18 results about "Working set" patented technology

The invention relates to a pointer analysis-based incremental data flow path extraction method. The method comprises the following steps of: firstly, acquiring an initial pointer variable related to an incremental code as a task variable set; then, performing reverse query on the task variables based on the pointing relation constraint graph to obtain a basic variable set and a memory object pointed by the basic variable set; and after the working set is initialized, executing pointing relation propagation, propagating the memory object to a pointing set of the task variable, and classifying a new association pointer found in propagation into an indirect variable set. And executing explicit dependency query on the indirect variable set, adding a new dependency variable into the task set, repeating the process until the constraint graph is stable, and finally obtaining pointing information of all variables. And an inter-process data dependence sub-graph is constructed, and an incremental data flow path is extracted. According to the method, through a locality principle and a reverse query strategy, an analysis range is limited to incremental code related variables, and the overhead and memory problems of global analysis are avoided.

The application belongs to the technical field of computer memory management, and discloses an active page replacement method and system based on regional reuse mode identification, which comprises the following steps: working set identification based on memory access sample clustering, obtaining memory access samples by sampling the memory access instructions of a process, and clustering the samples to obtain the working set in the current monitoring window in real time; adaptive memory region division based on working set integration, dividing the address space into regions according to virtual address continuity and adaptively splitting / merging; reuse mode identification based on state transition, determining the reuse mode of each region according to the switching between hot and cold states of each region and the state duration; active page replacement and prefetching method based on the reuse mode, prefeteching the regions that will reach the reuse time, and triggering the replacement of the regions that are in the cooling state for a long time. Compared with the existing passive page fault triggering method, the application can reduce the page fault and jitter, reduce the replacement traffic, and improve the application execution efficiency and system throughput.

[Problem] To provide a work setting system for setting work of a mounted function unit that is mounted on a moving body capable of autonomous movement at a movement destination of the moving body. [Solution] A work setting system according to the present invention comprises: a movable moving body; the mounted function unit that executes a function at a movement destination by being mounted on the moving body; a map display unit that displays, on a user interface, a map that indicates a region in which the moving body can move; a point setting unit that sets a point that constitutes the movement destination of the moving body on the map, said point being the location where the function of the mounted function unit is executed; a point function display unit that displays, for the point, functions that can be executed by the mounted function unit, said functions being displayed in a manner that makes it possible to select a function; and a link function unit that associates the execution of the function that is selected by the point function display unit with the point that is set by the point setting unit.

A working-set structure is used to organize cached data for storing to persistent storage, which includes leaf structures and page descriptors (PDs) for data pages to be persisted. Upon adding a new PD located in an address range of an existing leaf structure, a PD population count of the existing leaf structure is compared to a predetermined PD population threshold. When the count is below the threshold, the new PD is incorporated into an existing set of PDs for the existing leaf structure, and otherwise (a) a new leaf structure is created, and (b) the new leaf structure is used for the new PD and later-added PDs in the address range. Flush parallelism is enhanced by avoiding large differences in PD population across a set of leaf structures.

1. The name of the design product: the working setting graphical user interface of electronic equipment. 2. The use of the design product: for an electronic equipment. 3. The design points of the design product: the graphical user interface. 4. The picture or photo that best shows the design points: design 1 front view. 5. Design 1 is designated as the basic design. 6. The use of the graphical user interface: the present graphical user interface is used for working setting. Design 1 front view to design 8 front view are working setting interfaces, and the user can make relevant settings through touch operation according to the content prompt of the interface. Design 9 front view is a working setting interface, and the user can set more parameters by sliding up in the pop-up window, as shown in the interface change state diagram of design 9.

The application relates to a GPU progressive optimization method for neural operator backbone network reasoning and belongs to the electronic information technical field.The method comprises the following steps: step S1, initializing constraints and tolerances;step S2, obtaining baseline execution time and reference results through baseline measurement;step S3, obtaining an estimated split value based on a working set model and expanding;step S4, performing KSweep evaluation and consistency checking, and preferably obtaining; a process of scanning and measuring each candidate set is called KSweep; step S5, fixing, adding fusion reasoning optimization, and performing consistency verification, if the consistency verification is passed, outputting and saving the optimal configuration, and directly loading and executing in the running stage; otherwise, backtracking or adjusting and recalibrating, and returning to step S3. The application shortens the reuse distance between adjacent layers, reduces the round-trip overhead of intermediate activation between the video memory and the cache, and thus improves the reasoning efficiency.

Automatically generating definitions of materialized views. Generally, automated MV generation identifies a set of candidate MVs by examining the working set of query blocks. Once candidates are formed, the candidate MVs are further evaluated to compute the benefit to the candidate MVs. An improved method for generating a candidate set of MVs is described herein. The improved method is referred to as the extended covered sub-expression technique (ECSE). Under ECSE, various relationships between join sets other than strict equivalence are used to generate new result join sets. Such relationships include subsets, intersections, supersets, and unions, which are described in further detail below. In some cases, relationships between the result join sets and the initial join sets are considered to generate new result join sets. The final result join sets are then used to form a candidate set of MVs.

The present disclosure provides a cluster management method, device, electronic equipment and storage medium. The cluster management method comprises: initiating a work cluster addition and deletion task by a meta cluster, the work cluster addition and deletion task being used to instruct to add a specified work cluster in a cluster system or to instruct to delete the specified work cluster in the cluster system; wherein the cluster system comprises at least one cluster, the at least one cluster comprising one meta cluster and at least one work cluster, the meta cluster carrying operation and maintenance control of the cluster system, the operation and maintenance control comprising work cluster addition and deletion processing, and the work cluster carrying a business container to execute a business task; and according to indication information of the addition and deletion task, controlling the specified work cluster to implement corresponding cluster deployment or removal in the cluster system. According to the present application, the work cluster can be expanded or deleted according to actual needs through the meta cluster, thereby helping to improve the applicability of the cluster system.

The invention discloses a large model reasoning method and system based on active load awareness, and belongs to the technical field of large model reasoning services. Aiming at the problem that reasoning service is difficult to consider resource efficiency and service quality due to burstiness and heterogeneity of mixed loads in the prior art, the invention provides a closed-loop method of cycle planning, online scheduling and self-adaptive adjustment and optimization. In the period planning stage, prospective instance pre-configuration is carried out by utilizing long-term load prediction; in the online scheduling stage, the request is routed and scheduled through response length sensing and a heterogeneous scheduling strategy; and in the self-adaptive tuning stage, fine-grained correction is carried out on the instance scale based on peak working set pre-judgment. According to the method, resource supply and request scheduling are coordinated, the SLO is guaranteed, the delay peak is effectively inhibited, the GPU utilization rate is improved, and the method is suitable for an enterprise-level LMaaS platform.

The present application relates to the technical field of artificial intelligence computing power scheduling, in particular to a load balancing and performance guarantee method for multi-modal AI computing power service, comprising: firstly, analyzing request characteristics, extracting a computing work set and a minimum safe capacity to generate a local key-value cache baseline, and fitting to generate a minimal algebraic arrival curve of network direct writing traffic; then, combining the baseline, the curve and the network topology routing hop count to build a key-value cache non-eviction model and solve the physical cache capacity limit; further, determining the number of isolated paths according to the limit and the single-path capacity, calling the hardware register to set the capacity bitmask, and building a space-time physical isolation architecture; finally, when the traffic breaks through the upper bound of the curve, establishing an orthogonal damping constraint relationship based on equivalent arithmetic strength, and synchronously reducing the computing end micro-batch and the network end transmission unit. The present application solves the problem of physical squeezing of local cache by network direct writing through bottom-layer physical isolation and double-end orthogonal feedback.

Innovations in training a machine learning model using an accelerator pipeline with deferred updates to model parameters are described. An accelerator identifies one or more first-class micro-batches (“MBs”) and a second-class MB of a working set. A first-class MB contains, as inputs, frequently accessed embeddings stored in graphics processing unit (“GPU”) memory. The accelerator schedules the first-class MB(s) for training using one or more GPUs. During the training, the accelerator obtains the second-class MB, which contains, as inputs, non-frequently accessed embeddings stored in main memory. At least some updates to the model parameters from the training with the first-class MB(s) are deferred until after training with the second-class MB. The accelerator schedules the second-class MB for training. Finally, after the training with the second-class MB, the accelerator updates the non-frequently accessed values for the second-class MB. At this point, the GPU(s) also update model parameters, applying deferred updates.

This invention discloses a physically isolated minimal deterministic safety actuator and its control method for intelligent driving, belonging to the field of functional safety technology for intelligent driving. The device adopts a hardware architecture physically independent of the main decision-making unit, possessing independent power supply and timing. Its runtime working set is no larger than 10KB, and it has no operating system or dynamic memory. Deterministic control is achieved through hardware-fixed logic, with a worst-case response time of no more than 50ms. The device is equipped with independent sensing paths and hardwired execution interfaces, enabling direct actuator driving without relying on the vehicle bus. Cross-domain health monitoring enables multi-node status monitoring and computing power management. This invention can provide a safety fallback in the event of a system-wide failure, resisting common-cause failures, meeting ASIL D functional safety level, and features low cost, small size, easy integration, and suitability for mass production. It solves the problems of insufficient independence and poor real-time performance of traditional safety islands.

Antagonistic queries can have a high resource and time footprint triggering a range of issues, such as causing compilation performance degradation of other queries and machine failures. Described herein are techniques for automatically identifying antagonistic queries and redirecting the antagonistic queries to dedicated resources. This redirecting can help better balance the workload on different work clusters and to isolate antagonistic workloads from impacting the compilation and execution performance of other queries.

A task setting system is provided that sets tasks for onboard functional units mounted on an autonomously movable body at a destination of the autonomously movable body. [Solution] The work setting system of the present invention comprises a movable vehicle 2, an on-board function unit 4 that is mounted on the vehicle and executes functions at its destination, a map display unit 42 that displays a map on a user interface showing the area in which the vehicle can move, a point setting unit 44 that sets a point on the map that will become the destination of the vehicle and at which the function of the on-board function unit will be executed, a point function display unit 46 that displays functions that can be executed by the on-board function unit for each point so that they can be selected, and a link function unit 52 that links the execution of the function selected by the point function display unit to the point set by the point setting unit.

The invention relates to a GPU progressive optimization method for neural operator backbone network reasoning, and belongs to the technical field of electronic information. Comprising the steps of S1, initializing constraints and tolerances; s2, baseline measurement is carried out to obtain baseline execution time and a reference result; s3, obtaining a pre-estimated segmentation value based on the working set model and expanding the pre-estimated segmentation value; s4, carrying out KSweep evaluation and consistency check, and carrying out optimization; the process of scanning and measuring the candidate set one by one is called KSweep; s5, fixing, newly adding fusion reasoning optimization, carrying out consistency verification, if the consistency verification is passed, outputting and storing optimal configuration, and directly loading and executing in an operation stage; otherwise, returning or adjusting and recalibrating, and returning to the step S3. According to the method, the multiplexing distance between adjacent layers is shortened, and the round-trip overhead of intermediate activation between the video memory and the cache is reduced, so that the reasoning efficiency is improved.

The invention discloses an embedded system snapshot optimization method and system oriented to a resource-constrained environment. The method comprises the following steps: firstly, configuring snapshot parameters according to hardware parameters of the embedded equipment; secondly, constructing a page state management structure, and continuously capturing page writing events during scanning in a pre-copying stage so as to distinguish a high-writing-frequency page from a low-writing-frequency page; executing multiple rounds of traversal in the pre-copying stage to ensure the convergence of the working set, and judging whether to terminate the pre-copying stage or not according to the convergence; storing the identified working set pages instead of all dirty pages in the shutdown stage again, setting write protection for non-working set pages, and restoring the operation of the virtual machine; after the operation of the virtual machine is recovered, a non-working set is stored in the background; and finally, real-time redundancy elimination and writing are carried out. According to the method, the shutdown duration can be effectively shortened in an embedded scene with dense dynamic write-in and obvious working set change, the snapshot duration and the storage write-in burden are reduced, and the overall performance of an embedded virtualization operation framework is improved.

Systems and techniques for cost-based prefetching are described. In one example, a processor includes a cache system having a hierarchy of one or more cache levels and an instruction prefetcher associated with a cache level of the cache system. The instruction prefetcher encodes fetch requests to a prefetcher table and determines a cost associated with each entry based on a memory level from where the instructions associated with the fetch requests were fetched. The instruction prefetcher evicts entries in the prefetcher table based on the cost associated with one or more corresponding fetch requests in response to adding new entries. The described techniques hide instruction fetch latency that a processor frontend experiences with workloads that have code working sets that do not fit in higher levels of the cache system.

Methods, devices, and systems for information routing. A destination indication corresponding to data is received. The data is transmitted on a routing path based on routing information retrieved from a routing table bypass that includes a routing cache and a match mask. In some implementations, the destination indication comprises at least a portion of a network address. Some implementations include bypassing a routing table based on the routing information being present in the routing cache or the match mask. In some implementations, a total number of entries of the routing cache is fewer than a number of addresses in a working set. In some implementations, the match mask is dynamically programmable. In some implementations, an entry of the match mask includes a bit mask, a bit match and a routing destination associated with at least a portion of an address. In some implementations, the data comprises a packet.