Resource scheduling method, elastic computing system, computing equipment, medium and product

Abstract

The embodiments of this specification provide a resource scheduling method, an elastic computing system, a computing device, a medium, and a product. The method includes: obtaining resource information of system resources in the elastic computing system and event information of system events in the elastic computing system; aligning the timestamps of the resource information and the event information to construct a time series of system events, wherein the time series includes the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp; based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp, performing cross-modal time series fusion processing to obtain time series correlation features, and based on the time series correlation features, predicting resource scheduling strategies for future system events; scheduling system resources in the elastic computing system for future system events based on the resource scheduling strategies. By performing cross-modal time series fusion on the time series, it is possible to learn the dynamic change patterns of resource usage and deep semantic information, thereby improving prediction accuracy.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of cloud computing technology, and in particular to a resource scheduling method, elastic computing system, computing device, medium, and product. Background Art

[0002] Cloud computing is an internet-based computing model that centrally manages computing resources (such as servers, storage, and networks) and provides them to users as services. In cloud computing, users can acquire and release resources at any time based on their needs, flexibly adjusting the scale of computing resources and achieving elastic expansion or contraction. Elastic computing systems use virtualization technology to abstract physical resources into virtual resource pools, from which users can access resources such as computing instances, storage capacity, and network bandwidth on demand. In practical applications, elastic computing systems are widely used in big data processing, artificial intelligence training, web services, and other fields. For example, during AI model training, GPU resources can be dynamically allocated based on the complexity of the training task.

[0003] However, in complex scenarios, user demand for resources often exhibits volatility and uncertainty. Flexible resource demand forecasting using static configuration rules alone is difficult, and can easily lead to resource over-allocation or under-supply in actual operations, impacting service quality. Therefore, a more flexible and accurate resource scheduling method is urgently needed. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a resource scheduling method. One or more embodiments of this specification also relate to a resource scheduling method applied to model training scenarios, an elastic computing system, a computing device, a computer-readable storage medium, and a computer program product to address technical deficiencies in the prior art.

[0005] According to a first aspect of an embodiment of this specification, a resource scheduling method is provided, including:

[0006] Obtaining resource information of system resources in the elastic computing system and event information of system events in the elastic computing system;

[0007] Align resource information and event information with timestamps to construct a time series of system events, where the time series includes the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp;

[0008] Based on the timestamps of system events, resource information associated with the timestamps, and event information associated with the timestamps, cross-modal time series fusion processing is performed to obtain time series correlation features. Based on these time series correlation features, resource scheduling strategies for future system events are predicted.

[0009] Based on resource scheduling strategies, system resources in the elastic computing system are scheduled for future system events.

[0010] In this way, by obtaining resource information of system resources in the elastic computing system and event information of system events in the elastic computing system, resource usage and system status can be obtained in real time, providing a reliable data basis for subsequent predictions; by aligning the timestamps of resource information and event information and constructing a time series of system events, different types of information can be integrated together and aligned in the time dimension, thereby realizing multimodal modeling based on time series; by performing cross-modal time series fusion processing based on timestamps, resource information under timestamps and event information under timestamps, time series correlation features can be obtained, and time series data and text information in the system can be cross-modally fused, so that the system can simultaneously understand the dynamic changes in resource usage and deep semantic information, thereby deeply understanding the root causes of changes in resource demand, and then predicting resource scheduling strategies for future system events and improving the accuracy of prediction results; by scheduling system resources in the elastic computing system based on resource scheduling strategies, dynamic expansion and contraction can be achieved, and the timeliness of resource scheduling can be improved, thereby effectively reducing resource waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a system architecture diagram of an elastic computing system provided by one embodiment of this specification;

[0012] Figure 2 This is a flow chart of a resource scheduling method provided by one embodiment of this specification;

[0013] Figure 3 This is a process flow chart of a resource scheduling method provided by an embodiment of this specification;

[0014] Figure 4 This is a schematic diagram of a processing process of a resource management unit provided in one embodiment of this specification;

[0015] Figure 5 This is a schematic diagram of the structure of a resource scheduling device provided by an embodiment of this specification;

[0016] Figure 6 This is a schematic diagram of the structure of a resource scheduling device applied to a model training scenario provided by an embodiment of this specification;

[0017] Figure 7 This is a structural block diagram of a computing device provided by one embodiment of this specification. DETAILED DESCRIPTION

[0018] The following description sets forth many specific details to facilitate a thorough understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0019] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms "a," "the," and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0020] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0021] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0022] First, the terms involved in one or more embodiments of this specification are explained.

[0023] Temporal Point Process (TPP): is a statistical modeling framework used to analyze event sequences occurring in a continuous time dimension, which can capture the temporal dynamics and type correlation of events.

[0024] Large Language Model (LLM): A deep neural network model pre-trained on massive amounts of text data, capable of text understanding, reasoning, and generation.

[0025] Time series-text fusion representation technology: refers to a cross-modal feature extraction method that converts event timestamps into discrete byte tags and jointly encodes them with text descriptions.

[0026] Dynamic scaling: refers to automatically adjusting the size of the cloud computing cluster based on real-time predicted GPU resource demand, achieving dynamic matching of resource supply and demand.

[0027] Fine-grained resource usage pattern: This refers to identifying micro-fluctuations in GPU resource usage through event time prediction and text semantic association analysis.

[0028] Application Programming Interface (API): It can expose the functions or data of a software system in a specific way for use by other software systems.

[0029] Graphics Processing Unit (GPU): In fields such as deep learning and scientific computing, it can process large amounts of data in parallel and is suitable for tasks that require highly parallel computing.

[0030] The CPU (Central Processing Unit) is the core component of a computer. It is responsible for executing the computer's instruction set, performing complex logical operations, processing data, and coordinating the work of various computer components. Like the computer's "brain," it controls the entire computer system, including booting the operating system, loading and running applications, and storing and retrieving data.

[0031] This specification provides a resource scheduling method. This specification also relates to a resource scheduling method for model training scenarios, an elastic computing system, a computing device, a computer-readable storage medium, and a computer program product, each of which is described in detail in the following embodiments.

[0032] See also Figure 1 , Figure 1 The system architecture diagram of an elastic computing system provided by one embodiment of this specification is shown as follows: The elastic computing system 100 includes a resource management unit 102 and a graphics processing unit 104 .

[0033] Resource management unit 102: used to obtain resource information of the graphics processing unit 104 and event information of system events in the elastic computing system 100; align the resource information and event information with timestamps to construct a time series of model training events, wherein the time series includes the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp; based on the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp, perform cross-modal time series fusion processing to obtain time series correlation features, and based on the time series correlation features, predict the resource scheduling strategy for future system events; and schedule the graphics processing unit 104 based on the resource scheduling strategy.

[0034] Specifically, the elastic computing system 100 may include one or more graphics processing units 104. A graphics processing unit 104 can be specifically understood as a GPU. If the elastic computing system 100 includes multiple graphics processing units 104, these multiple graphics processing units 104 may constitute a GPU cluster. Scheduling a graphics processing unit 104 can be understood as scheduling GPU computing resources. This can include allocating computing cores, allocating video memory, adjusting computing frequency, and adjusting power states, and the specific requirements can be determined based on actual application needs.

[0035] By applying this embodiment, by obtaining resource information of system resources and event information of system events in the elastic computing system 100, resource usage and system status can be obtained in real time, providing a reliable data basis for subsequent predictions. By aligning the timestamps of resource information and event information to construct a time series of system events, different types of information can be integrated and aligned in the time dimension, thereby realizing multimodal modeling based on time series. By performing cross-modal time series fusion processing based on timestamps, resource information under timestamps, and event information under timestamps to obtain time series correlation features, time series data and text information in the system can be cross-modally fused, enabling the resource management unit 102 to simultaneously understand the dynamic changes in resource usage and deep semantic information, thereby deeply understanding the root causes of changes in resource demand, and then predicting resource scheduling strategies for future system events, thereby improving the accuracy of prediction results. By scheduling the graphics processing unit 104 based on the resource scheduling strategy, more intelligent dynamic scaling decisions can be achieved, and the timeliness of resource scheduling can be improved, thereby effectively reducing resource waste and improving the overall utilization efficiency of the GPU cluster.

[0036] See also Figure 2 , Figure 2 A flow chart of a resource scheduling method provided according to an embodiment of this specification is shown, which specifically includes the following steps.

[0037] Step 202: Obtain resource information of system resources in the elastic computing system and event information of system events in the elastic computing system.

[0038] The embodiments of this specification can be applied to an elastic computing system with resource allocation and management functions. The elastic computing system can be an elastic computing system in a cloud computing scenario and can be applied to a model training and deployment platform of a private cloud or a public cloud.

[0039] Specifically, an elastic computing system can be understood as a system that can dynamically adjust computing resources based on computing needs. It can flexibly allocate and reclaim resources to meet computing needs under different loads, improve resource utilization, and improve system responsiveness. System resources can be understood as the various resources available in an elastic computing system, such as hardware resources such as CPU, GPU, memory, disk I / O, and network bandwidth, as well as software resources such as operating systems and databases. Resource information can be understood as information describing the current status and usage of system resources, such as CPU utilization, GPU utilization data, memory usage, disk read and write speeds, etc. System events can be understood as various operations or state changes that occur in an elastic computing system, such as user requests, task start / end, and system failures. Event information can be understood as detailed information related to system events, including the time and type of the event, the resources involved, and so on.

[0040] According to an optional implementation, obtaining resource information of system resources in the elastic computing system may include: collecting resource information of system resources through a system monitoring tool or monitoring software.

[0041] Specifically, the system monitoring tool can be understood as the underlying command provided by the operating system, and the monitoring software can be third-party monitoring software that is called through a specific API.

[0042] Optionally, the system monitoring tool or monitoring software may periodically sample and record resource information and store it in a database or log file.

[0043] According to an optional implementation, obtaining event information of system events in the elastic computing system may include: extracting event information of system events from data sources such as system log files, application logs, and message queues.

[0044] Specifically, system log files and application logs can be considered the system operation logs of an elastic computing system. These logs can record a variety of event information, including task launches and completions, error reports, and exception information. In an elastic computing system, message queues can be used to transmit messages between different components. Message queues can store various types of information, including information related to system events.

[0045] Optionally, in an elastic computing system, when a system event occurs, the relevant component sends a message describing the event to a message queue. These messages typically contain key information such as the event type (e.g., task state change event, resource allocation event), the time of the event, and the objects involved (e.g., specific task ID, resource identifier, etc.). To extract event information from the message queue, a corresponding consumer program must first connect to the message queue. The consumer program filters the messages in the message queue based on pre-defined rules (e.g., message subject, tags, etc.) to identify messages related to system events. The consumer program then parses these qualifying messages to extract specific event information, such as the event timestamp and detailed event description from the message content. This allows accurate acquisition of elastic computing system event information from the message queue.

[0046] Obtaining resource information of system resources in the elastic computing system and event information of system events in the elastic computing system provides a data basis for subsequent resource scheduling strategy prediction.

[0047] Step 204: aligning the timestamps of the resource information and the event information to construct a time sequence of the system event, wherein the time sequence includes the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp.

[0048] In practical applications, after collecting data from elastic computing systems and obtaining resource information and event information, the data can be cleaned and standardized, and the resource information and event information can be timestamped to construct a time series of system events.

[0049] Specifically, timestamp alignment can be understood as matching and synchronizing resource information and event information in chronological order, ensuring that the resource and event information at each point in time correspond. A system event time series can be understood as a collection of system resource and event information arranged in chronological order, with each point in time containing a corresponding timestamp, resource information, and event information. Timestamps can also be understood as points in time, and the specific accuracy can be determined based on actual application requirements.

[0050] According to an optional implementation, aligning the timestamps of resource information and event information may include: extracting the timestamp field from the resource information and event information; sorting the resource information and event information according to the timestamps, and matching the resource information and event information at the same time point or a similar time point based on the timestamp field.

[0051] During the actual implementation process, you can choose to align resource information and event information with time accuracy such as seconds or milliseconds based on actual needs.

[0052] According to an optional implementation, constructing a time sequence of system events may include: for each time stamp, combining the time stamp, resource information associated with the time stamp, and event information associated with the time stamp to obtain a time sequence.

[0053] During the operation of an elastic computing system, time series are continuously updated over time. In practical applications, the size of a time series can be maintained at a preset fixed value, such as 1k or 10k, depending on the specific needs of the application. For example, when the time series size is 1k, the time series can include 1000 sets of timestamps, resource information associated with the timestamps, and event information associated with the timestamps. A time series can be understood as a fixed-size sliding window that continuously slides backward with subsequent events, ensuring that each prediction is based on the most recently collected data.

[0054] Step 206: Based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp, perform cross-modal time series fusion processing to obtain time series correlation features, and predict resource scheduling strategies for future system events based on the time series correlation features.

[0055] In practical applications, based on the construction of time series, cross-modal time series fusion processing can be performed based on the timestamps of system events, resource information under the timestamps, and event information under the timestamps to obtain time series correlation features. Based on the time series correlation features, the resource scheduling strategy for future system events can be predicted.

[0056] Specifically, cross-modal time series fusion can be understood as the cross-modal fusion of different types of information (such as timestamps, resource information, and event information) and the extraction of correlation features between these different types of information. These temporal correlation features can reflect the temporal relationships and patterns of change between system resources and events, such as the correlation between resource usage trends and event frequency. Resource scheduling policies can be understood as plans for rationally allocating and managing system resources based on the current state of the system and predicted future demand. Pre-trained models can be either standard or large models.

[0057] According to an optional implementation, a pre-trained model can be used to perform cross-modal time series fusion processing to obtain time series correlation features, and based on the time series correlation features, the resource scheduling strategy for future system events can be predicted.

[0058] After obtaining temporal correlation features, standard models typically output predicted future system events. After outputting these predictions, additional rules or pre-defined algorithms are needed to determine the corresponding resource scheduling policies. For example, based on past experience, the resource requirements for different event types in different scenarios can be summarized. Combined with the predicted future events, appropriate system resources can be allocated to these events according to established rules. Large models, however, are trained and learned from large amounts of data, enabling them to understand complex contextual information and patterns. Leveraging their powerful generalization capabilities, large models can directly output corresponding resource scheduling policies when presented with temporal correlation features derived from the time series of system events. Large models can perform complex reasoning and calculations internally, comprehensively considering various factors and directly delivering resource scheduling solutions that align with the system's actual conditions. This approach is more efficient and intelligent, better adapting to complex and changing system environments. Even without pre-defined rules, they can execute accurate resource scheduling policies in a timely manner.

[0059] Step 208: Based on the resource scheduling policy, schedule system resources in the elastic computing system for future system events.

[0060] In practical applications, when resource scheduling strategies for future system events are obtained based on the constructed time series predictions, system resources in the elastic computing system can be scheduled based on the resource scheduling strategies, thereby achieving dynamic expansion and contraction of the elastic computing system during operation.

[0061] According to an optional implementation, scheduling system resources in the elastic computing system based on a resource scheduling policy may be implemented through a specific API.

[0062] By applying this embodiment, resource information of system resources in an elastic computing system and event information of system events in the elastic computing system are obtained; timestamps of resource information and event information are aligned to construct a time series of system events, wherein the time series includes the timestamp of the system event, resource information under the timestamp, and event information under the timestamp; based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp, cross-modal time series fusion processing is performed to obtain time series correlation features, and based on the time series correlation features, resource scheduling strategies for future system events are predicted; based on the resource scheduling strategies, system resources in the elastic computing system are scheduled for future system events.

[0063] In this way, by obtaining resource information of system resources in the elastic computing system and event information of system events in the elastic computing system, resource usage and system status can be obtained in real time, providing a reliable data basis for subsequent predictions; by aligning the timestamps of resource information and event information and constructing a time series of system events, different types of information can be integrated together and aligned in the time dimension, thereby realizing multimodal modeling based on time series; by performing cross-modal time series fusion processing based on timestamps, resource information under timestamps and event information under timestamps, time series correlation features can be obtained, and time series data and text information in the system can be cross-modally fused, so that the system can simultaneously understand the dynamic changes in resource usage and deep semantic information, thereby deeply understanding the root causes of changes in resource demand, and then predicting resource scheduling strategies for future system events and improving the accuracy of prediction results; by scheduling system resources in the elastic computing system based on resource scheduling strategies, dynamic expansion and contraction can be achieved, and the timeliness of resource scheduling can be improved, thereby effectively reducing resource waste.

[0064] In an optional embodiment, the event information of the system event in the elastic computing system includes at least one of task execution information, error report, and exception information.

[0065] Specifically, system events can be understood as various operations, state transitions, or specific situations that occur during the operation of an elastic computing system. These events can reflect the system's operating status, performance, and potential issues. Event information can be understood as detailed descriptions of system events. Administrators and systems can often use event information to analyze, diagnose, and take appropriate action. Resource information for system resources in an elastic computing system can include GPU utilization data, specifically metrics such as video memory usage and computing load. Task execution information can be understood as information covering the entire process of a task from start to finish, such as the task's start time, completion time, execution progress, resource usage (such as CPU, memory, GPU, etc.), and the task's execution result (success or failure). This information can be used to understand task execution efficiency and resource consumption. Error reports can be understood as reports containing detailed error information generated when a system error occurs during operation. Error reports typically record the time the error occurred, the error type (such as a program crash, memory overflow, or file read / write error), the location of the error (such as the specific line of code or process ID), and the possible cause of the error, helping to quickly locate and resolve the problem. Exception information can be understood as information about unexpected situations that occur during system operation. Exception information may not be as serious as an error report, but it can still affect normal system operation. For example, a process's runtime may exceed the normal range, or resource utilization may suddenly fluctuate significantly. Exception information can help identify potential system issues in advance.

[0066] Optionally, the context description information of the system event can be composed through the resource information and event information.

[0067] By applying this embodiment, by obtaining task execution information, it is beneficial to subsequently mine the deep correlation between task execution status and timing information, thereby helping to predict the task execution time and resource requirements of future system events and reasonably arrange scheduling strategies; by obtaining error reports and exception information, it is beneficial to mine the deep correlation between potential system problems and timing information, thereby improving the accuracy of prediction of abnormal situations and peak loads and improving the accuracy of resource scheduling.

[0068] In an optional embodiment, based on the timestamp of the system event, the resource information under the timestamp and the event information under the timestamp, cross-modal time series fusion processing is performed to obtain time series correlation features, which may include: encoding the timestamp, the resource information under the timestamp and the event information under the timestamp respectively to obtain timestamp features, type features and text content features; encoding the timestamp to obtain timestamp features, encoding the resource information under the timestamp to obtain type features, encoding the event information under the timestamp to obtain text content features; performing attention feature calculation on the cross-modal time series fusion features to extract time series correlation features.

[0069] Specifically, a timestamp can be understood as a marker indicating the time an event occurred, typically a numerical value accurate to a specific time unit (such as seconds or milliseconds). In this embodiment, the timestamp is used to record the specific moment a system event occurred and is a key element in constructing a time series. Time series correlation features can be understood as features that reflect the temporal correlation and changing patterns of system events. These features can capture the dynamic relationships between system resources, events, and time series, providing a basis for subsequent predictions and decision-making. Resource information associated with a timestamp can include information related to resource usage, such as resource allocation, release, and errors. Since resource usage often corresponds to specific event types, resource information associated with a timestamp can also be understood as an event type. Event information associated with a timestamp can include resource usage metrics and detailed event information (such as task start and end, error messages, and exception content). Therefore, event information associated with a timestamp can also be understood as event context.

[0070] According to an optional embodiment, encoding the timestamp to obtain timestamp features, encoding resource information under the timestamp to obtain type features, and encoding event information under the timestamp to obtain text content features may include:

[0071] Encode the timestamp at the byte level to obtain the timestamp feature; encode the resource information under the timestamp in the text dimension to obtain the type feature; encode the event information under the timestamp in the text dimension to obtain the text content feature.

[0072] It should be noted that byte-level encoding of timestamps is performed to enable alignment and fusion of the encoded timestamp features with the type and content features in the text dimension. Since timestamps are often discrete data, byte-level encoding can be used to obtain timestamp features in the text dimension.

[0073] In actual implementation, because timestamps, resource information associated with timestamps, and event information associated with timestamps are different types of data, the specific encoding processes for each are different. The timestamp features, type features, and text content features obtained through encoding are all features in the text dimension, specifically text tokens.

[0074] According to an optional embodiment, feature fusion processing is performed on the timestamp features, type features, and text content features to obtain cross-modal temporal fusion features. This can be achieved using various methods, such as simple concatenation, weighted summation, and fusion using a fully connected layer. Taking concatenation as an example, the timestamp features, type features, and text content features can be sequentially concatenated into a longer vector to obtain the cross-modal temporal fusion features.

[0075] According to an optional embodiment, performing attention feature calculation on cross-modal temporal fusion features and extracting temporal correlation features may include: using an attention mechanism (such as a multi-head attention mechanism) to calculate the cross-modal temporal fusion features. First, the cross-modal temporal fusion features are input into an attention layer, and the attention weight of each feature element is calculated. Then, the feature elements are weighted and summed according to the attention weights to obtain a feature vector processed by the attention mechanism, i.e., the temporal correlation feature.

[0076] By applying this embodiment, by encoding timestamps, resource information and event information respectively, different types of data can be converted into feature representations under a unified text dimension, thereby realizing multimodal modeling and facilitating subsequent processing and analysis; by performing feature fusion processing on timestamp features, type features and text content features, cross-modal time series fusion features are obtained, attention features are calculated on the cross-modal time series fusion features, and time series correlation features are extracted, which can mine the potential correlation between time, resources and events, thereby overcoming the problem of separation between time series dynamics and text semantic information in the prediction process, and providing more comprehensive information for subsequent predictions and decisions.

[0077] In an optional embodiment, encoding the timestamp to obtain timestamp features, encoding resource information under the timestamp to obtain type features, and encoding event information under the timestamp to obtain text content features may include:

[0078] Discretely encode the timestamp to obtain the timestamp feature; semantically encode the resource information under the timestamp to obtain the type feature; semantically encode the event information under the timestamp to obtain the text content feature.

[0079] According to an optional embodiment, discretely encoding the timestamp to obtain the timestamp feature may include: converting the timestamp into a floating point number; converting the floating point number into a binary representation; and splitting the binary representation to obtain a byte-level encoding representation.

[0080] For example, the timestamp may be converted into a 32-bit floating point number, and the 32-bit floating point number may be further converted into a 4-byte token.

[0081] According to another optional embodiment, discretely encoding the timestamp to obtain the timestamp feature may also include: dividing the timestamp into different time intervals; mapping each timestamp to a corresponding time interval, and performing one-hot encoding on each time interval. For example, the time intervals may be divided by year, month, day, hour, etc.

[0082] According to an optional embodiment, semantically encoding the resource information under the timestamp to obtain the type feature may include: mapping the resource information under the timestamp into a natural language description, and obtaining a text encoding corresponding to the natural language description.

[0083] According to another optional embodiment, semantic encoding is performed on the resource information under the timestamp to obtain the type feature, which may also include: using a hash function to map the resource information under the timestamp to a fixed-length code.

[0084] According to an optional embodiment, semantic encoding the event information at the timestamp to obtain text content features may include: using a word segmenter to process context text corresponding to the event information at the timestamp to obtain text content features.

[0085] According to another optional embodiment, semantic encoding of the event information at the timestamp to obtain text content features may also include: using a pre-trained language model (such as BERT) to encode the event information at the timestamp.

[0086] In the actual implementation process, the timestamp feature obtained by discrete encoding is a byte token, and the type feature and text content feature obtained by semantic encoding the resource information and event information under the timestamp are text tokens.

[0087] By applying this embodiment, the timestamp feature is obtained by discretely encoding the timestamp, and the byte-level encoding strategy can be adopted to encode the timestamp into a byte token; by semantically encoding the resource information under the timestamp to obtain the type feature, and semantically encoding the event information under the timestamp to obtain the text content feature, the resource information and event information can be encoded into text tokens, which is conducive to the cross-modal feature fusion of different types of information in the text dimension. Through multimodal modeling, the deep correlation between different information can be extracted, so that the log text and time series information can be better combined, the root cause of the change in resource demand can be better understood, and the accuracy of prediction and scheduling can be improved.

[0088] In an optional embodiment, predicting resource scheduling strategies for future system events based on timing correlation features may include: predicting the probability distribution of occurrence of different system events based on timing correlation features; predicting future system events based on timing correlation features and event occurrence probability distribution, and generating resource scheduling strategies for future system events.

[0089] Specifically, time series correlation features can be understood as a quantitative representation of the interrelationships and changing patterns between system resources and events along the temporal dimension. These features integrate multiple data points, including timestamps, resource information, and event information, to reflect the dynamic changes in the system over time, providing a foundation for subsequent predictions and decision-making. Future system events can be understood as system events that have not yet occurred but, based on historical system data and current state, have a certain probability of occurring in the future, such as task launches, system failures, and resource bottlenecks. A future system event can be the next pending event corresponding to the currently constructed time series, or it can be a collection of multiple future events. The occurrence probability distribution can be understood as a probabilistic model that describes the likelihood of future system events. It mathematically presents the probability of each possible event, helps quantify uncertainty, and provides a probabilistic basis for resource scheduling. A resource scheduling policy can be understood as a resource scheduling plan for future system events, ensuring the rational allocation and use of system resources.

[0090] According to an optional embodiment, predicting the occurrence probability distribution of different system events based on time series correlation features may include: predicting the occurrence probability distribution of different system events based on time series correlation features through the intensity prediction layer of the resource prediction model.

[0091] According to another optional embodiment, predicting the occurrence probability distribution of different system events based on time series correlation features may also include: predicting the occurrence probability distribution of different system events based on time series correlation features through a long short-term memory network (LSTM).

[0092] Specifically, the intensity prediction layer calculates and outputs the probability of various possible future system events based on the received temporal correlation features. For example, for events such as changes in task load, fluctuations in resource demand, or system failures, the intensity prediction layer provides the probability of these events occurring at different time points, transforming the originally complex temporal correlation features into a quantitative representation of the likelihood of future events.

[0093] In actual implementation, the intensity prediction layer further processes and maps the input temporal correlation features. Through a series of mathematical operations and nonlinear transformations, the data in the feature space is mapped to the probability space, ensuring that the output meets the requirements of the probability distribution, that is, the sum of the probabilities of all possible events is 1, and the probability of each event is between 0 and 1.

[0094] According to an optional embodiment, based on time series correlation characteristics and occurrence probability distribution, predicting future system events and generating resource scheduling strategies for future system events may include: using a resource prediction model, based on time series correlation characteristics and occurrence probability distribution, outputting the next event to occur, the resource demand corresponding to the next event to occur, and the resource scheduling strategy.

[0095] According to another optional embodiment, based on the time series correlation characteristics and the probability distribution of occurrence, future system events are predicted and resource scheduling strategies for future system events are generated, which can also be achieved through rule engines, reinforcement learning, integrated learning, etc.

[0096] Alternatively, a rule engine can be understood as a set of rules defined based on the system's task execution logic and requirements. By inputting temporal correlation features and probability distribution into the rule engine, the engine automatically matches matching rules. Based on these matching rules, it predicts future system events and generates corresponding resource scheduling policies.

[0097] Alternatively, reinforcement learning algorithms (such as Q-learning and deep Q networks) can be used to train the agent to learn the optimal decision-making strategy. In practical applications, the agent predicts future system events based on the current system state and generates corresponding resource scheduling strategies.

[0098] Optionally, you can also select multiple different types of base learners (such as decision trees, support vector machines, etc.); use training data to train each base learner; and synthesize the prediction results of the base learners (such as voting, weighted averaging, etc.) to obtain the final future system event prediction results and resource scheduling strategies.

[0099] Optionally, based on the time series correlation feature, the event type and timestamp of the next event to occur can be output; based on the occurrence probability distribution, the resource demand and resource scheduling strategy corresponding to the next event to occur can be output.

[0100] Specifically, the resource prediction model can be understood as a large model with generalization capabilities, and the next event can be understood as the next predicted event that may occur at the current time point in the current time series. The resource scheduling strategy can include the execution action for the next event.

[0101] It should be noted that the next event to occur may change over time. For example, if the last event in the time series does not change (i.e., no new events are collected), the predicted next event may be different if the time interval between the last event's timestamp and the current time is 10 seconds or 40 seconds.

[0102] According to another optional embodiment, when the resource prediction model is a common model rather than a large model, based on the time series correlation characteristics and the probability distribution of occurrence, future system events are predicted, and a resource scheduling strategy for future system events is generated. It may also include: through the resource prediction model, based on the time series correlation characteristics and the probability distribution of occurrence, outputting the event type and timestamp of the next event to occur; according to the event type and timestamp of the next event to occur, based on the mapping relationship between the event type and the resource demand, determining the resource demand corresponding to the next event to occur; and determining the corresponding resource scheduling strategy based on the event type, timestamp and resource demand of the next event to occur.

[0103] By applying this embodiment, the probability distribution of future system events is predicted based on the time series correlation characteristics, and the next possible system event can be determined based on the probability distribution, so that resources can be planned for the system event in advance and corresponding resource scheduling strategies can be generated, thereby realizing dynamic resource allocation and scheduling, improving the timeliness and flexibility of resource allocation, and through the time series correlation characteristics, more accurate results can be predicted based on the potential correlation between time series information and resources and events, thereby improving the accuracy of resource scheduling.

[0104] In an optional embodiment, the resource scheduling strategy includes the timestamp, event type and resource requirement of the future system event; generating the resource scheduling strategy for the future system event may include: predicting the timestamp and event type of the future system event based on the time series correlation characteristics; and predicting the resource requirement of the future system event based on the probability distribution of occurrence.

[0105] Specifically, the timestamp of a future system event can be understood as the specific time at which a future system event is predicted to occur, which can be accurate to seconds, milliseconds, and other time units. It is used to locate events in the time dimension and help the system prepare for resource allocation in advance. The event type of a future system event can be understood as a classification of system events that may occur in the future, such as task start, task end, system failure, resource expansion request, etc. Different event types have different resource requirements and processing methods. The resource demand of a future system event can be understood as the number of various system resources (such as the number of GPU cores, the number of CPU cores, memory size, disk storage space, network bandwidth, etc.) required during the execution of the predicted future system event.

[0106] According to an optional implementation, predicting the timestamps of future system events and the event types of future system events based on time series correlation features may include: using a resource prediction model to predict the timestamps of future system events and the event types of future system events based on time series correlation features.

[0107] In the actual implementation process, the resource prediction model can output the timestamp token and event type token of future system events based on the time series correlation features. In addition, it can also output the description information token for future system events, etc.

[0108] According to an optional implementation, predicting resource requirements of future system events based on the occurrence probability distribution may include: using a resource prediction model to predict resource requirements of future system events based on the occurrence probability distribution.

[0109] According to an optional implementation, generating a resource scheduling strategy for future system events based on timestamps, event types, and resource requirements may include: utilizing a resource prediction model to output a resource scheduling strategy for future system events based on temporal correlation features and occurrence probability distribution.

[0110] If the resource prediction model is a large model, it can leverage its generalization capabilities to output prediction results such as timestamps, event types, and resource requirements, along with corresponding resource scheduling policies. If the resource prediction model is a standard model, the system can determine a matching resource scheduling policy based on the prediction results, such as timestamps, event types, and resource requirements, output by the resource prediction model.

[0111] Optionally, the predicted timestamps, event types, and resource requirements for future system events can be combined to form a complete resource scheduling policy data structure. For example, a table or JSON-formatted data structure containing these three fields can be constructed. The generated resource scheduling policy can be stored in a database or file system for subsequent system access and execution. Furthermore, the policy can be visualized for system administrators to review and adjust.

[0112] By applying this embodiment, the occurrence time and event type of future events can be accurately predicted through the time series correlation characteristics, so that the required amount of resources can be accurately allocated to specific types of events at the right time, avoiding over-allocation or under-allocation of resources, improving resource utilization, and reducing resource waste; by predicting the resource demand of future system events based on the probability distribution of occurrence, the system can arrange maintenance tasks and prepare backup resources in advance, reduce the impact of system failures on operation, and enhance the stability and reliability of the system.

[0113] In cloud computing scenarios, based on resource prediction results and resource scheduling strategies for future events, corresponding resource expansion and contraction operations can be performed at the appropriate time, thereby enabling early response to possible resource over-allocation or undersupply situations.

[0114] In an optional embodiment, the resource scheduling policy includes resource requirements for future system events; scheduling system resources in the elastic computing system based on the resource scheduling policy may include:

[0115] When the resource demand exceeds the current resource capacity of the system, perform resource expansion operations on the system resources;

[0116] When the resource demand does not exceed the current resource amount of the system resources, a resource scaling operation is performed on the system resources.

[0117] Specifically, the current amount of system resources can be understood as the total amount of various resources actually owned and available by the elastic computing system at the current moment. Examples include the currently available number of GPU cores, remaining memory size, and free disk space. Resource expansion can be understood as a series of actions taken to increase system resources to meet resource demands when the system predicts that future events will require more resources than the current amount. This can include adding physical hardware devices (such as increasing the number of GPU cores, adding server memory modules, adding new hard drives, etc.) or leasing additional virtual resources from a cloud service provider (such as adding virtual machine instances). Resource reduction can be understood as reducing system resources to avoid idle resources when resource demands do not exceed the current amount. Examples include reclaiming excess GPU cores, shutting down some idle virtual machines, freeing up excess memory space, or reducing leased cloud storage capacity.

[0118] According to an optional implementation, when the resource demand exceeds the current resource amount of the system resources, performing a resource expansion operation on the system resources may include: when the resource demand exceeds the current resource amount of the system resources, using a specific API interface of the cloud service provider to expand the resources.

[0119] It's important to note that after adding new resources, they must be configured so that they can be properly recognized and used by the system. For example, a newly added hard drive must be partitioned, formatted, and mounted to the operating system's designated directory. Newly started virtual machines must also undergo network configuration and have necessary software dependencies installed to ensure seamless integration of the new resources with existing system resources.

[0120] According to another optional implementation, a target time point or target time period for executing the resource expansion operation may be determined based on the timestamp of the future event, and the resource expansion operation may be executed when the target time point or target time period is reached.

[0121] According to an optional implementation, for example, if the resource to be reclaimed is a GPU resource, and if the resource demand does not exceed the current resource demand of the system resources, performing a resource scaling operation on the system resources may include: determining the resource to be reclaimed based on the resource scheduling policy; terminating non-critical tasks or processes related to the resource to be reclaimed, and releasing the GPU's video memory and computing resources. Furthermore, the video memory and computing resources may be marked as available for reallocation to other tasks in need.

[0122] According to another optional implementation, a target time point or target time period for executing the resource scaling operation may be determined based on the timestamp of the future event, and the resource scaling operation may be executed when the target time point or target time period is reached.

[0123] This embodiment, through dynamic resource expansion and contraction, ensures that system resources always match actual demand. When resource demand is high, capacity can be expanded promptly to meet system operation requirements; when demand is low, capacity can be contracted to avoid idle resources, significantly improving resource utilization efficiency and reducing operating costs.

[0124] In an optional embodiment, after scheduling system resources in the elastic computing system based on the resource scheduling policy, the method may further include: collecting resource information of system resources in future system events to obtain execution monitoring results for the resource scheduling policy.

[0125] Specifically, resource information for future system events can be understood as the resource usage of system resources in the event of a future system event, including metrics such as GPU utilization. Execution monitoring results can be understood as conclusions about the actual execution effectiveness of resource scheduling policies, obtained through the collection and analysis of resource information from future system events. These conclusions can reflect whether the resource scheduling policy effectively meets the resource requirements of the system event and whether there is any resource waste or shortage.

[0126] According to an optional implementation, collecting resource information of system resources in future system events and obtaining execution monitoring results for resource scheduling strategies may include: determining monitoring indicators; collecting resource information corresponding to the monitoring indicators; comparing the resource information with the predicted resource demand in the resource scheduling strategy to obtain execution monitoring results for the resource scheduling strategy.

[0127] Optionally, determining monitoring indicators may include determining key resource indicators. In practical applications, key resource indicators can be determined based on the specific needs of different scenarios. For example, in a cloud computing scenario, a key resource indicator used in a model training scenario may be a GPU utilization indicator.

[0128] Furthermore, monitoring indicators can be further refined based on different future system event types. For example, for task start events, we can monitor the response time of resource allocation; during task execution, we can monitor the stability and sustainability of resources; for task end events, we can monitor the timeliness of resource release.

[0129] Optionally, resource information corresponding to monitoring indicators can be collected by using system tools, or by deploying third-party monitoring software and using cloud platform APIs.

[0130] Optionally, after collecting resource information corresponding to the monitoring indicators, the collected raw resource data can be cleaned to remove outliers, duplicate data, and erroneous data. The data can then be organized according to dimensions such as chronological order and event type to facilitate subsequent analysis.

[0131] Furthermore, the collected resource information can be compared with the expected resource usage in the resource scheduling policy. The deviation between actual and expected resource usage is calculated to analyze whether resource utilization is within a reasonable range and determine whether the resource scheduling policy has achieved its intended goals. Based on the comparative analysis results, an execution monitoring report can be generated. This report can include various resource usage metrics, comparisons with expectations, the strengths and weaknesses of the resource scheduling policy, and recommendations for improvements to address any deficiencies.

[0132] According to another optional implementation, collecting resource information of system resources in future system events and obtaining execution monitoring results for resource scheduling strategies may also include: using machine learning algorithms to perform anomaly detection on system resource information and evaluating the execution effect of resource scheduling strategies.

[0133] Optionally, historical resource information about system resources in future system events can be collected and divided into training and test sets. The data can be standardized to ensure consistency. An appropriate anomaly detection algorithm, such as Isolation Forest or Local Outlier Factor (LOF), can be selected. The model can be trained using the training set data. The collected real-time resource information is fed into the trained model to detect anomalies. Detecting anomalies indicates a potential problem with the resource scheduling strategy. The time, type, and severity of the anomaly can be analyzed to generate an execution monitoring results report.

[0134] By applying this embodiment, by collecting resource information and generating execution monitoring results, the effect of the resource scheduling strategy in actual application can be intuitively understood. It is clear whether the strategy has successfully guaranteed the resource requirements of future system events, providing a strong basis for the optimization and adjustment of the strategy. Based on the execution monitoring results, the resource scheduling strategy can be optimized to make resource allocation more accurate and reasonable, improve resource utilization, and reduce costs. By real-time monitoring of the operating status of system resources in future system events, it is also beneficial to timely discover potential system failures or performance bottlenecks. Once abnormal fluctuations in resource usage are found or exceed the normal range, measures can be taken quickly to make adjustments to avoid system failures due to resource problems and ensure the stable operation of the system.

[0135] In an optional embodiment, based on the timestamp of the system event, the resource information under the timestamp and the event information under the timestamp, cross-modal time series fusion processing is performed to obtain time series correlation features, and based on the time series correlation features, the resource scheduling strategy of the future system events is predicted. It can include: inputting the time series of the system event into the resource prediction model, and performing cross-modal time series fusion processing based on the timestamp of the system event, the resource information under the timestamp and the event information under the timestamp by the resource prediction model to obtain time series correlation features, and based on the time series correlation features, predicting the resource scheduling strategy of the future system events.

[0136] Specifically, the resource prediction model is a pre-trained model, which can be a common model or a large language model.

[0137] According to an optional implementation, the resource prediction model can be trained through the following steps:

[0138] Obtaining a sample time series and a tag resource scheduling policy corresponding to the sample time series, wherein the sample time series includes a sample timestamp of a system event of the elastic computer system, resource information under the sample timestamp, and event information under the sample timestamp;

[0139] The sample timestamp, resource information under the sample timestamp, and event information under the sample timestamp are input into the initial resource prediction model, and cross-modal time series fusion processing is performed to obtain sample time series correlation features. Based on the sample time series correlation features, the predicted resource scheduling strategy for future system events is predicted; based on the label resource scheduling strategy and the predicted resource scheduling strategy, the initial resource prediction model is trained to obtain a trained resource prediction model.

[0140] Furthermore, predicting a prediction resource scheduling strategy for future system events based on sample time series correlation features may include: predicting a prediction resource scheduling strategy for future system events and an event probability distribution of future system events based on sample time series correlation features.

[0141] Accordingly, based on the label resource scheduling strategy and the prediction resource scheduling strategy, training the initial resource prediction model to obtain the trained resource prediction model can include: calculating the first loss value based on the label resource scheduling strategy and the prediction resource scheduling strategy; calculating the second loss value based on the event probability distribution and the sample time series; training the initial resource prediction model according to the first loss value and the second loss value to obtain the trained resource prediction model.

[0142] According to an optional implementation, the reasoning process of the resource prediction model may include: receiving a time series of system events, performing cross-modal time series fusion processing based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp to obtain time series correlation features; and predicting the resource scheduling strategy for future system events based on the time series correlation features.

[0143] In practical applications, the resource prediction model can output prediction results for the sample time series and calculate the first loss value based on the prediction results and labels; the resource prediction model can also output the event probability distribution corresponding to the sample time series, and calculate the second loss value based on the event probability distribution and the target event probability distribution corresponding to the sample time series, so as to complete the model training by adjusting parameters based on the first loss value and the second loss value.

[0144] By applying this embodiment, through pre-training the resource prediction model, the time series process can be deeply integrated with the model, breaking through the limitations of the traditional GPU resource scheduling system in which time series data and text information are separated. By uniformly encoding multimodal information such as GPU usage changes, system logs and error reports into event sequences, and with the help of a specially designed byte-level time encoding mechanism, the system can simultaneously understand the dynamic changes in resource usage and deep semantic information. This fusion processing method not only overcomes the defects of existing solutions that rely only on single numerical predictions or static rules, but also can deeply understand the root causes of changes in resource demand by analyzing text information such as task descriptions and error logs. The model can accurately capture abnormal situations and load peaks during training, predict changes in resource demand in advance, and thus achieve more intelligent dynamic expansion and contraction decisions, improve the accuracy of resource predictions and the timeliness of scheduling, effectively reduce resource waste, and improve the overall utilization efficiency of the GPU cluster.

[0145] The following combined Figure 3 , taking the application of the resource scheduling method provided in this specification in the model training scenario as an example, the resource scheduling method is further explained. Figure 3 A flowchart of a resource scheduling method provided by an embodiment of the present specification is shown, which is applied to a resource management unit of an elastic computing system and specifically includes the following steps.

[0146] Step 302: Obtain resource information of a graphics processing unit in the elastic computing system and event information of a model training event in the elastic computing system.

[0147] Step 304: Align the timestamps of the resource information and the event information to construct a time series of the model training event, wherein the time series includes the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp.

[0148] Step 306: Based on the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp, perform cross-modal time series fusion processing to obtain time series correlation features, and based on the time series correlation features, predict the resource scheduling strategy for future model training events.

[0149] Step 308: Based on the resource scheduling policy, schedule the graphics processing unit in the elastic computing system for future model training events.

[0150] Specifically, the resource management unit can be understood as a key module or component in an elastic computing system responsible for unified management and scheduling of various resources. It can be specifically used to manage graphics processing units (GPUs). An elastic computing system may include one or more graphics processing units (GPUs). Specifically, a graphics processing unit can be understood as a GPU. If an elastic computing system includes multiple GPUs, these multiple GPUs can form a GPU cluster within the elastic computing system. Scheduling GPUs can be understood as scheduling GPU computing resources, which can include allocating computing cores, allocating video memory, adjusting computing frequency, and adjusting power states. The specific requirements can be determined based on actual application needs. A model training event can be understood as any event that may occur during the processing of model training tasks using the elastic computing system. These include new training task initiation events, training task completion events, training task pause events, training task resumption events, abnormal training task interruption events (such as hardware failure or software crash), insufficient training task resource warning events, training data loading failure events, and performance bottleneck triggering events during training (such as memory overflow or excessive network latency).

[0151] It should be noted that the specific implementation of steps 302 to 308 is the same as that of the above-mentioned steps 202 to 208. The specific implementation of steps 302 to 308 can refer to steps 202 to 208, and will not be repeated in this specification.

[0152] By applying this embodiment, multimodal information such as GPU usage changes, system logs, and error reports are uniformly encoded into an event sequence, and with the help of a specially designed byte-level time encoding mechanism, the system can simultaneously understand the dynamic changes in resource usage and deep semantic information. This fusion processing method not only overcomes the defects of existing solutions that rely only on single numerical predictions or static rules, but also can deeply understand the root causes of changes in resource demand by analyzing text information such as task descriptions and error logs. The model can accurately capture abnormal situations and load peaks during training, predict changes in resource demand in advance, and thus achieve more intelligent dynamic expansion and contraction decisions, improve the accuracy of resource predictions and the timeliness of scheduling, effectively reduce resource waste, and improve the overall utilization efficiency of GPU clusters.

[0153] See also Figure 4 , Figure 4 The resource management unit 400 includes a processing layer 402 , an output layer 404 , and an execution layer 406 .

[0154] Processing layer 402: used to collect GPU usage data and system operation log data in the elastic computing system; clean and standardize the GPU usage data and system operation log data, and align them according to time series; construct an event sequence based on timestamp, event type and event description information; encode the timestamp, event type and event description respectively to obtain timestamp features, type features and text content features, and input the timestamp features, type features and text content features into the resource prediction model; the resource prediction model performs cross-modal feature fusion processing on the timestamp features, type features and text content features to obtain time series correlation features, and obtains event probability distribution based on the time series correlation features.

[0155] Specifically, GPU usage data can include metrics such as video memory usage and computing load; system operation log data can include event information such as task startup, completion, and errors, as well as error reports and exception information. Each event in the constructed event sequence contains three core elements: a timestamp, which records the precise time of the event; an event type, which can reflect information such as resource allocation, release, and errors; and an event description, which contains detailed contextual information about the event.

[0156] Optionally, the resource prediction model can be understood as a large, trained language model. The resource prediction model can include a pre-trained encoder, decoder, hidden state extraction layer, and intensity prediction layer. Optionally, the decoder can use a QwenLM decoder, or other decoders can be used based on actual application requirements. This specification does not impose any restrictions on this.

[0157] In practical applications, the resource prediction model can use an encoder to adopt a byte-level encoding strategy for timestamps, converting 32-bit floating-point numbers into 4-byte tokens; map different types of system events to natural language descriptions to obtain text tokens corresponding to the event types; and use the built-in word segmenter of the large language model to process log text and obtain text tokens corresponding to event description information.

[0158] Furthermore, the resource prediction model can align and fuse timestamp features, type features, and text content features in the text dimension to obtain cross-modal temporal fusion features. The hidden state extraction layer then extracts the temporal dependencies between time, type, and contextual text content in the cross-modal temporal fusion features to obtain temporal correlation features. The resource prediction model can also estimate the probability distribution of future events through the intensity prediction layer to obtain event probability distributions.

[0159] Output layer 404: used to predict the next event and resource demand according to the temporal dependency, obtain the next event prediction result and resource demand prediction result, and output dynamic scheduling decision based on the next event prediction result and resource demand prediction result.

[0160] Specifically, the next event prediction result may include the timestamp and event type of the next event. The resource demand prediction result may include the resource demand corresponding to the next event. The dynamic scheduling decision may include the resource scheduling operation to be performed for the next event, which may specifically include resource expansion or resource reduction operations.

[0161] Execution layer 406: triggers resource expansion in advance when high load is predicted; performs resource recovery in a timely manner when load reduction is predicted; and monitors the effect of policy execution for the next event.

[0162] For example, in a large-scale model training scenario, the system may encounter the following sequence of events:

[0163] Receive a new training task start event, including information such as model scale and dataset size;

[0164] A rapid increase in GPU usage was detected.

[0165] Capture warning logs caused by insufficient video memory;

[0166] The resource management unit automatically predicts the next resource demand trend;

[0167] Start new GPU resource allocation in advance before resource shortage occurs;

[0168] After the training task is completed, it is predicted that the resource demand will decrease and resources will be automatically recycled.

[0169] By applying this embodiment, by deeply integrating the time series process with the large language model, it is possible to break through the limitations of the traditional GPU resource scheduling system in which time series data and text information are processed separately. By uniformly encoding multimodal information such as GPU usage changes, system logs and error reports into event sequences, and using a specially designed byte-level time encoding mechanism, the system can simultaneously understand the dynamic changes in resource usage and deep semantic information. This fusion processing method not only overcomes the defects of existing solutions that rely only on single numerical predictions or static rules, but also can deeply understand the root causes of changes in resource demand by analyzing text information such as task descriptions and error logs. The model can accurately capture abnormal situations and load peaks during training, predict changes in resource demand in advance, and thus achieve more intelligent dynamic expansion and contraction decisions, improve the accuracy of resource predictions and the timeliness of scheduling, effectively reduce resource waste, and improve the overall utilization efficiency of GPU clusters.

[0170] Corresponding to the above method embodiment, this specification also provides a resource scheduling device embodiment, Figure 5 FIG. 1 shows a schematic diagram of the structure of a resource scheduling device provided by an embodiment of this specification. Figure 5 As shown, the device includes:

[0171] The first acquisition module 502 is configured to acquire resource information of system resources in the elastic computing system and event information of system events in the elastic computing system.

[0172] The first constructing module 504 is configured to align the timestamps of the resource information and the event information to construct a time sequence of the system event, wherein the time sequence includes the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp.

[0173] The first prediction module 506 is configured to perform cross-modal time series fusion processing based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp, obtain time series correlation features, and predict the resource scheduling strategy of future system events based on the time series correlation features.

[0174] The first scheduling module 508 is configured to schedule system resources in the elastic computing system for future system events based on a resource scheduling policy.

[0175] Optionally, the first prediction module 506 is further configured to: encode the timestamp to obtain the timestamp feature, encode the resource information under the timestamp to obtain the type feature, encode the event information under the timestamp to obtain the text content feature; perform feature fusion processing on the timestamp feature, type feature and text content feature to obtain cross-modal time series fusion feature; perform attention feature calculation on the cross-modal time series fusion feature to extract time series correlation feature.

[0176] Optionally, the first prediction module 506 is further configured to: discretely encode the timestamp to obtain timestamp features; semantically encode the resource information under the timestamp to obtain type features; and semantically encode the event information under the timestamp to obtain text content features.

[0177] Optionally, the first prediction module 506 is further configured to: predict the occurrence probability distribution of different system events based on the time series correlation characteristics; predict future system events based on the time series correlation characteristics and the occurrence probability distribution, and generate resource scheduling strategies for future system events.

[0178] Optionally, the resource scheduling strategy includes the timestamp, event type and resource requirement of future system events; the first prediction module 506 is further configured to: predict the timestamp and event type of future system events based on time series correlation characteristics; and predict the resource requirement of future system events based on the occurrence probability distribution.

[0179] Optionally, the resource scheduling strategy includes the resource demand of future system events; the first scheduling module 508 is further configured to: when the resource demand exceeds the current resource amount of the system resources, perform a resource expansion operation on the system resources; when the resource demand does not exceed the current resource demand of the system resources, perform a resource reduction operation on the system resources.

[0180] Optionally, the resource scheduling device further includes a monitoring module configured to collect resource information of system resources in future system events and obtain execution monitoring results for the resource scheduling strategy.

[0181] Optionally, the event information of the system event in the elastic computing system includes at least one of task execution information, error report and exception information.

[0182] Optionally, the first prediction module 506 is further configured to: input the time series of system events into the resource prediction model, perform cross-modal time series fusion processing based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp, obtain time series correlation features, and predict the resource scheduling strategy of future system events based on the time series correlation features.

[0183] By applying this embodiment, system resources in an elastic computing system are scheduled based on a resource scheduling policy, which can achieve dynamic expansion and contraction, improve the timeliness of resource scheduling, and thus effectively reduce resource waste.

[0184] The above is a schematic scheme of a resource scheduling device of this embodiment. It should be noted that the technical scheme of the resource scheduling device and the technical scheme of the resource scheduling method described above are of the same concept. For details not described in detail in the technical scheme of the resource scheduling device, please refer to the description of the technical scheme of the resource scheduling method described above.

[0185] Corresponding to the above method embodiment, this specification also provides an embodiment of a resource scheduling device applied to a model training scenario, Figure 6 FIG1 shows a schematic diagram of a resource scheduling device for a model training scenario provided by an embodiment of the present specification. Figure 6 As shown, the device includes:

[0186] The second acquisition module 602 is configured to acquire resource information of a graphics processing unit in the elastic computing system and event information of a system event in the elastic computing system.

[0187] The second construction module 604 is configured to align the timestamps of the resource information and the event information to construct a time series of the model training event, wherein the time series includes the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp.

[0188] The second prediction module 606 is configured to perform cross-modal time series fusion processing based on the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp, obtain time series correlation features, and predict the resource scheduling strategy of future system events based on the time series correlation features.

[0189] The second scheduling module 608 is configured to schedule the graphics processing unit in the elastic computing system based on the resource scheduling policy.

[0190] By applying this embodiment, multimodal information such as GPU usage changes, system logs, and error reports are uniformly encoded into an event sequence, and with the help of a specially designed byte-level time encoding mechanism, the system can simultaneously understand the dynamic changes in resource usage and deep semantic information. This fusion processing method not only overcomes the defects of existing solutions that rely only on single numerical predictions or static rules, but also can deeply understand the root causes of changes in resource demand by analyzing text information such as task descriptions and error logs. The model can accurately capture abnormal situations and load peaks during training, predict changes in resource demand in advance, and thus achieve more intelligent dynamic expansion and contraction decisions, improve the accuracy of resource predictions and the timeliness of scheduling, effectively reduce resource waste, and improve the overall utilization efficiency of GPU clusters.

[0191] The above is a schematic scheme of a resource scheduling device for a model training scenario according to this embodiment. It should be noted that the technical scheme of the resource scheduling device for a model training scenario and the technical scheme of the resource scheduling method for a model training scenario are based on the same concept. For details not described in detail in the technical scheme of the resource scheduling device for a model training scenario, please refer to the description of the technical scheme of the resource scheduling method for a model training scenario.

[0192] Figure 7 7 shows a block diagram of a computing device 700 according to one embodiment of the present disclosure. Components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0193] Computing device 700 also includes an access device 740 that enables computing device 700 to communicate via one or more networks 760. Examples of such networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. Access device 740 may include one or more of any type of network interface (e.g., a network interface controller (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, or a near field communication (NFC) interface.

[0194] In one embodiment of the present specification, the above components of the computing device 700 and Figure 7 Other components not shown in the figure may also be connected to each other, for example, via a bus. Figure 7 The computing device structure block diagram shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art may add or replace other components as needed.

[0195] Computing device 700 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, personal digital assistant, laptop computer, notebook computer, netbook computer, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or personal computer (PC). Computing device 700 can also be a mobile or stationary server.

[0196] The processor 720 is configured to execute the following computer-executable instructions, which implement the steps of the above method when executed by the processor.

[0197] The above is a schematic solution of a computing device of this embodiment. It should be noted that the technical solution of the computing device and the technical solution of the above method belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the above method.

[0198] An embodiment of the present specification further provides a computer-readable storage medium storing computer-executable instructions, which implement the steps of the above method when executed by a processor.

[0199] The above is a schematic solution of a computer-readable storage medium of this embodiment. It should be noted that the technical solution of the storage medium and the technical solution of the above method belong to the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the above method.

[0200] An embodiment of the present specification further provides a computer program product, comprising a computer program / instruction, which implements the steps of the above method when executed by a processor.

[0201] The above is an illustrative solution of a computer program product of this embodiment. It should be noted that the technical solution of the computer program product and the technical solution of the above method belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the above method.

[0202] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0203] The computer instructions include computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium may be appropriately increased or decreased based on the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0204] It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0205] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0206] The preferred embodiments disclosed above are intended only to help illustrate this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made based on the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that those skilled in the art can better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A resource scheduling method, comprising: Acquire resource information of system resources in the elastic computing system and event information of system events in the elastic computing system, wherein the resource information is information describing the current state and usage of the system resources; Performing time stamp alignment on the resource information and the event information to construct a time sequence of the system event, wherein the time sequence includes the time stamp of the system event, the resource information under the time stamp, and the event information under the time stamp; Based on the timestamp of the system event, the resource information under the timestamp, and the event information under the timestamp, cross-modal time series fusion processing is performed to obtain time series correlation features, and based on the time series correlation features, resource scheduling strategies for future system events are predicted, wherein the resource scheduling strategies for future system events based on the time series correlation features include: predicting the occurrence probability distribution of different system events based on the time series correlation features; predicting the future system events based on the time series correlation features and the occurrence probability distribution, and generating resource scheduling strategies for the future system events; Based on the resource scheduling policy, system resources in the elastic computing system are scheduled for the future system event.

2. The method according to claim 1, wherein the performing cross-modal time series fusion processing based on the timestamp of the system event, the resource information associated with the timestamp, and the event information associated with the timestamp to obtain time series correlation features comprises: Encoding the timestamp to obtain timestamp features, encoding resource information under the timestamp to obtain type features, and encoding event information under the timestamp to obtain text content features; Performing feature fusion processing on the timestamp feature, the type feature, and the text content feature to obtain a cross-modal temporal fusion feature; Attention features are calculated on the cross-modal temporal fusion features to extract the temporal correlation features.

3. The method according to claim 2, wherein encoding the timestamp to obtain timestamp features, encoding resource information under the timestamp to obtain type features, and encoding event information under the timestamp to obtain text content features, comprises: Discretely encode the timestamp to obtain a timestamp feature; Performing semantic encoding on the resource information under the timestamp to obtain type features; Semantic encoding is performed on the event information under the timestamp to obtain text content features.

4. The method according to claim 1, wherein the resource scheduling policy includes a timestamp, event type, and resource requirement of the future system event; The resource scheduling strategy for generating the future system event includes: Predicting the timestamp of the future system event and the event type of the future system event based on the time series correlation feature; Based on the occurrence probability distribution, resource requirements of the future system events are predicted.

5. The method according to claim 1, wherein the resource scheduling strategy includes resource requirements of the future system events; Scheduling system resources in the elastic computing system based on the resource scheduling policy includes: When the resource demand exceeds the current resource amount of the system resource, performing a resource expansion operation on the system resource; In a case where the resource demand does not exceed the current resource amount of the system resource, a resource scaling operation is performed on the system resource.

6. The method according to claim 1, after scheduling system resources in the elastic computing system based on the resource scheduling policy, further comprising: Resource information of the system resources in the future system events is collected to obtain an execution monitoring result of the resource scheduling strategy. 7 . The method according to claim 1 , wherein the event information of the system event in the elastic computing system includes at least one of task execution information, error report and exception information.

8. The method according to claim 1, wherein the performing cross-modal time series fusion processing based on the timestamp of the system event, the resource information associated with the timestamp, and the event information associated with the timestamp to obtain time series correlation features, and predicting resource scheduling strategies for future system events based on the time series correlation features, comprises: The time series of the system events is input into the resource prediction model, and the resource prediction model performs cross-modal time series fusion processing based on the timestamp of the system events, the resource information under the timestamp and the event information under the timestamp to obtain time series correlation features. Based on the time series correlation features, the resource scheduling strategy of future system events is predicted.

9. A resource scheduling method for a model training scenario, applied to a resource management unit of an elastic computing system, comprising: Obtaining resource information of a graphics processing unit in an elastic computing system and event information of a model training event in the elastic computing system, wherein the resource information is information describing a current state and usage of the graphics processing unit; Performing timestamp alignment on the resource information and the event information to construct a time series of the model training event, wherein the time series includes the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp; Based on the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp, cross-modal time series fusion processing is performed to obtain time series correlation features, and based on the time series correlation features, resource scheduling strategies for future model training events are predicted, wherein the resource scheduling strategies for future system events are predicted based on the time series correlation features, including: predicting the occurrence probability distribution of different system events based on the time series correlation features; predicting the future system events based on the time series correlation features and the occurrence probability distribution, and generating resource scheduling strategies for the future system events; Based on the resource scheduling policy, a graphics processing unit in the elastic computing system is scheduled for the future model training event.

10. An elastic computing system comprising a resource management unit and a graphics processing unit; The resource management unit is used to obtain resource information of the graphics processing unit and event information of the model training event in the elastic computing system, wherein: The resource information is information describing the current status and usage of the system resources; The resource information and the event information are timestamp aligned to construct a time series of the model training event, wherein the time series includes the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp; based on the timestamp of the model training event, the resource information under the timestamp, and the event information under the timestamp, cross-modal time series fusion processing is performed to obtain time series correlation features, and based on the time series correlation features, the resource scheduling strategy of future model training events is predicted, wherein the resource scheduling strategy of future system events is predicted based on the time series correlation features, including: predicting the occurrence probability distribution of different system events based on the time series correlation features; predicting the future system events based on the time series correlation features and the occurrence probability distribution, and generating a resource scheduling strategy for the future system events; scheduling the graphics processing unit for the future model training event based on the resource scheduling strategy.

11. A computing device comprising: memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, the steps of the method according to any one of claims 1 to 9 are implemented.

12. A computer-readable storage medium storing a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

13. A computer program product comprising a computer program / instruction, which implements the steps of the method according to any one of claims 1 to 9 when executed by a processor.

Images