CN114466057A - Automatic optimization method for service configuration based on space search - Google Patents

Abstract

The invention relates to a service configuration automatic optimization method based on space search. Firstly, acquiring candidate configuration items by using a gridding sampling method under a given configuration constraint condition; then, running test is carried out on the target application, a performance result corresponding to configuration is collected, and the performance result is converted into scalar performance measurement through a utility function; and finally, searching to obtain the configuration with the optimal performance by using a recursive bounded search method. Therefore, under the condition of resource limitation, automatic configuration and optimization are realized for wide cloud application and service.

Description

Automatic optimization method for service configuration based on space search

Technical Field

The invention relates to a service configuration automatic optimization method based on space search, and belongs to the technical field of software.

Background

The number and scale of cloud services are increasing, the types are diversified, and users need to face more and more configuration parameters. The large number of configuration parameters leads to the increasing complexity of configuration, which puts a great strain on users, developers and administrators, and configuration errors are more likely to occur, thereby causing performance degradation. Therefore, in order to mine the performance potential of the service and guarantee the service quality, a user needs to find a reasonable configuration through configuration optimization. However, it becomes difficult to adjust the conventional service configuration due to the following problems. Firstly, cloud services have diversity, including data analysis services, database services, and business logic processing services, and various services can be deployed and configured in a cloud computing environment. Also, user related performance goals may be throughput, latency, run time, etc., some need to be maximized and some need to be minimized. In the literature (Dana Van Aken, et al 2017. Automatic Database Management System Tuning Through Large-scale Machine learning. In Proceedings of the 2017 ACM International Conference on Management of data. ACM, 1009 + 1024.) only the Database Management System and the response delay performance objective are considered, but In a cloud computing environment, the configuration Tuning process must also take into account various combinations of services, performance objectives and loads of the application. Second, cloud services have complexity. Given different performance goals and different loads, a deployed service may have different performance for a given set of configuration parameters, and different services may have complex performance. Third, configuring cloud services has a higher operational overhead. Configuration tuning involves solving the problem of high dimensional parameter space, typically requiring large sample sets, since there is no performance simulator for regular services, samples can only be generated by testing deployed services. The literature (Yuq Zhu, et al 2017. ACTS In Need: Automatic Configuration Tuning with scaling guidelines. In Proceedings of the 8th SIGOPS Asia-Pacific Workshop systems. ACM.) uses Configuration data obtained In a small amount of simulation environment for optimization, but In the cloud computing actual deployment environment, the Configuration Tuning sample collection and optimization process has higher computation and time overhead. The literature (chuning Tang, et al 2015. horizontal configuration management at facebook. In Proceedings of the 25th Symposium on Operating Systems principles. ACM 328. 343.) proposes a specific automatic configuration optimization method for distributed application Systems, but has no versatility for various cloud services and is difficult to solve the above challenges In a cloud computing environment. In particular, statistical or machine learning models are proposed for distributed systems, but these models are usually developed based on simulation environments, are not suitable for complex cloud computing environments with high-dimensional parameters, and rarely consider the computing overhead of actual environments.

Disclosure of Invention

The purpose of the invention is as follows: in order to meet the performance optimization requirements of users on cloud applications and services and simplify the optimization problem, the configuration capable of optimizing the performance of the deployed system under the condition of specific application programs and loads is automatically searched within given resource limits.

The principle of the invention is as follows: the resource limit is the number of tests allowed to run during the tuning process, provided as input to the tuning process. Other inputs include the configuration parameter set and its lower/upper bound as configuration constraints. And combining a plurality of performance optimization requirements into a maximization target by adopting a utility function method. The tuning process is in a closed loop state and can run in any number of cycles allowed by resource constraints. The output is the configuration, with the best performance within given resource constraints.

The technical scheme of the invention is as follows: a service configuration automatic optimization method based on space search is characterized by comprising the following implementation steps:

in a first step, given configuration constraints, a number of configurations are generated based on resource limitations, and then the configurations of the target application are updated using these configurations. All types of parameters can be handled, including boolean, enumerated, and numeric. The generated samples must cover a wide spatial range of parameters. To guarantee the scalability of the resource, the sampling method must also guarantee a better coverage of the entire parameter space if the user allows to run more tuning tests. Therefore, the sampling method generates a sample set that must satisfy the following three conditions: (1) the sample set has wide coverage on a high-dimensional space of configuration parameters; (2) the set is small enough to meet resource constraints and reduce test cost; (3) if the resource constraints are expanded, the set can be scaled to have greater coverage.

To ensure broad coverage of the high dimensional parameter space, the present invention divides the space into subspaces and then randomly selects a point from each subspace, each subspace being represented by one sample. For random sampling without subspace partitioning, it is likely that some subspaces are not represented, especially when the dimensionality of the space is high. Given n parameters, the invention can divide the range of each parameter into k intervals and collect combinations of these intervals, for a total of kn combinations, and thus kn subspaces and samples. Through subspace partitioning, gridding ensures complete coverage of the entire parameter space. It also produces a sample set with a larger cardinality that is exponential to the parameter dimension.

The invention reduces the number of subspaces to be sampled. The impact of the value of one influencing parameter on the performance can be demonstrated by a comparison of the performance without taking into account the values of the other parameters. The present invention does not require checking all combinations of parameter values with other parameter values. Instead, it is only necessary to check once which value of each parameter is likely to stand out and compare the resulting performance with other samples. Thus, in the resource-limited case, each interval of one parameter is considered only once, rather than a complete combination of all intervals. After dividing the parameter range into k intervals, not all intervals are fully combined. Instead, the intervals for each parameter are arranged, and then the interval arrangement for each parameter is aligned and k samples are obtained. For a given sample set size, the sample point set is maximally dispersed by representing each interval represented by each parameter exactly once.

Since the configuration tuning process is in a closed loop state, multiple samples can be run. For scalability and coverage, a new sampling process is not restarted by repartitioning the entire space. In contrast, in the case of resampling, the original divisions to the entire sample space and the sample points in the subspace that were not considered before are reused, while the sample points are spread out as much as possible. Although a continuous range of parameters is required, it can be applied to a parameter of a boolean type or a category type by converting the type to a parameter having a continuous range of values.

Second, for each configuration, a test is run on the target application and corresponding performance results are collected, which are then converted to scalar performance metrics by a utility function. The optimization is performed for a scalar performance indicator having only a single value, the performance indicator being defined by a utility function with a user focused performance goal as input. The utility function is an identification function if only one performance objective is of interest, such as throughput or latency. If multiple performance goals are considered simultaneously, the utility function may be defined as a weighted sum. The user can define and implement the utility function of the user through a performance interface.

And thirdly, using all the sample pairs of the performance measurement and the corresponding configuration by the performance optimization algorithm to find the configuration with the best performance. If the resource constraints allow more tests and sampling, the optimization algorithm will record the found configuration and output a new set of configuration constraints for the next tuning cycle. Otherwise, the tuning process will end, outputting the best performing configuration found so far. By maximizing the performance metric defined by the utility function, the user may have a performance goal that needs to be minimized, but may translate the minimization problem into a maximization problem. The performance index is maximized according to a given number of samples. Requiring that output configuration settings must improve system performance over given configuration settings, which may be default settings or manually adjusted by the user. In order to optimize the output of a function/system, the optimization algorithm must satisfy the following conditions: (1) even if the number of samples is limited, answers can be found; (2) if more sample sets are provided, a better answer can be found; (3) if there are enough resources, it does not fall into a local sub-optimal area and it is possible to find a global optimum.

Since the performance map has parameters with values and continuous ranges, the performance surface is a continuous surface. Given a continuous surface, it is likely that other points with similar or better performance will be found near the point with the best performance in the sample set. The above observation holds true even though the continuous performance surface may not be smooth, or is continuous only when projected into a particular dimension or constrained to a particular subspace. On this basis, the invention gives an initial sample set, finds the best performing point, then samples another set of points in a bounded space around the changed point, and samples in a limited space around it. This bounded sampling step is performed recursively until no better performing point is found in the sample set.

There is a problem in this process how large a bounded space should be. If the parameter value has a positive effect on the performance, he should also have a higher performance in the sample set. By recursively performing the boundary sampling step at the best performing point in the sample set, the best performing point can be found even on a limited sample set. Each boundary sampling step is a round, and the size and the number of rounds of the sample set can be adjusted to meet the resource limitation requirement. If a greater number of tests are allowed, then more boundary sampling steps may be performed to perform a finer search in the possible bounded subspace. To avoid trapping in a suboptimal bounded subspace, if no better performing point can be found within the boundary, this measure can find a better answer when providing a larger sample set by using a sampler to sample in the complete parameter space, thus restarting the search.

Compared with the prior art, the invention has the following advantages: the method is applied to cloud computing environment-oriented application and service, automatic configuration and tuning are achieved, and a dispersed sampling method and a recursive boundary searching method are used, so that the conventional system can be configured and tuned under the condition that resources are limited. The method can be widely applied to various cloud applications and services, has low time overhead, and automatically realizes configuration and tuning without manual participation.

Drawings

Fig. 1 is a technical route diagram of an automatic optimization method for service configuration based on spatial search.

Detailed Description

The present invention is described in detail below with reference to specific embodiments and accompanying drawings, as shown in fig. 1, a method flow of the embodiments of the present invention:

to handle the variety of deployed systems and loads, the system architecture design has flexibility in being composed of numerous loosely coupled components connected by data streams. The main components include a configuration sampler, a performance optimizer, a system manipulator and a load generator.

A sampler is configured: implementing an extensible sampling method, inputting a configuration set to be sampled to a system manipulator;

a performance optimizer: a telescopic optimization algorithm is realized, and a new configuration constraint condition is adaptively input into a configuration sampler;

a system manipulator: updating the configuration of the target application, monitoring and testing the state of the target application, operating the target application, and inputting a performance configuration pair sample into a performance optimizer;

a load generator: the generated application workload may be a benchmark test system or a test system provided by the user, and the actual application workload may be regenerated.

The components in the architecture are loosely coupled, interacting with each other only through configuration constraints, configuration and performance metric data streams, allowing different scalable sampling methods and scalable optimization algorithms to be inserted into the configuration tuning process. Only the system manipulator and the load generator need to be adjusted to handle different target applications or loads. With an extensible architecture, future-appearing systems can be optimized with only minor changes to the system manipulator and load generator. The system manipulator component is an interface to interact with a target application deployed in a target environment, while the load generator component allows easy input of any target load.

Claims (1)

1. The method is characterized by comprising the following implementation steps:

the method comprises the steps of firstly, giving configuration constraints, generating a plurality of configurations according to resource limitations, then using the configurations to update the configuration of a target application, and using a gridding sampling method to obtain candidate configuration items;

secondly, for each configuration, running and testing a target application, collecting corresponding performance results, and converting the performance results into scalar performance measurement through a utility function;

and thirdly, using all the sample pairs of the performance measurement and the corresponding configuration by the performance optimization algorithm, and searching by using a recursive bounded search method to obtain the configuration with the optimal performance.

Images