Dynamically meeting performance objectives for multiple services on a service mesh Forough Shahab Samaniyand Rolf Stadlery

2025-05-03 0 0 935.05KB 9 页 10玖币

侵权投诉

Dynamically meeting performance objectives for

multiple services on a service mesh

Forough Shahab Samani †and Rolf Stadler†

†Dept. of Computer Science, KTH Royal Institute of Technology, Sweden

Email: {foro, stadler}@kth.se

October 11, 2022

Abstract—We present a framework that lets a service provider

achieve end-to-end management objectives under varying load.

Dynamic control actions are performed by a reinforcement

learning (RL) agent. Our work includes experimentation and

evaluation on a laboratory testbed where we have implemented

basic information services on a service mesh supported by the

Istio and Kubernetes platforms. We investigate different manage-

ment objectives that include end-to-end delay bounds on service

requests, throughput objectives, and service differentiation. These

objectives are mapped onto reward functions that an RL agent

learns to optimize, by executing control actions, namely, request

routing and request blocking. We compute the control policies not

on the testbed, but in a simulator, which speeds up the learning

process by orders of magnitude. In our approach, the system

model is learned on the testbed; it is then used to instantiate

the simulator, which produces near-optimal control policies for

various management objectives. The learned policies are then

evaluated on the testbed using unseen load patterns.

Index Terms—Performance management, reinforcement learn-

ing, service mesh, digital twin

I. INTRODUCTION

End-to-end performance objectives for a service are difﬁcult

to achieve on a shared and virtualized infrastructure. This

is because the service load often changes in an operational

environment, and service platforms do not offer strict resource

isolation, so that the resource consumption of various tasks

running on a platform inﬂuences the service quality.

In order to continuously meet performance objectives for a

service, such as bounds on delays or throughput for service

requests, the management system must dynamically perform

control actions that re-allocate the resources of the infras-

tructure. Such control actions can be taken on the physical,

virtualization, or service layer, and they include horizontal and

vertical scaling of compute resources, function placement, as

well as request routing and request dropping.

The service abstraction we consider in this paper is a

directed graph, where the nodes represent processing functions

and the links communication channels. This general abstrac-

tion covers a variety of services and applications, such as a

network slice on a network substrate, a service chain on a

softwarized network, a micro-service based application, or a

pipeline of machine-learning tasks. We choose the service-

mesh abstraction in this work and apply it to micro-service

based applications.

In this paper, we propose a framework for achieving end-

to-end management objectives for multiple services that con-

currently execute on a service mesh. We apply reinforcement

learning (RL) techniques to train an agent that periodically per-

forms control actions to reallocate resources. A management

objective in this framework is expressed through the reward

function in the RL setup. We develop and evaluate the frame-

work using a laboratory testbed where we run information

services on a service mesh, supported by the Istio and Kuber-

netes platforms [1],[2]. We investigate different management

objectives that include end-to-end delay bounds on service

requests, throughput objectives, and service differentiation.

Training an RL agent in an operational environment (or

on a testbed in our case) is generally not feasible due to

the long training time, which can extend to weeks, unless

the state and action spaces of the agent are very limited.

We address this issue by computing the control policies in

a simulator rather than on the testbed, which speeds up the

learning process by orders of magnitude for the scenarios

we study. In our approach, the RL system model is learned

from testbed measurements; it is then used to instantiate

the simulator, which produces near-optimal control policies

for various management objectives, possibly in parallel. The

learned policies are then evaluated on the testbed using unseen

load patterns (i.e. patterns the agent has not been trained on).

We make two contributions with this paper. First, we present

an RL-based framework that computes near-optimal control

policies for end-to-end performance objectives on a service

graph. This framework simultaneously supports several ser-

vices with different performance objectives and several types

of control operations.

Second, as part of this framework, we introduce a simulator

component that efﬁciently produces the policies. Through

experimentation, we study the tradeoff of using the simulator

versus learning the policies on a testbed. We ﬁnd that while

we lose some control effectiveness due to the inaccuracy of

the system model we gain by signiﬁcantly shortening the

training time, which makes the approach suitable in practice.

To the best of our knowledge, this paper is the ﬁrst to

advocate a simulation phase a part of implementing a dynamic

performance management solution on a real system.

Note that, when developing and presenting our framework,

we aim at simplicity, clarity, and rigorous treatment, which

helps us focus on the main ideas. For this reason, we choose

a small service mesh for our scenarios (which still includes

key complexities of larger ones), we consider only two types

arXiv:2210.04002v1 [cs.LG] 8 Oct 2022

of control actions, etc. Our plan is to reﬁne and extend the

framework in future work, as we lay out in the last section of

the paper.

II. PROBLEM FORMULATION AND APPROACH

TABLE 1

Notation for formalizing the problem.

Concept Notation

Service index i

Node index j

Offered load of service i li

Carried load of service i lc

i=li(1 −bi)

Utility of service i ui

Response time objective Oi

Response time di

Routing weight of service i pij

towards node j

Blocking rate bi

We consider application services built from microservice

components, whereby each component performs a unique

function. A service request traverses a path on a directed graph

whose nodes offer microservices. Figure 1(a) shows a directed

graph with a general topology of a microservice architecture,

which we call a service mesh. A service is realized as a

contiguous subgraph on the service mesh.

While the framework we present in this paper applies to the

general service mesh of Figure 1(a), we focus the discussion

and experimentation on the smaller conﬁguration shown in

Figure 1(b). An incoming service request from a client is

processed ﬁrst by the front node, before being routed to one

of the two processing nodes. The processing result is sent to

the responder node which sends it to the client.

We consider two services Siwith different resource require-

ments on the processing nodes. In the conﬁguration in Figure

1(b), both processing nodes can process requests from both

services. We express the service quality in terms of end-to-

end delay diof a service request, the carried load of a service

i(which we also call the throughput of the service), or the

utility uigenerated by a service.

In order to control the service quality, our framework

includes control actions. In this work, we consider request

routing expressed by pij ∈[0,1] which indicates the fraction

of requests of service irouted to processing node j. In

addition, we consider blocking service requests at the front

node at the fraction bi∈[0,1].

Central to this work are management objectives, which

capture the end-to-end performance objectives for the services

on a given service mesh. These objectives include client

requirements, as well as provider priorities. To present and

evaluate our framework, we focus on the following three

management objectives. See Table 1 for notation.

Management Objective 1 (MO1): the response time of a

request of service iis upper bounded by Oiand the overall

carried load is maximized:

maximize X

iwhile di< Oi(1)

(a) Directed graph of a service mesh

(b) Use case in this work

Fig. 1. (a) General service mesh; each processing node runs one or more

microservices; links are communication channels for service requests. (b)

Service mesh of the use case; dij is the processing delay of a request of

service ion node j;pij is the routing weight of request itowards node j.

Management Objective 2 (MO2): the response time of a

request of service iis upper bounded by Oiand the sum of

service utilities is maximized:

maximize X

uiwhile di< Oi(2)

Management Objective 3 (MO3): the response time of a

request of service iis upper bounded by Oi, the carried load

iof service iis maximized, while service kis prevented from

starving (i.e., by specifying a lower threshold lmin for service

k):

maximize lc

iwhile di< Oiand lc

k> lmin i6=k(3)

For the management objective M1, M2, or M3, we solve

the following problem for the conﬁguration in Figure 1(b).

Given the offered load of service i, and the response time di,

we need to ﬁnd the control parameters pij and bi. We obtain

these parameters through reinforcement learning where they

represent the optimal policy.

Following the reinforcement learning approach, we formal-

ize the above problem as a Markov Decision Process (MDP),

which is a well-understood model for sequential decision

making [3][4]. The elements of this formalization for the

scenario in Figure 1(b) are:

State space: The state at time tincludes the response time

and offered load of the services running on the service mesh:

st∈ {(di,t, li,t)∈R2m|i= 1, ..., m}=S, where m is the

number of services.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DynamicallymeetingperformanceobjectivesformultipleservicesonaservicemeshForoughShahabSamaniyandRolfStadleryyDept.ofComputerScience,KTHRoyalInstituteofTechnology,SwedenEmail:fforo,stadlerg@kth.seOctober11,2022AbstractWepresentaframeworkthatletsaserviceproviderachieveend-to-endmanagementobjectivesund...

展开>> 收起<<

Dynamically meeting performance objectives for multiple services on a service mesh Forough Shahab Samaniyand Rolf Stadlery.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Dynamically meeting performance objectives for multiple services on a service mesh Forough Shahab Samaniyand Rolf Stadlery

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: