An example method of managing a plurality of hardware accelerators in a computing system includes executing workload management software in the computing system configured to allocate a plurality of jobs in a job queue among a pool of resources in the computer system; monitoring the job queue to determine required hardware functionalities for the plurality of jobs; provisioning at least one hardware accelerator of the plurality of hardware accelerators to provide the required hardware functionalities; configuring a programmable device of each provisioned hardware accelerator to implement at least one of the required hardware functionalities; and notifying the workload management software that each provisioned hardware accelerator is an available resource in the pool of resources.

By obtaining metadata for transactions submitted by a service to a log-first distributed database of a provider network, a metrics manager may determine database replica performance for those transactions and notify clients of potential performance issues. When an instance of the service submits a write transaction to the log-first distributed database, the transaction may include a host name and a timestamp for the submission of the transaction. At a later point in time, a write applier may obtain the transaction and apply it to a local database replica, along with an additional timestamp for the application of the transaction to the replica. A metrics manager may obtain the transaction timestamps from the replica and calculate a latency metric for the propagation of the transaction from the particular service instance/instance host to the replica. The latency metric may be stored or transmitted to an endpoint (e.g., a client or administrator).

A method and an apparatus for dynamically managing a shared memory pool are provided, to determine, based on different service scenarios, a shared memory pool mechanism applicable to a current service scenario, and then dynamically adjust a memory pool mechanism based on the determining result. The method for dynamically managing a shared memory pool includes: determining a first shared memory pool mechanism, where the first shared memory pool mechanism is a fixed memory pool mechanism or a dynamic memory pool mechanism; determining a second shared memory pool mechanism suitable for a second service scenario based on the second service scenario, where the second shared memory pool mechanism is a fixed memory pool mechanism or a dynamic memory pool mechanism; and when the second shared memory pool mechanism is different from the first shared memory pool mechanism, adjusting the first shared memory pool mechanism to the second shared memory pool mechanism.

A resource scheduling method, an apparatus and a readable storage medium. In a process in which a terminal device participates in federated learning, a network device first generates first information, the first information including configuration information, and specifically, the configuration information is used to indicate characteristics of sample data, and/or, characteristics of a model to be trained for the terminal device to perform model training; and then, the network device sends the first information to the terminal device, and the terminal device acquires the configuration information in the first information. In the present embodiment, the network device determines, according to a data processing capability and/or a data transmission capability of the terminal device, that the terminal device participates in a current round of federated learning, and generates and sends the configuration information of a training parameter that matches the terminal device.

[Problem] When resource reserved in a resource sharing system become unavailable, the reservation is efficiently transferred. [Solution] In a resource sharing system 10, a plurality of users 20 (user terminals) share a plurality of resources 30. A resource reservation management device 42 includes: a reservation setting unit 402 that accepts a reservation request including a usage condition of the plurality of resources 30 from the user 20 and sets a usage reservation according to the usage condition to a first resource predetermined 30 in the resource sharing system 10; and a reservation changing unit 404 that re-sets the usage reservation to a second resource 30 being different from the first resource 30 in the resource sharing system 10 when a reserved resource 30 becomes unavailable. When a resource capacity of the second resource 30 is insufficient for the usage reservation to be re-set, the reservation changing unit 404 changes the usage condition and re-sets the usage reservation to the second resource 30.

Apparatuses, systems, and techniques to execute one or more application programming interface (API) functions to facilitate parallel computing. In at least one embodiment, one or more APIs are to indicate information about one or more storage locations using various novel techniques described herein.

Disclosed are a FAAS distributed computing method and apparatus. The method includes: decomposing a computation task into multiple steps with correlation and execution order, and constructing multiple mirror images and multiple method groups respectively; creating multiple dockers in a process and allocating hardware resources according to the mirror image corresponding to a specific step; transferring, according to a dynamic pointer of the corresponding method group, a data processing result to the method group corresponding to the next step, and simultaneously storing a generated intermediate result in a distributed memory file system; modifying dynamic pointers of the multiple method groups in real-time to dynamically adjust the correlation and execution order of the multiple steps; monitoring a running state of each method in the multiple method groups, restarting the docker of the method or loading latest docker snapshot of the method and expanding or reducing a capacity of the method groups.

Embodiments of the present disclosure may provide dynamic and fair assignment techniques for allocating resources on a demand basis. Assignment control may be separated into at least two components: a local component and a global component. Each component may have an active dialog with each other; the dialog may include two aspects: 1) a demand for computing resources, and 2) a total allowed number of computing resources. The global component may allocate resources from a pool of resources to different local components, and the local components in turn may assign their allocated resources to local competing requests. The allocation may also be throttled or limited at various levels.

A computing architecture providing for rapid analysis and control of an environment via edge computing nodes is disclosed. Input data streams may be captured via one or more data stream independent CPU threads and prepared for processing by one or more machine learning models. The machine learning models may be trained according to different use cases to facilitate a multi-faceted and comprehensive analysis of the input data. The evaluation of the input data against the machine learning models may be facilitated via independent GPU threads (e.g., one thread per model or use case) and the outputs of the models may be evaluated using control logic to produce a set of outcomes and control data. The control data may be utilized to generate one or more command messages that may provide feedback to a remote device or user regarding a state of a monitored environment or other observed condition.

A solution is proposed for controlling invocations of a target component by multiple source components in a software application. A corresponding method comprises associating a plurality of source components in a software application with one or more corresponding source rates for invoking a target component in the software application; monitoring corresponding invocations of the target component by a number of instances of the plurality of source components; receiving an enablement request for a new invocation of the target component from a current instance of a current source component; verifying an enablement of the new invocation; estimating a serving probability indicative of a residual capability of the target component to serve the new invocation; and enabling the new invocation according to the serving probability.