 a task graph specification program for execution, executing 904 control flow instructions specified in the task graph specification program to traverse the task graph specification program, generating 906 nodes based on path instructions of the task graph specification program, generating 908, in a task graph, resources and directed edges between the generated nodes based on resource utilization of each node, and outputting 910 the task graph to a command scheduler for scheduling on an acceleratedprocessingdevice (APD) or other device.SELECTED DRAWING: Figure 9

This disclosure provides methods, components, devices and systems for workload scheduling in edge computing. Some aspects more specifically relate to workload scheduling in distributed computing scenarios, which may include various devices including one or more workload scheduler devices, one or more host devices, and one or more client devices. A workload scheduler device may receive a workload request from a client device and may assign the workload to a host device from a set of host devices, with such a set of host devices including host devices capable of satisfying expectations of the workload. To assign the workload, the workload scheduler device may perform joint evaluations for the set of host devices, obtain scores for the set of host devices in accordance with the joint evaluations, rank the set of host devices in accordance with their respective scores, and assign the workload to the host device having the greatest ranking.

The application discloses a time-varying work load scheduling method based on priority and resource perception, comprising the following steps: 1) determining the cloud server cluster state and the work load needing to be scheduled; 2) performing equivalent class mapping on the work load; 3) constructing a priority queue; 4) using a parallel GAT layer for resource perception; 5) using a strategy network to generate a scheduling action; 6) scheduling the work load at the head of the priority queue to the server specified by the action; 7) state transition, generating a reward, and recording the corresponding triple in track 1; 8) updating the work load to be scheduled; 9) initializing the cloud server cluster state and the work load needing to be scheduled; 10) performing multiple simulation experiments and recording the corresponding track; and 11) updating the network weight according to the track. The application solves the problem of reasonably and effectively scheduling the time-varying work load continuously arriving under the condition of limited cloud server cluster resources.

This specification provides an application workload scheduling method and apparatus. The application workload scheduling method includes: acquiring observability data of a target application workload in an initial node; determining, based on the observability data, related application workloads that the target application workload depends on from all nodes; generating scheduling configuration data based on the dependency relationship between the target application workload and the related application workloads; and scheduling the target application workload and the related application workloads to corresponding target nodes based on the scheduling configuration data. By determining the related application workloads that the target application workload depends on from all nodes and generating scheduling configuration data based on the dependency relationship between application workloads to schedule the application workloads to the corresponding nodes, the communication latency between dependent application workloads is reduced because the dependency relationship between application workloads is taken into account during scheduling.

The invention discloses a directional migration method for stateless workload scheduling unit containers and a medium, and relates to the technical field of cloud computing. In a scheduling unit container scheduling management platform, a migration controller is indicated to receive and analyze a current scheduling unit container migration request of a user in real time to obtain a current to-be-migrated scheduling unit container and a target scheduling unit container migration node; establishing a current hot replacement scheduling unit container corresponding to the current to-be-migrated scheduling unit container on the target scheduling unit container migration node, and performing configuration according to a configuration rule; and if it is detected that the current hot replacement scheduling unit container is in a preparation completion state, according to the standard label and the ownership relationship, indicating the migration controller to modify the current hot replacement scheduling unit container, and managing the number of the scheduling unit containers. The problems that migration of the scheduling unit container is complex, the manpower efficiency is low and the business performance capacity is limited are solved, and the directional migration efficiency and accuracy of the scheduling unit container are improved.

Aspects of the present disclosure relate to an apparatus comprising a plurality of processing elements having a spatial layout, and control circuitry to assign workloads to said plurality of processing elements. The control circuitry is configured to, based on a timing parameter, determine one or more active processing elements to deactivate; determine, based on the spatial layout, one or more inactive processing elements to activate; and deactivate said one or more active processing elements and activate said one or more inactive processing elements.

Systems and methods are provided for using historic input power periodic data from a server in an IT data center to train a machine learning (ML) model to obtain forecasted power consumption data of the server for a future time period. Time windows of hotspots or coldspots are then identified in the forecasted power consumption data, hotspots being defined as areas or regions of over-utilization in a time series data, and coldspots being defined as areas or regions of under-utilization in a time series data. The hotspots and coldspots are identified by calculating an exponential mean average (EMA) of the forecasted power consumption data, taking points above the EMA as hotspots and points below the EMA as coldspots. The identified hotspots and coldspots can be used to schedule workloads for a server or a data center, to more efficiently plan existing workloads, or to introduce new workloads at more optimal time periods.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for scheduling workloads on computing resources using a high priority queue and a low priority queue. The high priority queue maintains pending high priority workloads to be scheduled for execution, and the low priority queue maintains pending low priority workloads to be scheduled for execution. The computing system as described in this specification schedules the pending low priority workloads for execution by utilizing computing resources provided by the system only when the high priority queue is empty.

A method for workload scheduling by a host device in communication with a head-mounted display includes prioritizing processing of a first head-mounted display (HMD) task in response to determining that the host device is within a target wake-up time (TWT) window. The method also includes prioritizing processing of the host task in response to determining that the host device is outside the TWT window. The method may also include evaluating whether increasing the processing core voltage and / or increasing the processing core clock frequency will allow host tasks to be completed in time to allow a second head mounted display (HMD) task to have sufficient time to complete processing before the end of the next TWT window.

An apparatus to facilitate workload scheduling is disclosed. The apparatus includes one or more clients, one or more processing units to processes workloads received from the one or more clients, including hardware resources and scheduling logic to schedule direct access of the hardware resources to the one or more clients to process the workloads.

A graphics processing apparatus includes a workload execution circuit to execute workloads and a performance counting circuit to count instances of performance metrics. A workload handling circuit receives commands and responds to performance counter sampling commands that indicate performance counter sampling contexts comprising performance metrics to be sampled and sampling intervals. The workload handling circuit monitors sampling intervals and triggers the workload execution circuit to write out sample values for performance metrics upon interval elapse. A driver receives performance metric sampling indications, allocates memory for sample values, generates performance counter sampling commands, and provides these to the workload handling circuit. The workload handling circuit writes out workload scheduling metadata, configures sampling according to sampling contexts, and manages the writing of sample values either directly to memory or back to the workload handling circuit with associated timestamp information.

Aspects of the present disclosure relate to an apparatus comprising a plurality of processing elements having a spatial layout, and control circuitry to assign workloads to said plurality of processing elements. The control circuitry is configured to, based on a timing parameter, determine one or more active processing elements to deactivate; determine, based on the spatial layout, one or more inactive processing elements to activate; and deactivate said one or more active processing elements and activate said one or more inactive processing elements.

Disclosed are systems and methods for workload scheduling in compute clusters using coolant health monitoring to optimize performance. In-situ sensors measure coolant properties in liquid cooling loops of compute nodes. A processing unit analyzes sensor data to determine coolant health levels and reallocates workloads from nodes with degraded coolant to nodes with higher coolant health levels, preempting thermal failures. A machine learning model processes coolant sensor data and performance metrics to generate cooling efficiency scores for each node. A cluster management module dynamically distributes computational tasks based on cooling system assessments, optimizing cluster efficiency and maintaining performance.

Certain aspects of the present disclosure provide techniques for artificial intelligence (AI) workload scheduling. An example method to distribute execution of at least one AI workload across a plurality of processing cores, generally includes obtaining first characteristics for the plurality of processing cores, obtaining second characteristics for the at least one AI workload, computing one or more metrics for each of the processing cores, based on the first characteristics and the second characteristics, and scheduling the at least one AI workload on at least one of the processing cores, based on the computed metrics and one or more conditions.

A computer-implemented method enables workload scheduling in a storage system for optimized deduplication. The method includes determining a dynamic deduplication correlation between workload processes in a previous time window. The workload process includes one or more tasks having defined execution timing parameters. The method further includes determining a deduplication ratio based on the deduplication correlation between the workload processes. The method further includes scheduling a plurality of workload processes based on a highest determined deduplication ratio among the determined deduplication ratios.