Scholarly article on topic 'Simulation of Kanban-based Scheduling for Systems of Systems: Initial Results'

Simulation of Kanban-based Scheduling for Systems of Systems: Initial Results Academic research paper on "Computer and information sciences"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"lean systems engineering" / "kanban scheduling" / "scheduling simulation" / scheduling}

Abstract of research paper on Computer and information sciences, author of scientific article — Alexey Tregubov, Jo Ann Lane

Abstract Systems engineering processes for evolving systems of systems (SoS) are often software-driven and software-intensive. At the same time, SoS have multiple levels of abstraction that correspond to the various levels in the SoS hierarchy where these SoS engineering processes take place. Multiple levels of management make it difficult to capture the actual state and relative value of work in these kinds of environments. The Kanban-based scheduling system (KSS) applies lean concepts to coordinate work queues to better to address these issues. The motivation to apply agile methodologies in multi-organizational multi-level environments is based on lean principles that encourage increased visibility of work in progress, limited work in progress, and identification of issues causing blocked work. Current research is focused on formulating the KSS principles and estimating expected performance of the KSS. This paper describes the KSS work flow, Kanban scheduling principles and a simulation model designed to estimate the KSS performance.

Academic research paper on topic "Simulation of Kanban-based Scheduling for Systems of Systems: Initial Results"

(8)

CrossMark

Available online at www.sciencedirect.com

ScienceDirect

Procedía Computer Science 44 (2015) 224 - 233

2015 Conference on Systems Engineering Research

Simulation of Kanban-based scheduling for systems of systems:

initial results

Alexey Tregubova*, Jo Ann Laneb

a University of Southern California, Los Angeles, CA, 90089, USA bUniversity of Southern California, Los Angeles, CA, 90089, USA

Abstract

Systems engineering processes for evolving systems of systems (SoS) are often software-driven and software-intensive. At the same time, SoS have multiple levels of abstraction that correspond to the various levels in the SoS hierarchy where these SoS engineering processes take place. Multiple levels of management make it difficult to capture the actual state and relative value of work in these kinds of environments. The Kanban-based scheduling system (KSS) applies lean concepts to coordinate work queues to better to address these issues. The motivation to apply agile methodologies in multi-organizational multi-level environments is based on lean principles that encourage increased visibility of work in progress, limited work in progress, and identification of issues causing blocked work. Current research is focused on formulating the KSS principles and estimating expected performance of the KSS. This paper describes the KSS work flow, Kanban scheduling principles and a simulation model designed to estimate the KSS performance.

© 2015TheAuthors.Publishedby ElsevierB.V.This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of the Stevens Institute of Technology.

Keywords: lean systems engineering; kanban scheduling; scheduling simulation; scheduling

1. Introduction and background

In the last several years, the effectiveness of on-demand, value-based scheduling has been proven for various software development processes. Agile methodologies have been successfully adopted by thousands of software development teams across the world1-7. However, the development of large, enterprise-wide integrated software

* Corresponding author. Tel.: +1-716-319-4731. E-mail address: tregubov@usc.edu

1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.Org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of the Stevens Institute of Technology.

doi: 10.1016/j.procs.2015.03.004

systems also known as systems of systems (SoS) have faced obstacles adopting agile methodologies. According to recent studies5'8 major obstacles for agile development in an SoS environment include a lack of visibility of capability status at the top levels of the SoS organization and a lack of insights into issues preventing the completion of SoS capabilities. As a result, a Systems Engineering Research Center (SERC) research team conceptualized an approach based on lean principles, the Kanban-based scheduling system (KSS), to address these challenges. This research8 proposed to use a multi-tiered Kanban-based process for these interdependent enterprise-wide systems or SoS. In order to further evolve and mature this concept and evaluate the potential effectiveness of the proposed Kanban-based scheduling system, the University of Southern California (USC) Center for Systems and Software Engineering (CSSE) developed a KSS simulation model. This paper describes the KSS concepts, the proposed KSS simulation model to evaluate potential KSS performance, and early results from the KSS simulator.

To formulate the KSS principles and develop a KSS simulation model, the research team needed better insights into SoS environments and common obstacles. Next, the team needed to better characterize both SoS evolution and how a Kanban-based scheduling system should be applied to SoS.

System of Systems Engineering (SoSE) distinguishes four types of SoS with respect to management at the SoS level. These types, in order of increasing authority and responsibility, are virtual, collaborative, acknowledged, and directed9. The only types that have an SoS engineering team to oversee the development of SoS capabilities are the acknowledged and directed SoS. Therefore, the KSS described in this paper is only intended to apply to the acknowledged and directed software-intensive SoS.

Next, the KSS research team studied current management approaches in acknowledged and directed SoS. The SERC research team found that often there are multiple levels of management in the SoS, some responsible for SoS capabilities and performance and others responsible for constituent system capabilities and performance. These multiple levels of management with different responsibilities and focus, sometimes overlapping each other, and often unaware of what other constituent system engineering teams are doing, make it difficult to capture the actual state and relative value of work at the SoS level. Therefore, SoS development becomes less and less deterministic and controllable. To mitigate these obstacles and improve capability delivery to the users, the research8 proposed a KSS that would provide for:

• More effective integration and use of scarce systems engineering resources.

• Improved visibility and coordination.

• Improved flexibility without reducing predictability across systems.

• Increased value delivered earlier.

• Lower governance overhead.

2. Kanban Scheduling System for SoS

The KSS for SoS, proposed in the paper8, applies original lean concepts to coordinate work queues. The KSS provides a set of guidelines and work prioritization techniques based on lean concepts for single system development, but extends it to multiple teams and levels within the SoS. Key to this approach was defining a mechanism to relate single system tasks to original SoS capabilities and to assign value to a task based on both its value to the single system as well as its associated SoS value.

In general, the KSS for SoS is implemented as a hierarchical set of teams using Kanban-based scheduling. These teams are associated with SoS constituent systems and may represent different products and system domains. Each team uses Kanban boards to identify and track incoming work items (backlog) and work in progress. The capacity of each queue is limited; these limits depend on resources available within the teams. Work backlogs are usually formed collaboratively by the upstream customer(s) (another team in the hierarchy) and the team responsible for execution of the backlog. At this point in time, a work item (WI) is placed in the work backlog queue for a given team and all significant dependencies, date-certain events, and other special concerns should be identified.

Apart from managing work queues the KSS also introduces notion of Classes of Service (CoS) - one of the most important attributes used in work prioritization. For the KSS we selected CoS based on the book Kanban10. Each WI in a backlog is given a CoS that defines how the WI must be handled in the queue. For example, a WI with an expedite class of service in the queue requires an immediate response, and it allows a team to interrupt work that is

already in progress All other CoSs do not allow interruptions for work already started unless a new dependency external to the team is discovered. Table 1 defines five general CoSs that are used in the KSS.

Table 1. CoSs description.

CoS Description

Critical A Critical Expedite WI represents something that fixes an existing or imminent issue within the system. Safety,

Expedite security, or other emergency WIs are assigned this CoS. It is disruptive and requires all appropriately skilled resources to suspend their current activities and work on the Critical WI. It allows a team to interrupt work that is already in progress to resolve the Critical Expedite WI. It also applies to every allocated WI associated with the WI assigned to this CoS.

This CoS is assigned to very high priority WIs where the speed of completion is such that this work should take priority over all other work in the ready queue. It is not disruptive, because all work in progress is allowed to finish before the important work begins.

A Date certain WI reflects work that must be completed by a specific date or there will be significant consequences. Regulatory implementation deadlines, COTS upgrade preparation, or integration/deployment dependencies are candidates for this class of service. It operates essentially like a Standard CoS, but as the date becomes closer, it may elevate to an Important or Critical Expedite COS based on workload.

This is the normal CoS for the development organizations' work. A high percentage of work should be assigned at this level to provide the desired outcomes.

Background work is work that must go on but is usually not time critical. It includes things like architectural enhancements, low priority problem fixes research and environmental scanning. It is usually prioritized by its length of time in the queue.

In addition, Kanban boards in a KSS allow us to address the lack of process visibility in multilevel SoS environments. Each team on each level has a virtual Kanban board for tracking work in progress and a demand backlog that is integrated with other team Kanban boards at higher levels in the SoS hierarchy.

Before we can start a discussion of the simulation model in detail, we need to provide some insights into a typical structure of a SoS and the KSS proposed in the Turner-Lane paper8 that was based upon a healthcare SoS example, a medical information management set of integrated systems that consists of hardware, several million lines of source code, numerous commercial-off-the shelf (COTS) software products, and communications networks that support the administration and delivery of health care in networked set of several hundred hospitals and clinics. The key custom software constituent systems within the health care SoS include user access management, patient management, pharmacy, laboratory, radiology, and patient telemetry. The constituent systems share a single data repository that maintains the information for all of the patients and personnel related to a given health care site. Some of the key constituents use COTS products tailored and integrated into the health care system. In addition, there are interfaces to other health care systems maintained by the parent organization at various sites. The interfacing systems include custom legacy systems and COTS products and devices.

As was mentioned earlier, a typical enterprise SoS such as the healthcare example has three major levels:

> Executive/Stakeholder Management (ESM) - level that determines what SoS capabilities should be developed, their values, and desired schedule.

> System Engineering (SE) - level that contains all enterprise/SoS capability-related software engineering activities such as developing engineering approaches and technical requirements based on customer and management requests, monitoring system performance and initiating changes when performance falls below acceptable thresholds, allocating requirements across the products and other engineering domains, and coordinating changes across multiple constituent systems.

> Product/Domain Engineering (PDE) - constituent systems level, where each product or domain team in the system has a separate KSS Kanban board. This is the level where requirements are decomposed into smaller WIs and implemented by software development or specialty engineering (e.g., network management, information/data management) teams.

The overall KSS in SoS may be viewed as a network of Kanban boards on these three levels. Each Kanban board represents a demand backlog (all WIs that need to be implemented, but have not yet been started), work in progress, and completed WIs. Connections between Kanban boards are defined by relationships between WIs on these boards.

Important

Date Certain

Standard Background

The simple SoS structure is illustrated in Figure 1. It shows all three major levels (their Kanban boards pictured as "KSS" in Figure 1). Executive/Stakeholder Management level and System Engineering team use their Kanban boards for tracking capabilities and requirements. As shown in Figure 1 these boards also serve as an aggregation level of downstream teams' work queues. For example, the System Engineering team's Kanban board shows a requirements backlog (in addition to their own WIs backlog). Kanban boards of the Product/Domain Engineering level are responsible for local work queues. These queues mostly consist of product/domain specific WIs, which eventually implement related requirements. It is also possible that a product/domain engineering team coordinates work of several software development teams that work on a given product. For example, the healthcare SoS in Figure 1 has several software development teams working on that system.

Fig. 1. SoS structure and work flow.

3. Simulation model overview

This section describes a discrete-event simulation model used to explore the KSS. The model was implemented in the KSS Simulator, a software tool developed by the authors. Figure 2 illustrates general simulation work flow. A Scenario Generator allows user to build a scenario with given characteristics such as a total number of teams, a total number of WIs in the model, etc. The KSS simulator implements the discrete-event simulation engine and prioritization algorithms in JAVA. The Scenario Generator allows the user to create stochastic events in a scenario that change value, CoS and other WIs' properties over time. However, once the scenario is generated, it is deterministic. Uncertainty of different events can be modeled via generation of multiple scenarios (multiple alternatives) for the same set of constraints such as a total number of teams, a total number of WIs in the model, etc.

Fig. 2. Simulation work flow.

The model describes engineering processes in a SoS as a discrete sequence of timeframes. All the system engineering activities in the model are represented as a set of WIs grouped by aggregation nodes such as SoS capabilities and system requirements. Together the aggregation nodes, the WIs, and relationships between the WIs form a network of WIs. There are two main types of relationships the network of WIs: precedence and causality. Precedence relationships define prerequisites for a WI. Causal links between WIs define how one WI affects others. For example, completion of one WI can cause updates in other WIs (e.g. change of value, estimated effort).

The way the WI network evolves is defined by the event scenario and other input parameters such the scheduling algorithm and team resources. The event scenario is a set of events that describes a network of multiple WIs and how it changes over course of the scenario execution. Overall structure of the WI network as well as causal and precedence relationships between WIs is deterministic and predefined in the event scenario once it is generated. However, order in which various events occur can be different depending on the prioritization technique used.

One of the main purposes of the model is to compare amount of value delivered to the stakeholders over time using various work prioritization techniques. There are two algorithms that are compared in the KSS simulation model now: value-neutral (random) work selection from the backlog and value-based prioritization technique known as Kanban Scheduling or KSS-scheduling for short.

The following subsections briefly introduce the key entities of the model: WIs and aggregation nodes, kanban boards, resources.

3.1. Work Items and Aggregation Nodes

A WI is a task that requires real work to be completed. Every WI is assigned to a team, which means it is added to the team's Kanban board. WIs have the following attributes: id, creation time, effort required to accomplish it, and status. A WI's status changes over time. Figure 3 shows the general life cycle of the WI.

CREATED

IN PROGRESS

SUSPENDED

Fig. 3. WI life cycle.

An aggregation node is an upper level representation of related sub-activities in the Kanban network. For example, a capability is an aggregation node that is decomposed in one or more requirements and a requirement is an aggregation node that is decomposed into one or more WIs.

In Figure 4, four capabilities are decomposed into six requirements. These requirements are allocated across three Product/Domain teams. Each team decomposes assigned requirements into WIs. Figure 4 can also be viewed as an example of a work breakdown structure. The model implements three types of relationships between WIs work decomposition structure (Figure 4), WI prerequisites, and causal relationships.

4-J TO

C2 C3 C4

R1 R2 R3 R4 R5 R6

Fig. 4. Decomposition levels (capabilities into requirements and requirements into WIs).

3.2. Kanban boards

The team Kanban board is an abstract concept that aggregates a list of resources and a list of related WIs. The purpose of Kanaban boards is to provide visibility of work in progress. The network in the model has all three major levels: Executive-Stakeholder Management level, Systems Engineering level and Product/Domain Engineering level.

3.3. Resources

A resource is an agent that can be assigned to work on a WI. The effectiveness of each resource working on a WI depends on total number of resources working on the same WI: the more people that work together on one WI, the less productive each one is.

The resource effectiveness can also depend on a number of interruptions. Every time a WI is resumed, the resource assigned to it has to switch his or her work context from previous WI. This context switching usually takes some extra amount of time. A typical value for this context switching is 1-2 timeframes if each timeframe is 1 hour.

4. Simulation inputs and outputs

The simulation manages the evolution of a Kanban network during the user-specified duration, applying the prioritization algorithm and re-calculating properties for each timeframe.

The inputs for the simulation are the scenario script and the simulation parameters. The simulation parameters include:

• Resources allocation

• Prioritization algorithms

• Stop condition

A scenario is a set of if-then rules, also called triggers, that defines the initial state (number of capabilities, resources and their properties, etc.) and the way in which the Kanban network evolves over time. Every scenario describes the creation of new WIs (that includes their properties, dependencies and prerequisites), and their changes over the course of their life cycle. In the KSS simulator (program implementation of the simulation model), each scenario is an xml file that contains these triggers.

The simulation output is a set of indicators that show how much time was spent to complete the work, how much work was completed, resource allocation statistics, and so on. Currently implemented indicators include work completeness of each capability over time, value delivered over time, team workload (number of busy resources and underutilized resources), number of blocked tasks.

5. Simulation experiment

In order to demonstrate the operation of the simulation model we use a simple enterprise SoS example. This simple SoS example is a simplified version of a real health care SoS described in the paper8. However, in the simple example SoS we limited number of constituent product/domain teams to system engineering team, network team, patient management system, and database team.

In this experiment we run the scenario, described above, two times. Each time we used different prioritization algorithm. The first algorithm was KSS-scheduling, and the second was Value-neutral (random) work selection. The resource configuration was the same in both runs.

Simulation results allow us to compare various process performance indicators. Table 2 shows overall indicators such as schedule and effort.

Table 2. Schedule and effort comparison.

Indicator KSS-scheduling Value-neutral

work selection

Time spent, schedule (days) 24 28

Total effort (person-days) 120 121

Average time of WI suspension 3 3.4

The KSS-scheduling algorithm saved 4 days in terms of schedule and one person-day of effort. Every time when an unfinished WI is interrupted, it takes some extra effort to finish it (effort penalty). This effort penalty is caused by reduced resource performance due to context switching between different WIs. Effort associated with the KSS-scheduling algorithm was saved because of fewer of context switching events between different WIs. Time was saved because of more efficient work dependency resolution. To better explore how limiting work in progress can affect schedule and effort, the simulation scenario must be larger in terms of number of WIs and length of the simulation. The simple SoS example has 4 teams and 22 WIs. This example was used as a proof of concept for small agile teams. Typical number of WIs in larger scenarios is between 100 and 2000 WIs. Additional experiments have been run with such scenarios. The outputs were relatively consistent with smaller cases.

Figure 5 shows cumulative value delivered over time. The KSS-scheduling, which uses value based approach, delivers value faster than Value-neutral work prioritization. The value of each WI is determined by parent capabilities values.

Value comparison

1 600 >

KSS Scheduling

Value-neutral (random selection

15.00 20.00

Time (days)

Fig. 5. Delivered value comparison.

In general, the KSS-scheduling has a smaller number of suspended tasks. When tasks are suspended, the KSS-scheduling interruption time is usually shorter than in Value-neutral work selection. That can be explained by a more efficient WI dependency resolution in the KSS-scheduling.

6. Prioritization Algorithms

Each transition between states of the network begins by applying a prioritization algorithm that defines which WIs will be done next. The prioritization algorithm analyzes last state of the KSS network and assigns WIs to resources. There are different ways to perform this allocation.

The KKS-scheduling algorithm analyzes the following properties:

> Dependencies between tasks - this property separates the WIs in two groups: the blocked items and available items. Every item in the blocked group has one or more dependency to another WI that is still to be completed.

> Class of service (CoS) - WIs are arranged according to their CoS. Each CoS is handled according to its semantic definition. For example, a WI with an expedite class of service in the queue requires immediate response, and it enables team to interrupt work that is already in progress. All other CoSs do not allow any interruptions of work already in progress. Another example is a date-certain CoS. A date-certain WI's priority depends on current date, day due, and WI estimated size and value. The closer the due date the higher priority of the WI in the backlog. All other CoSs define priority level.

> Value - all tasks are prioritized according to their values.

After that, if there are items in the expedite group, they are assigned to resources starting with the task with the highest value. If there are not enough resources available, we pull out busy resources. Context switching is inevitable here: resources already assigned will switch from their current task to the new expedite task. Resources with proper skills (specialties) are chosen from the least valuable WI in progress in the least important class of service group.

In the case of expedited analysis tasks at the system engineering level, the prioritization algorithm has to look for available resources not only in the team the task belongs to, but also among all other teams (other product teams). This exception allows us to bring experts from different teams together and work on that analysis task together.

The KSS-scheduling algorithm was compared to Value-neutral work selection. The main difference between them is how they determine the most valuable work to do. The scheduling algorithms discussed above were chosen, because quite often such algorithms are used de facto in real life.

A Value-neutral (random) work selection algorithm is a simplified variation of the KSS-scheduling, which ignores WIs' value (all WIs are equally valuable) and acknowledges only expedite CoS. Value-neutral work selection allows us to simulate decision-making process in an extreme case of business value unawareness due to lack of visibility.

7. Conclusions

The simulation model presented above was designed to explore an effect of using the KSS in multilevel systems. The model implements discrete-event simulation of the software-intensive system engineering processes. Simulation outputs allow us to estimate how the KSS-scheduling can achieve predicted benefits in terms of delivered value over time and schedule.

In addition to the KSS performance estimation, the simulation model and the KSS simulator can be used for broader purposes as well. For example, the KSS simulator can also be used as a tool to help in business decision making in SoS environments. It can help to assess:

• the impact of using the different resources for a given task;

• understanding how a task affects the completion of a higher level capability to more quickly achieve full capability value;

• what provides more value early in the incremental development process;

• visibility into what is blocking task/capability completion;

• impacts of context switching between tasks (as tasks become blocked/unblocked);

• amount of resources required in a given specialty area for a given work profile.

8. Future work

The key next steps are to pilot the KSS with several organizations and fine-tune the simulator based on pilot results. Piloting the KSS would require adjusting the SE processes for purposes of the Kanban scheduling and customizing it for a given SoS development organization. Pilot results can be used to improve the simulator through verification and calibration using empirical data and pilot organizations' feedback. Based on the size of the pilots and their current scheduling algorithms, it may also be possible to further explore the scalability of KSS as well as compare KSS performance to other scheduling approaches. The further simulation experiments will also include:

• comparison of the KSS-scheduling with other prioritization techniques such as First-In-First-Out (FIFO) where WIs are worked in the order received and Last-In-Last-Out (LIFO) where the most recently created WI is perceived as the one with the most value;

• modeling uncertainty and sensitivity analysis of the KSS-scheduling.

Acknowledgements

This material is based upon work supported by the U.S. Department of Defense through the Systems Engineering Research Center (SERC) under Contract H98230-08-D-0171 and documented in technical reports SERC-TR-022. SERC is a federally funded University Affiliated Research Center managed by Stevens Institute of Technology.

References

1. Dingsoyr, T., Nerur, S., Balijepally, V., Moe, N. A decade of agile methodologies: Towards explaining agile software development. J Syst Software 2012; 1213-1221.

2. Boehm, Barry and Turner, Richard. Balancing Agility and Discipline: A Guide for the Perplexed. Boston, MA: Addison Wesley; 2004.

3. Larman C. and Vodde, B. Scaling Lean & Agile Development. Boston, MA: Addison Wesley; 2009.

4. Poppendiek, Mary. Implementing Lean Software Development. Boston, MA: Addison Wesley; 2007.

5. National Defense Industrial Association. Top Systems Engineering Issues In US Defense Industry. Systems Engineering Division Task Group Report, http://www.ndia.org/Divisions/Divisions/SystemsEngineering/Documents/Studies/Top%20SE%20Issues%202010%20Report%20v11%20F INAL.pdf. September, 2010.

6. Turner, Richard, Shull F., et al. "Evaluation of Systems Engineering Methods, Processes and Tools on Department of Defense and Intelligence Community Programs: Phase 1 Final Technical Report," Systems Engineering Research Center; 2009 Sept. Report No SERC-

2009-TR002.

7. Turner, Richard, Shull F., et al. "Evaluation of Systems Engineering Methods, Processes and Tools on Department of Defense and Intelligence Community Programs: Phase 2 Final Technical Report," Systems Engineering Research Center, 2009 Dec. Report No. SERC-

2010-TR004.

8. Turner, R., Lane, J. Goal-question-Kanban: Applying Lean Concepts to Coordinate Multi-level Systems Engineering in Large Enterprises.

Proc. of 11th Confe on Syst Eng Res CSER 2013, 512-521.

9. Maier, M. Architecting principles for systems-of-systems. Syst Eng 1998; 267-284.

10. Anderson, David. Kanban: Successful Evolutionary Change for Your Technology Business. Sequim, WA: Blue Hole Press; 2010.