A work ow cloud management framework with process-oriented case-based reasoning

Abstract

Work ow execution in the cloud is a novel eld for processoriented case-based reasoning (PO-CBR). In contrast to traditional POCBR approaches where the focus is usually on single work ow instances, an entire set of work ow instances is considered that are currently running on cloud resources. While traditional methods such as running a work ow management tool monolithically on cloud resources lead to over- and under-provisioning problems, other concepts include a very deep integration, where the options for changing the involved work ow management tools and clouds are very limited. In this work, we present the architecture of WFCF, a connector-based integration framework for work ow management tools and clouds to optimize the resource utilization of cloud resources for work ow by Case-Based Reasoning. Experience reuse contributes to an optimized resource provisioning based on solutions for past resource provisioning problems. The approach is illustrated by a real sample work ow from the music mastering domain.

INTRODUCTION

Resource provisioning for work ow execution is a well known issue in workow management. It has been solved for on-premise systems by load balancing components, for instance. However, in cloud computing, resources are provided on-demand. Thus, work ow management in the cloud has to deal with scalable resources. Standard load balancing approaches are not capable to deal with this. Novel business concepts for work ow execution in the cloud emerge. One of these concepts is work ow as a Service (WFaaS) as introduced by [ 17, 9 ]. The Work ow Management Coalition [ 18 ] de nes a work ow as \the automation of a business process, in whole or part, during which documents, information or tasks are passed from one participant to another for action, according to a set of procedural rules". A task, also called activity, is de ned as \a description of a piece of work that forms one logical step within a process. An activity may be a manual activity, which does not support computer automation, or a work ow (automated) activity. A work ow activity requires human and/or machine resources(s) to support process execution" [ 18 ]. The idea of WFaaS is to execute activities within a cloud. A cloud vendor [ 3 ] is a company that o ers services Copyright © 2017 for this paper by its authors. Copying permitted for private and academic purpose. In Proceedings of the ICCBR 2017 Workshops. Trondheim, Norway in the cloud, for example the execution of a work ow. However, the vendor is not always a cloud provider. Even if renting the required cloud resources by a third party provider, the vendor is responsible for maintaining the service level agreements (SLA) for the own costumers. An SLA de nes agreements between the provider and the customer about di erent aspects of the quality of service. For example, an SLA can be speci ed for the execution time of the work ow. To prevent an SLA violation, the vendor may rent more resources than required (over-provisioning) but this will reduce the pro t. On the other hand, if the vendor rents less resources than required (under-provisioning) this can lead to violations of the SLA. Violations of an SLA create high costs and a loss of reputation [ 16 ]. Thus, the optimal management of resources is an important aspect for cloud computing [ 6 ] in general and, particularly, for WFaaS vendors. It is challenging to nd a good balance between over- and under-provisioning of resources [ 4 ]. A straight-forward solution to provide resources is the static way. This means, the system does not adjust itself to a changing situation. Obviously, this will lead to under- or over-provisioning [ 16 ]. A more dynamic approach is preferable. Existing approaches range from rather simple, rule-based solutions, such as observing the number of open connections to a cloud resource [ 14 ] to sophisticated, algorithmic solutions [ 15 ].

Knowledge and experience management methods [ 7 ] provide an alternative solution approach focusing on the reuse of experience. In this paper, we investigate Case-based reasoning (CBR) as a method for optimizing the provisioning of cloud resources by experience reuse. This work is an extended version of the approach we introduced in ICCBR 2014 [ 13 ]. We will introduce the architecture of WFCF (Work ow Cloud Framework) a connector-based integration framework for workow management tools and clouds that aims to optimize the resource utilization of cloud resources for work ows by means of CBR. WFCF follows a shallow integration approach, i.e. it is independent of the chosen work ow management tools and cloud systems. The idea is to have a set of WFCF components that are independent of the work ow management tools and cloud systems. A tool-speci c set of connectors interacts with the actually used tools and system. Further, we will present the details how the problem solving component of the architecture is realized by means of PO-CBR. The bene ts of using PO-CBR is twofold namely the reduction of costs for the vendor by reducing over-provisioning and SLA violations and, second, a better cost estimation based on experience. 2

WFCF ARCHITECTURE

In this section, we will explain the architecture of WFCF and its components. Starting with the overall architecture, we show the details of the monitoring and management components and how they interact. 2.1

Overall architecture

The main components of WFCF work independently from the actually used environment. To work properly, WFCF needs information about the status of the actually running work ow instances and the resource utilization of the cloud. The basic idea is, to connect the used work ow engine and cloud to WFCF with connectors. This allow the usage of di erent engines and clouds without or just small adaptions. For further details, please have a look at our previous work [ 12 ]. 2.3

Management

Whereas the monitoring component observes the environment, the management component con gures it. This means, the management component starts and stops virtual machines or PaaS container, scales resources and migrates content. Figure 2shows the management component in more detail. After CWorkload has built the WFCF CloudWF Status, CProblem is the part of WFCF which interprets the current status of the environment that is recorded as the WFCF CloudWF Status. Besides the CloudWF status, there is another archive, the Global SLA // Constraint Archive, where global constraints and SLA's are stored. The Global SLA // Constraint Archive contains SLA's and constraints that are valid for all work ows of a user. There are several di erent problems that can occur and which CProblem will identify, e.g., violated SLA's. We are planning that CProblem does not only check the current situation, but also do a forecast to identify upcoming problems and over-provisioning. A work ow de nition contains all information about the structure of the work ow. For example, the name of the tasks and their order. Via the work ow de nitions, for example, CProblem can recognize if a certain web service is going to be used in the future by a currently running work ow instance. If not, WFCF can shut down the VM or container to save money. Another possible scenario could be that currently, there is no violated SLA, but in the near future, several tasks with high resource demand will be started, which can probably lead to a SLA violation, so WFCF should scale up the resources to avoid this problem. Forecasting SLA violations, however, could be a di cult task. To decide if the start of some resource intensive tasks lead to a SLA violation is not as easy as to recognize if a web service has not started yet. A simulations seems a proper way to identify these kind of problems. Therefore, CProblem interacts with CSimu. We are planning to use CloudSim [ 1 ] as the core of our simulation part. CSimu will simulate the execution of the tasks with the current cloud status and will show if this will lead to a SLA violation. If any problem is unidenti ed, CProblem extends the CloudWF status with annotations about the problems. This new annotated model is the WFCF CloudWF Problem. Such annotations could be, for example, web service x is not longer needed or SLA y is currently violated. Whereas CWorkload is the core of the monitoring component, the WFCFSolver is the core of the management. Similar to CWorkload, the solver has two jobs. First, the solver searches for a new cloud con guration that solves the current problems. Then it nds a recon guration path from the current cloud con guration to the new solution. In the last step, the solver sends the recon guration steps to the WFCF Con gurator as shown in Figure 1. The recon gurator then will do the recon guration job. There are several possible approaches to nd a new cloud con guration. We will choose Case-Based Reasoning (CBR) as our solving strategy. 3

CBR FOR PROBLEM SOLVING

In this section we take a closer look how the WFCFSolver will solve the cloud management problems with CBR methods. As mentioned in Section 1, the idea of CBR is that similar problems have similar solutions. If a problem situation occurs the system retrieves experience by searching a similar situation from the past. In our case a problem situation is a cloud con guration with a problem, such as violated SLA's. This is the retrieval step. The key to experience retrieval is a good notion when some kind of experience is relevant for a certain situation. This knowledge is captured in the similarity measure [ 7 ]. The reuse step of CBR is to use the solutions from the past for the current problem. In our case, the solution contains re-con guration steps. This for example could be the to start new VM's or to migrate containers to another VM.

A problem situation is recorded as WFCF CloudWF Problem. Figure 3 shows an example of a simple CloudWF Problem. This example contains one VM, two containers for the required web services and a bunch of work ow instances currently being executed. The image depicts not the entire work ows but the tasks that are currently active within the instances. Most of the work ow instances are derived from the same work ow de nition and are in the same state of execution. At this point, the task Task 1 uses the web service web service 1 while Task 2 uses web service 2. In addition, there is another work ow instance (in the bottom right corner). This instance is probably from a di erent work ow de nition, or the instance is in a di erent state of execution. The current task of this instance is task 217 and for its proper execution, a web service that has not yet started is required. This example also includes the constraint that the average resource utilization must not extend 75% for reasons of performance. The example CloudWF Problem includes also three problems. The resource utilization of the CPU and memory of VM1 is too high and a new web service must start for Task 217. More complex CloudWF Problems may involve several VM's, containers and work ow instances.

A case base is an archive of previous problems and their solutions. The case base is not depicted in Figure 2, because it is part of the solving strategy and not part of WFCF itself. The solver will search the case base for similar problems in the past. In our previous work [ 11 ], we have introduced the idea of a similarity function for cloud con gurations. For the similarity of a cloud con guration, we consider the following aspects as important.

The provided resources. Two VM's are similar, if they have a similar set of resources available. For example, two VM's with a quad core processor should be more similar than a VM with a dual core processor and a VM with a quad core. The idea is, that VM's with a similar set of resources should handle general workload similar, where VM's with a di erent set of resources maybe lead to other results, for example you can not migrate a container that requires a quad core, if the VM only have a dual core. The same applies to containers. The resource utilization. VM's with a similar resource utilization, for example average CPU usage, should be considered as similar. If the utilization di ers signi cantly, a solution that is valid for one case could be invalid for the recent case. For example if the disk space utilization for a VM vm1is 20% and for an another vm2 100%, the system can not migrate a container to vm2, because of the lack of free disk space, while a migration to vm1 is feasible. The same applies to containers.

The assigned SLA's and whether they are violated or not. If two cloud con gurations have a similar set of SLA's, the con gurations should be considered as similar. Di erent SLA's or the violation of di erent SLA's can lead to a situation, where a problem of the one case is not a problem in another case. For example if a cloud con guration includes an SLA on availability and the other doesn't, the availability can be a problem in the rst case while it is not in the second. That leads to the situation, that a solution that mends the availability problem for one case is not applicable for the other case.

The executed work ow instances and their work ow de nitions. The number of the started instances and the structure of the work ow de nitions can have a high impact on the requirements for resources and for started web services. For example, if an instance of a work ow is started that requires a certain web service, every solution that does not include this web service is not valid. The structure of the work ow de nitions also speci es which tasks will be started next.

To determine the similarity of two cases, we use a composite, distance-based similarity function based on the aspects introduced before. The similarity of each aspect in two cases is computed by a particular local similarity function. The local similarity values are aggregated by means of a sum of weighted aspects. For example, the similarity function of the resources provided for a VM is based on a taxonomy, and analog for containers. For the size of the provided resources, we have been inspired by Amazon EC2 instances [ 5 ] for nodes and OpenShift [ 2 ] for containers. For other aspects, we use mainly standard distance functions. For example to determine the distance between the resource utilization for VMs vmuti, we use the Euclidean distance for the resource vectors of CPU, memory, storage, network tra c, and so on. The utilization values are provided in percentage. The distance of the resource utilization vmutil is calculated by the s n Euclidean distance vmutil(p; q) = P (qivm pivm)2, where p is the vector of i=1 n utilization values for the rst case and q for the second case. For example, qvm = 50 is the utilization of the CPU qvm with a value of 50%. p2 is the uti1 1 lization of the memory and so on.

The similarity function for the work ow aspect of our approach is ongoing work. Each work ow instance has 0 to n active tasks. These are the tasks that are currently executed. We are planning to consider the currently active tasks, as a bag of tasks in our similarity function. In addition, another relevant set of tasks can be derived from the work ow de nition namely the set of tasks that will be active in the near future. We call this the bag of tasks approaching next. We assume, these two bags of task should be an important part within the similarity function. The similarity of two individual tasks will be determined by its service characterization, the size of its input data and the name of the task. Two tasks are similar if they have the same characterization (for example CPU intensive) and if the size of the input data and the name of the task are similar. However, we have not decided yet how to implement the similarity function for bags of tasks nally.

For the reuse step, a solution is a cloud con guration without problems. The solver will search for a similar problem and use the solution for this old problem or the solution can serve as a starting point for a new solution. Anyways, the solver will send the solution back to CProblem to check if the solution comes up with new problems. CProblem will check and simulate the solution and give feedback to the solver. This will be repeated until a solution is found or another condition is ful lled. This could be, for example, a time limit. In this case, the solution with the least signi cant problem will be chosen. The usage of CBR also opens the opportunity for post-mortem analysis and improvement of the stored solution, while WFCF is otherwise idle. This lazy learning can also be used, if there is no similar soluation, or when the case base is empty. In such a case, a simple rule based approach can generate a rst placement for the current situation and a post-mortem analysis can improve the result afterward, for the next time, a similar situation approaches. In addition to the case base, there is the WFCF Cloud Resources and Service Archive. This archive contains information about the available type of containers, VM's, web services and so on. This archive helps the solver to nd valid solutions. Similar to the connectors in the monitoring part, the Cloud Service Explorer is a connector to the cloud to discover available sizes and services and store them in the Resources and Service Archive. 4

EXAMPLE

To demonstrate the idea of WFCF, we will give a running example. As our example domain, we chose music work ows to mastering music. The purpose of such a work ow is to transform and process a music le. This includes to normalize and limit the volume of the sound, increase or reduce the sample rate, convert from mono to stereo or reverse and adding special e ects like fading and compressing the size of the music le. Figure 4 shows an example work ow. The work ow is modelled in BPMN [ 8 ]. To simplify the image, gure 4 does not show the input and output les of the web services. The work ow starts with the Init Work ow Parameter tasks to initialize the work ow by a human. The user chooses some parameter for the later mastering. The following two tasks are also human tasks require along with the rst one no cloud resources. The following tasks are all based on web services and alter the music le each time. For example, the task normalize normalizes the volume of the music le, while the task fading adds a fade-out e ect to the end of the music. Let us assume that task choose le is currently active.

CProblem realizes that, in the near future, the task normalize will start. This task requires the web service normalize web service that is not available at the moment and this is a problem. CProblem prepares the WFCF CloudWF Problem and annotates that this web service is required. Because of the simple cloud con guration and because no SLA's are involved, no simulation from CSimu is needed. The WFCFSolver searches its case base for a case where a web service is required and no container is currently started. Let us assume that the WFCFSolver nds such a solution and this solution includes to start a container with the needed web service. The solver will send this solution back to CProblem to check if the solution includes new problems. This, however, is not the case. The solver can now start to plan the recon guration. After the solver is done, the WFCF Con gurator starts a container with the web service. 5

CONCLUSION

In this paper, we introduced the architecture of WFCF, a connector-based integration framework for work ow management tools and clouds. The goal of WFCF is to provide a way to integrate di erent work ow management tools and clouds, while also optimizing the resource utilization of the used cloud resources by PO-CBR. To achieve this goal, WFCF uses multiple concepts. The connector's concept allows in a modular way to integrate work ow tools and clouds by using their usual management and monitoring concepts and without the need for special requirements to the used tools. The monitoring component of WFCF analyzes the run time behavior and resource usage of tasks for a better understanding of their needs and also combines information of the work ow management tool and the cloud to a status model for future analysis and forecast of problems. The management component analyzes this status model for problems by using a combination of simulation and static methods. When a problem occurred or can be forecasted, the management component uses CBR to nd a similar problem in the past and solve the problem based on the past solution. WFCF aims at a shallow integration of cloud and work ow management tools for exible combination of tools and the optimization of resource usage. We believe that the use of PO-CBR will lead to the reduction of costs for the vendor by reducing over-provisioning and SLA violations and, second, o er the opportunity for a better cost estimation due to experience, while the approach should be less compute intensive and therefore faster as other solutions. Currently, we are working on a prototypical implementation of the of the architecture to evaluate the concept in future. For our future evaluation, we are planing to compare WFCF with Cloud Socket. An open issue is to design the similarity functions in detail and the WFCF CloudWF Status model in a universal way without dependencies of the actually used tools. Another future task is the acquisition of a larger set of problems that should be recognized and solved and also to investigate how strong is the impact of di erent optimization goals (for example, reduce costs or reduce SLA violations), for di erent solutions.

References

1. The CLOUDS lab: Flagship projects - gridbus and cloudbus ( 2016 ), http://www.cloudbus.org, 2016- 12 -08

2. OpenShift ( 2016 ), https://www.openshift.com/, 2016- 12 -08

3. Antonopoulos , N. , Gillam , L. : Cloud computing . Springer ( 2010 )

4. Armbrust , M. , Fox , A. , Gri th , R., Joseph , A.D. , Katz , R. , Konwinski , A. , Lee , G. , Patterson , D. , Rabkin , A. , Stoica , I. , Zaharia , M.: A view of cloud computing 53(4 ), 50 { 58 ( 2010 )

5. AWS: Amazon web services (AWS) - cloud computing services ( 2016 ), http://aws.amazon.com/, 2016- 12 -08

6. Baun , C. , Kunze , M. , Nimis , J. , Tai , S. : Cloud Computing - Web-Based Dynamic IT Services. Springer ( 2011 )

7. Bergmann , R.: Experience management: Foundations, development methodology, and Internet-based applications . Springer Verlag ( 2002 )

8. Chinosi , M. , Trombetta , A. : BPMN: An introduction to the standard 34(1 ), 124 { 134 ( 2012 )

9. Korambath , P. , Wang , J. , Kumar , A. , Hochstein , L. , Schott , B. , Graybill , R. , Baldea , M. , Davis , J.: Deploying kepler work ows as services on a cloud infrastructure for smart manufacturing 29, 2254 { 2259 ( 2014 )

10. Kubler, E., Minor , M. : Towards cross-layer monitoring of cloud work ows . In: Helfert, M. , Ferguson , D. , Muoz , V.M. (eds.) CLOSER 2015 , Lisbon, Portugal, 20 -22 May, 2015 . pp. 389 { 396 . SciTePress ( 2015 )

11. Kubler, E., Minor , M. : Towards a case-based reasoning approach for cloud provisioning . In: CLOSER 2016 , Rome, Italy 23- 25 April , 2016 . vol. 2 , pp. 290 { 295 . SciTePress ( 2016 )

12. Kubler, E., Minor , M.: WFCF - a work ow cloud framework . In: CLOSER 2017 , Porto, Portugal 24 - 26 April , 2017 . pp. 518 { 523 . SciTePress ( 2017 )

13. Minor , M. , Schulte-Zurhausen , E. : Towards process-oriented cloud management with case-based reasoning . In: Proc. ICCBR 2014 . pp. 303 { 312 . LNCS 8766, Springer ( 2014 )

14. Pousty , S. , Miller , K. : Getting Started with OpenShift. " O'Reilly Media , Inc." ( 2014 )

15. Quiroz , A. , Kim , H. , Parashar , M. , Gnanasambandam , N. , Sharma , N.: Towards autonomic workload provisioning for enterprise grids and clouds . In: Grid Computing , 2009 10th IEEE/ACM International Conference on. pp. 50 { 57 . IEEE ( 2009 )

16. Shoaib , Y. , Das , O. : Performance-oriented cloud provisioning: Taxonomy and survey abs/1411.5077 ( 2014 )

17. Wang , J. , Korambath , P. , Altintas , I. , Davis , J. , Crawl , D. : Work ow as a service in the cloud: Architecture and scheduling algorithms 29 , 546{ 556 ( 2014 )

18. Work ow Management Coalition: Work ow management coalition glossary & terminology ( 1999 ), http://www.wfmc.org/resources 2016- 12 -15