Interference-aware scheduling system based on task classification and interference scoring

Abstract

This invention provides an interference-aware scheduling system based on task classification and interference scoring, including a BE task classifier, a host interference quantifier, and an interference-aware scheduler. The task classifier infers the category of BE task instances to obtain the task category. The host interference quantifier evaluates the interference level of candidate hosts based on the BE distribution on the target host to be scheduled, obtaining the interference level of different candidate hosts. The interference-aware scheduler schedules BE task instances to the candidate hosts corresponding to the specified interference level. This invention performs interference-aware scheduling by predicting new task categories and quantifying the contribution of different categories to performance interference, overcoming the difficulties in studying and analyzing performance interference between applications and mitigating the interference to existing services when scheduling new tasks in large-scale, highly dynamic cluster environments.

Description

Technical Field

[0001] This invention relates to the fields of task scheduling and data mining technology, specifically to an interference-aware scheduling system based on task classification and interference scoring. Background Technology

[0002] In today's large-scale data centers, applications of different priorities are often deployed together on the same host to improve overall resource utilization. A single physical host simultaneously runs user-facing, latency-critical online services (LC services) and a large number of low-priority, best-effort batch processing jobs (BE jobs). The resources used by LC services are positively correlated with queries per second (QPS). Typically, user requests fluctuate periodically, dropping to a minimum around midnight, freeing up a significant amount of idle resources. To utilize the underutilized resources of LC services, BE jobs are scheduled to run on the same host. BE jobs are mostly compute-intensive batch processing tasks with no strict performance requirements and are permissible to have significant latency or be terminated and restarted midway.

[0003] Deploying latency-sensitive applications and batch processing tasks together on the same machine is a primary method for improving resource utilization in large data centers. However, due to resource contention between applications, this co-deployment can introduce performance interference issues. Performance interference reduces the effective utilization of resources and is detrimental to all running applications, especially LC services with stringent Quality of Service (QoS) requirements.

[0004] Existing performance interference studies typically begin with paired application deployments and predict runtime performance in finite-scale scenarios by pre-characterizing the characteristics of applications under controlled-intensity interference. However, for large data centers, the interference problem is highly complex. The diversity of running applications, the uncertainty of interference sources, and the variability of host resource utilization all increase the difficulty of modeling and analyzing performance. Interference changes dynamically with cluster states. Therefore, previous analyses of specific application combinations under finite constraints have not performed well in real-world production clusters. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide an interference-aware scheduling system based on task classification and interference scoring. This system comprehensively considers the characteristics of different types of batch processing tasks and the varying performance changes of multiple online services on a single host when subjected to interference. It enables the scheduling of batch processing jobs in large-scale data centers to reduce interference with online services. Furthermore, this system can effectively reduce the performance impact on online services by scheduling new batch processing tasks in an interference-aware manner based on information from historically run batch processing tasks.

[0006] This invention provides an interference-aware scheduling system based on task classification and interference scoring, including a BE task classifier, a host interference quantifier, and an interference-aware scheduler.

[0007] The task classifier performs category inference on BE task instances to obtain task categories;

[0008] The host interference quantizer evaluates the interference level of candidate hosts based on the BE distribution on the target host to be scheduled, and obtains the interference level of different candidate hosts.

[0009] The interference-aware scheduler schedules BE task instances to candidate hosts corresponding to the specified interference level.

[0010] Preferably, the BE distribution includes all BE task instances running on the host that contain the task category.

[0011] Preferably, it also includes a training data storage and a BE distributed storage;

[0012] The training data storage device is used to store historical data for training the task classifier;

[0013] The BE distribution memory is used to store the BE distribution of the host.

[0014] Preferably, the task categories include: CPU utilization, memory utilization, and the duration of the task instance.

[0015] Preferably, the task classifier infers the category of a BE task instance based on its features using a multi-classification model, thereby obtaining the task category. The features of the BE task instance include the similarity of the DAG and metadata.

[0016] Preferably, the host interference quantizer includes interference classification for LC services, interference level quantization for hosts, and a prediction quantizer;

[0017] The interference classification of the LC service divides the interference into several levels on an average basis according to the actual scenario, so as to achieve a balance between the accuracy and predictability of interference assessment.

[0018] The interference level quantification of the host is based on the weighting of the host interference situation according to different LC services, and the candidate host interference score is obtained.

[0019] The predictive quantizer outputs the interference level based on the interference score.

[0020] Preferably, the interference score Host Defined as:

[0021] ScoreHost =kDIV(∑(w) i *Score i ))

[0022] Where i represents a service instance on the host, w i Using weights, the kDIV() function represents the process of classifying LC service interference levels.

[0023] Preferably, the interference average division divides the performance indicators of each LC service into several levels based on the corresponding percentiles.

[0024] Preferably, the weight is set as the correlation coefficient between the response time of the LC service and the total CPU utilization of the BE task instances on the host.

[0025] Preferably, the interference-aware scheduling system can be integrated into other schedulers.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. The scheduling system proposed in this invention is designed based on patterns discovered in real large-scale data centers. This system is suitable for scenarios where different types of tasks are deployed in a large-scale mixed manner, avoiding the complex process of pre-modeling each application.

[0028] 2. Compared with existing mainstream cluster interference-aware scheduling algorithms, this invention fully considers the diversity and complexity of applications in large-scale scenarios, analyzes interference problems in scheduling scenarios that are common in clusters, and avoids the limitations of modeling specific application combinations or specific interference sources.

[0029] 3. Compared with existing mainstream cluster interference-aware scheduling algorithms, this invention pays more attention to the universality of interference and the versatility of interference-aware algorithms, enabling the system to achieve significant results in dynamic and complex scenarios.

[0030] 4. This invention performs interference-aware scheduling by predicting new task categories and quantifying the contribution of different categories to performance interference. This overcomes the difficulties in studying and analyzing performance interference between applications and mitigating the interference to existing services when scheduling new tasks in a large-scale, highly dynamic cluster environment.

[0031] 5. Based on the conventional scheduling algorithm, this invention adds interference scoring to candidate hosts to schedule tasks in a way that minimizes interference. This allows the BE task instances to be scheduled to run on the machines with the lowest interference levels, effectively reducing the proportion of machines with high interference levels and increasing the proportion of machines with low interference levels, thereby reducing the average interference intensity of the entire cluster. Attached Figure Description

[0032] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0033] Figure 1 This is a schematic diagram of the logical flow of the interference perception and scheduling system based on task classification and interference scoring of the present invention.

[0034] Figure 2 This is a distribution map of host resource usage collected during the research and development of this invention;

[0035] Figure 3 This invention focuses on the different performance characteristics of the LC service under conditions with and without BE operation;

[0036] Figure 4 This serves as the classification basis for the BE task classifier implemented in this invention;

[0037] Figure 5 The prediction result is the interference quantizer implemented in this invention;

[0038] Figure 6 The scheduling result is the result of the scheduler implemented in this invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0040] This invention provides an interference-aware scheduling system based on task classification and interference scoring, including a BE task classifier, a host interference quantizer, and an interference-aware scheduler; it also includes a training memory and a BE distribution memory; wherein, the training memory is used to store historical data for training the task classifier; and the BE distribution memory is used to store the BE distribution of the host.

[0041] The task classifier infers the category of BE task instances to obtain the task category; the task category includes: CPU utilization, memory utilization, and the duration of the task instance. Thus, each BE task instance can be represented by the following triples:

[0042] [CPU,MEM,Makespan]

[0043] CPU, MEM, and Makespan represent CPU utilization, memory utilization, and the duration of the BE task instance, respectively.

[0044] The task classifier infers the category of a BE task instance based on its features using a multi-classification model. The features of a BE task instance include the similarity of its Directed Acyclic Graph (DAG) and metadata. Metadata such as submitter, submission time, and the DAG (the dependencies between tasks in a job) are available when a new BE task instance is submitted. Furthermore, the similarity of the DAG is determined using the Weisfeiler-Lehman Kernel, a graph kernel-based machine learning method. After mapping the DAG to vectors, the similarity between two DAGs is judged, serving as a crucial prerequisite for determining whether two jobs are the same task.

[0045] Furthermore, the similarity of the DAG and the metadata of the task are combined as features of the BE task. These features are then used to train a BE task classifier, which includes three SVM models to classify CPU, MEM and Makespan respectively, and finally obtains the triplet of task category.

[0046] Specifically, this invention analyzes Business Environment (BE) task instances from the perspective of resource usage and categorizes different BE task instances into multiple classes. The resource contention effects of tasks of the same type can be considered similar, thus simplifying the interference effect of each BE task instance to the interference effect of each class. Furthermore, considering the cyclical nature of BE task instances commonly found in production environments, a graph algorithm is used to predict which category a newly submitted BE task instance belongs to, thereby assessing the magnitude of interference that running this task will cause to applications on the corresponding host. This invention employs a multi-classification method to evaluate the scoring model of interference experienced by online services; this model can be used to measure the degree of performance interference experienced by online services or the overall host.

[0047] The host interference quantizer evaluates the interference level of candidate hosts based on the BE distribution on the target host, obtaining the interference level of different candidate hosts. The BE distribution includes all BE task instances running on the host, containing task categories. The host interference quantizer includes interference classification for the LC service, interference level quantization for hosts, and a prediction quantizer.

[0048] The interference classification of LC services divides interference into several levels on an average basis according to the actual scenario, so as to balance the accuracy and predictability of interference assessment; the average interference classification divides the performance indicators of each LC service into several levels on an average basis according to the corresponding percentile.

[0049] Specifically, the interference is divided into k levels, where k represents the granularity of quantization. This granularity can be adjusted according to the actual scenario. If only the presence or absence of interference is determined (k=2), the prediction accuracy will be high, but it will be detrimental to scheduling decisions. Conversely, if k is too large, although the interference can be quantified in more detail, accurate prediction will be difficult.

[0050] Interference is categorized into k levels based on the percentile values ​​of each LC service's performance metric, response time. Each 100 / k percentile represents level 1, and the `kDIV()` function represents this categorization process. First, the total response time of the target LC service is retrieved from historical data. Then, the 100 / k percentile is calculated. Response times less than the 100 / k percentile are classified as interference level 1. Next, the 2*100 / k percentile is calculated. Response times greater than or equal to 100 / k but less than 2*100 / k percentiles are classified as interference level 2, and so on. The pseudocode for `kDIV()` is as follows:

[0051]

[0052] The interference level quantification of a host depends on all LC services running on the host. Since different services have varying sensitivities to interference, their weights in contributing to the host's interference level differ. Therefore, host interference level quantification assigns corresponding weights based on the interference levels of different LC services, obtaining candidate host interference scores; the weights are set as the correlation coefficient between the response time of the LC service and the total CPU utilization of the BE task instances on the host. Interference Score Host Defined as:

[0053] Score Host =kDIV(∑(w) i *Score i ))

[0054] Where i represents a service instance on the host, w i Using weights, the kDIV() function represents the process of classifying LC service interference levels.

[0055] Furthermore, after interference partitioning and weight contribution by the LC service, the performance of different BE task instances is mapped to interference scores between [1, k] using the kDIV() function.

[0056] The predictive quantizer outputs the interference level based on the interference score. This invention can train decision tree models as predictive quantizers at both the service level and the host level to predict interference. Specifically, the input of the decision tree model is the BE distribution on the host, and the output is the interference level of a service or a host.

[0057] The interference-aware scheduler schedules BE task instances to candidate hosts corresponding to a specified interference level. Specifically, the interference level can be set according to the actual situation, generally using the lowest interference level. This ensures that the BE task instances to be scheduled are run on the machines with the lowest interference level, effectively reducing the proportion of machines with high interference levels and increasing the proportion of machines with low interference levels, thereby reducing the average interference intensity of the entire cluster.

[0058] Furthermore, this system can be seamlessly integrated into any other scheduler, such as load balancing schedulers and resource reservation schedulers. When integrated with other schedulers, it is only necessary to first use the other scheduler to select a certain number of candidate hosts, then use the host interference quantifier of this invention to score their interference levels, and finally schedule the task instance to the host with the lowest score, thus achieving interference-aware scheduling. This can be dynamically applied to different scenarios and is superior in terms of flexibility. Compared with existing mainstream cluster interference-aware scheduling algorithms, this invention pays more attention to the universality of interference and the versatility of interference-aware algorithms, thus demonstrating that this invention can also have significant effects in dynamically changing and complex scenarios.

[0059] The working principle of this invention is as follows:

[0060] This invention discloses an interference-aware scheduling system based on task classification and interference scoring, comprising three modules: a BE task classifier, a host interference quantizer, and an interference-aware scheduler. Additionally, it includes two data storage devices. One stores historical data used to train the classifier. The other stores the host BE distribution, i.e., the composition of BE task instances running on each host, which can be quickly accessed by the interference quantizer. The BE distribution is updated only when a new BE task instance is deployed or when a running instance completes.

[0061] In addressing interference issues, this invention primarily focuses on the interference experienced by LC services, as their QoS must be guaranteed. Furthermore, on the same host, LC services are simultaneously affected by other LC services and BE jobs. Since LC services are typically tied to a subset of CPU cores, and BE jobs consume significant host resources, BE jobs are more likely to compete for resources with LC services. Therefore, BE jobs, i.e., BE task instances, are the primary root cause of interference with LC services.

[0062] like Figure 1As shown, the workflow of the Interference Aware Scheduling System (PISCO) based on task classification and interference scoring in this invention is as follows: Users submit BE jobs to the BE job queue, awaiting scheduling; when a BE task is ready to run, this framework first uses a BE task classifier to infer its category; after obtaining the task category and target host, if one or more instances of the task are ready to be scheduled to that host, the host interference quantifier evaluates the host's interference level based on the BE distribution on the host; by evaluating a set of candidate hosts, the interference-aware scheduler can schedule BE task instances to the candidate host with the least interference. This invention can be seamlessly integrated into any scheduler, requiring only interference evaluation after other schedulers provide candidate hosts.

[0063] 1. Interference Source Analysis

[0064] Figure 2 The figure shows the overall CPU utilization of all hosts in the cluster, as well as the individual CPU utilization of the LC service and BE tasks, collected during the research and development of this invention. The host CPU utilization ranges from 20% to 80%, primarily driven by the BE tasks. The LC service's CPU utilization is below 20%, with very little variation. The BE tasks are the main factor affecting host CPU utilization, and the two are essentially linearly correlated.

[0065] Figure 3 The diagram illustrates the performance of the LC service, a focus of this invention, with and without BE tasks running. When BE tasks are disabled in the cluster, the LC service performance remains stable. However, when BE jobs consume a large amount of resources without restriction, the LC service suffers severe interference, resulting in significant performance fluctuations and a pronounced long-tail phenomenon. Therefore, resource-intensive BE tasks are the primary source of performance degradation for the LC service. This invention focuses on the impact of BE tasks on the LC service.

[0066] 2. BE Task Classifier

[0067] Analysis of cluster data reveals that the resource usage of BE tasks fluctuates frequently during their operation and is easily affected by other applications on the same host, making it difficult to accurately estimate BE resource usage. Conversely, estimating the resource usage range of BE tasks (e.g., one or two CPU cores) is highly feasible. This invention measures three performance metrics: CPU utilization, memory utilization, and task instance duration, as these represent the CPU and memory pressure generated by the task instance. Based on the characteristics of the studied cluster environment, this invention categorizes CPU into 3 categories, memory into 5 categories, and duration into 4 categories, which can be used to... Figure 4As observed, in different types of tasks (SQL / Dup / SQLRT / Algo), the dividing points of each category are also inflection points on the cumulative distribution (CDF) graph, indicating a small number of tasks at these points. Specifically, CPU usage is divided into three categories (0, 1, and 2) based on two thresholds: 0.5 and 1.5. That is, when a BE task instance uses less than or equal to 0.5 CPU cores, it belongs to category 0; when it uses more than 0.5 cores but less than or equal to 1.5 cores, it belongs to category 1; and when it uses more than 1.5 cores, it belongs to category 2. This classification method is derived from the distribution of CPU usage by BE tasks. Near the thresholds of 0.5 and 1.5, the number of BE tasks is small, reducing the probability of tasks being classified into different categories due to natural fluctuations during repeated program execution, thus achieving the best accuracy in predicting BE tasks. Memory is divided into 5 categories (0, 1, 2, 3, 4) based on 4 thresholds: 0.5GB, 1GB, 2GB, and 6GB; and duration (Makespan) is divided into 4 categories based on 3 thresholds: 10s, 20s, and 60s.

[0068] Each BE task can be represented by the following triple:

[0069] [CPU,MEM,Makespan]

[0070] CPU, MEM, and Makespan represent CPU utilization, memory utilization, and the duration of a task instance, respectively.

[0071] In modern large-scale clusters, the repetitive execution of business logic (BE) jobs is a common feature, which can be used to classify new BE tasks. When a new BE job is submitted, its metadata, such as the submitter, submission time, and DAG (Directed Acyclic Graph, i.e., the dependencies between each task in the job), is readily available. This invention uses the Weisfeiler-Lehman Kernel, a machine learning method based on the Graph Kernel, to map the DAG into vectors and then determine the similarity between two DAGs, serving as an important prerequisite for determining whether two jobs are the same task. Combining the similarity of the DAG and the task's metadata as features of the BE task, these features are then used to train a BE task classifier. This classifier includes three simple multi-classification models (SVM is used in this invention) to classify CPU, MEM, and Makespan respectively, finally obtaining a triplet of task categories.

[0072] Based on a task classification scheme, this invention abstracts the BE environment (number of BE tasks of different types) on a host into a BE distribution, encompassing all running task instances on this machine to assess their interference with the LC service on the same host. The host BE distribution is represented as follows:

[0073] [n1,n2,…,n 60 ]

[0074] Where n i This represents the number of task instances of category i on the host, with each category i corresponding to a triplet.

[0075] 3. Host interference quantizer

[0076] (1) Interference level of LC service

[0077] To balance the accuracy and predictability of interference assessment, this invention divides interference into k levels on average, where k represents the granularity of quantification. This granularity can be adjusted according to the actual scenario. If only the presence or absence of interference is determined (k=2), the prediction accuracy will be high, but it will be detrimental to scheduling decisions. Conversely, if k is too large, although interference can be quantified in more detail, accurate prediction will be difficult. This invention uses k=10.

[0078] Interference partitioning divides the performance metric (response time) of each LC service into k levels based on the corresponding percentiles. Each 100 / k percentile represents one level, and the kDIV() function represents this partitioning process. After interference partitioning, the performance of different services is mapped to interference scores between [1, k].

[0079] (2) Interference level of the host

[0080] Quantifying host interference depends on all LC services running on the host. Since different services have varying sensitivities to interference, their weights in contributing to the host interference level differ. Based on analysis of large-scale clusters, this invention selects the correlation coefficient between the response time of LC services and the total CPU utilization of BE tasks on the host as the weight. This correlation coefficient remains relatively stable for a given service. The weight w for each service instance running on the host is determined. i Then, the host's interference score. Host Defined as:

[0081] Score Host =kDIV(∑(w) i *Score i ))

[0082] Where i includes all service instances on a host, and the kDIV() function represents the process of classifying LC service interference levels.

[0083] (3) Predictive quantizer

[0084] This invention trains decision tree models as quantizers at both the service level and host level to predict interference. The model's input is the BE distribution on a host, and the output is the interference level for a service or a host. Figure 5 As can be seen, the accuracy of the quantizer implemented by this invention is such that the service-level interference prediction results are concentrated near the actual results, and the prediction will basically not have serious errors.

[0085] 4. Interference-aware scheduler

[0086] The scheduler implemented in this invention adds a scoring step for candidate hosts to the original cluster scheduling. Through interference prediction using a quantizer, this invention schedules the BE (Breakpoint and Entry) tasks to the machine with the lowest interference level. Figure 6 The scheduling interference scoring results show that this invention can effectively reduce the proportion of machines with high interference levels and increase the proportion of machines with low interference levels, thereby reducing the average interference intensity of the entire cluster. This invention can be seamlessly integrated into any other scheduler, such as load balancing schedulers and resource reservation schedulers. To integrate this invention with other schedulers, it is only necessary to first use other schedulers to select a certain number of candidate hosts, then use the host interference quantifier of this invention to score their interference levels, and finally schedule the task instances to the host with the lowest score.

[0087] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0088] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An interference-aware scheduling system based on task classification and interference scoring, characterized in that, This includes a BE task classifier, a host interference quantizer, and an interference-aware scheduler. The task classifier performs category inference on BE task instances to obtain task categories; The host interference quantizer evaluates the interference level of candidate hosts based on the BE distribution on the target host to be scheduled, and obtains the interference level of different candidate hosts. The interference-aware scheduler schedules BE task instances to candidate hosts corresponding to the specified interference level, so that the BE task instances to be scheduled are scheduled to run on the machine with the lowest interference level. The host interference quantizer includes interference classification for LC services, interference level quantization for hosts, and a prediction quantizer. The interference classification of the LC service divides the interference into k levels on average according to the actual scenario, so as to balance the accuracy and predictability of interference assessment. k=10, where k represents the granularity of quantification. The interference level quantification of the host is based on the weighting of the host interference situation according to different LC services, and the candidate host interference score is obtained. The predictive quantizer outputs the interference level based on the interference score.

2. The interference sensing and scheduling system based on task classification and interference scoring according to claim 1, characterized in that, The BE distribution includes all BE task instances running on the host that contain the task category.

3. The interference sensing and scheduling system based on task classification and interference scoring according to claim 1, characterized in that, It also includes training data storage and BE distributed storage; The training data storage device is used to store historical data for training the task classifier; The BE distribution memory is used to store the BE distribution of the host.

4. The interference sensing and scheduling system based on task classification and interference scoring according to claim 1, characterized in that, The task categories include: CPU utilization, memory utilization, and the duration of the task instance.

5. The interference sensing and scheduling system based on task classification and interference scoring according to claim 1, characterized in that, The task classifier, based on the features of BE task instances, uses a multi-classification model to infer the category of BE task instances, thereby obtaining the task category. The characteristics of the BE task instance include DAG similarity and metadata.

6. The interference sensing and scheduling system based on task classification and interference scoring according to claim 5, characterized in that, The interference score Defined as: in For a specific service instance on the host, The weight for each service instance running on the host. The function represents the process of classifying LC service interference levels; Interference classification divides the performance metrics of each LC service, i.e., response time, into k levels based on the corresponding percentiles. Each 100 / k% is level 1. First, the total response time of the target LC service is obtained from historical data. Then, the 100 / k percentile is calculated. All response times less than the 100 / k percentile are considered interference level 1. Then, the 2*100 / k percentile is calculated. Percentiles greater than or equal to 100 / k but less than 2*100 / k percentiles are considered interference level 2, and so on.

7. The interference sensing and scheduling system based on task classification and interference scoring according to claim 5, characterized in that, The interference average division divides the performance indicators of each LC service into several levels based on the corresponding percentiles.

8. The interference sensing and scheduling system based on task classification and interference scoring according to claim 5, characterized in that, The weight is set as the correlation coefficient between the response time of the LC service and the total CPU utilization of the BE task instances on the host.

9. The interference sensing and scheduling system based on task classification and interference scoring according to claim 1, characterized in that, The interference-aware scheduling system can be integrated into other schedulers.

Images