[Workload balancing capabilities of load based pull production control systems] 1

(1)

[Workload balancing capabilities of load based pull production control systems – An evaluation of various load based mechanisms using discrete event simulation]

[Hendrik Wilz] [S3050106]

Master thesis, MSc Technology & Operations Management University of Groningen, Faculty of Economics and Business

January 30, 2017

Supervisor: Nick Ziengs, MSc

(2)

Abstract

Pull systems can be simple but effective production control systems. However, most pull systems do not perform well with random processing times and have a limited applicability for the MTO sector. The systems do not achieve a decent reduction of total throughput times (TTT) or only under special conditions. The unsolved problem is how to design a load-based system that is able to perform well with random processing times. To overcome this problem a load-based pull system with an adapted release mechanism is proposed. Pull systems need to balance the workload effectively over the different routes and stations in order to reduce TTT. By

prioritizing orders regarding their size, the adapted release mechanism balances the workload within the routes and between the different routes. A discrete event simulation with several sets of experiments was conducted to test this release mechanism in a load-based system and

compare it to various other pull systems. The results show an improved performance of the proposed system compared to the other systems. The magnitude of this performance improvement strongly depends on the used ratio of large to small orders.

Keywords: [production control], [pull system], [load-based], [workload balancing],

(3)

Every order has a specific amount of needed processing times at the station along its route. Load-based cards display a certain amount of processing times. Orders that need less or equal processing times than displayed by one load-based card are called small orders and need only 1 card to be released. Orders with more needed processing times than displayed by one card are called large orders and need 2 cards to be released.

Production Planning and Control Systems (PPCS):

PPCS are tools to control and plan the production. This is done over functions like order scheduling or material and capacity planning. The goal is to reduce WIP, production and lead times (Stevenson et. al., 2005).

Pull Systems:

Pull systems control the production by limiting the amount of WIP on the shop floor. Signals e.g. in form of tokens are used to signal free capacity on the shop floor. Releases of orders are regulated through this signals to keep the WIP under the limit (Hopp & Spearman, 2004).

Workload Balancing Capability

(9)

1. Introduction

In the manufacturing industry controlling the workload as well as the input/output is important to realize short lead times and total throughput times (TTT) (Moreira & Alves, 2009). Production planning and control systems (PPCS) are used to gain an advantage over competitors by limiting the work in progress (WIP) and reducing shop floor throughput times (STT) and TTT. Especially in the increasingly important make-to-order (MTO) sector it is most importance to have these advantages in lead times as products cannot be produced to stock and the

competition in this sector is high (Stevenson et. al., 2005).

The described advantages can be achieved by using card-based pull systems to control the production and throughput times by limiting the WIP (Germs & Riezebos, 2010). But limiting the WIP does not automatically reduce the TTT. For the reduction of TTT the pull systems need to be able to effectively balance the workload over the different workstations on the shop floor. For each workstation, the workload should be kept on a target level. In order to do this, the WIP has to be balanced between the different routes and between the different station within each route (Land & Gaalman, 1998).

(10)

Previous researches from Germs & Riezbos (2010) and Ziengs et. al. (2012) have shown that widely used straightforward unit-based pull systems are able to balance the workload effectively by dividing the WIP between the different specific routes. But the systems failed to achieve reduced TTT when random processing times were simulated.

Bodnar (2016) studied a load-based pull system with a simple release mechanism that balances the workload between the routes. This system showed effective workload balancing capabilities and was even able to achieve reduced TTT with random processing times, but only under specific conditions.

It can be said, that the applicability of the most card-based pull systems is limited for MTO companies (Thurer et. al., 2016). Thus, further research of load based pull systems is needed to overcome the problem with processing time variability. The purpose of this research should be to improve the applicability and robustness of card-based pull systems in the MTO environment. A system that shall overcome this problem has to be able to balance the workload effectively between and within the routes under a variety of processing time distributions.

In this research, a load-based pull system will be investigated that is supposed to be able to balance the workload between the routes and within the routes. This shall be possible by prioritizing the orders by the sum of the needed processing times at the different stations. It releases orders, which’s added processing times are smaller than a certain threshold, first onto the shop floor, whereas the system researched by Bodnar (2016) releases the orders according to the FIFO rule.

(11)

The gap is that simple types of pull systems using different prioritizations regarding processing times were not researched yet.

The aim of this research is to explore the proposed load-based release mechanism and to investigate if this release mechanism is able to effectively balance the workload under the influence of processing time variability. Further, the performances of the different described unit- and load-based pull system will be assessed. The purpose is to research how currently in use load-load-based pull systems need to be designed or adapted to be able to cope with processing time variety.

For this, a discrete event simulation with unit- and load-based CONWIP and m-CONWIP systems will be conducted to investigate the performance of the systems. The throughput time performances of those systems will be compared in various environments to analyze the workload balancing capabilities.

By this, the research will contribute new insights to the literature about how load-based pull systems react to and perform under processing time variability regarding their workload balancing capabilities, and if they are more useful than the unit-based systems. For the managerial contribution, finding suitable pull systems for the MTO environment would enable companies to use simple pull systems or adapt the systems currently in use to control the production of products with processing time variety.

(12)

(13)

2. Background

In chapter 2, first workload balancing will be discussed and then the relevant literature for production planning and control systems (PPCS) and pull systems will be reviewed. Further, the in this study used CONWIP and m-CONWIP pull systems will be described.

2.1 Workload Balancing

According to Land and Gaalman (1998), control systems capability to reduce the TTT depends on the workload balancing capability. Total throughput time consists out of different times and these times get influenced differently by the limitation of WIP and workload balancing.

Figure 1: total throughput times components (adapted from Germs & Riezebos, 2010)

(14)

time (OPT), which is between acceptance and the release onto the shop floor, and the shop floor throughput time (STT), which is all the time the order spends in the queues and getting processed at the different processing stations.

Germs & Riezebos (2010) stated, workload balancing improves the control of arrival times of orders at the different stations on the shop floor. Because of this control, the required queue length to achieve a certain utilization level becomes shorter. This will result in shorter STT, but the orders may have to spend more time in the order pool, as the amount of orders allowed on the shop floor is limited and thus the orders have to wait longer until they get released. The workload balancing is only effective if the TTT gets reduced, which implies that the reduction in STT needs to be bigger than the increase in OPT (Ziengs et. al., 2012).

(15)

Figure 2: Influence of card count on TTT and STT (adapted from Germs & Riezebos, 2010)

The points on the graph display different card counts, from few cards on the left to an increasing number on the right. As it can be seen for a system that has no constrains on the WIP (push system), the STT is the highest. With limiting the amount of WIP the STT and TTT get reduced. But if the card count becomes to small the TTT increases, which is, as explained above, due to an increasing OPT that overcomes the STT reduction. This graph displays why it is important to find the right number of cards in order to reduce TTT.

(16)

workload imbalance between the routes (Ziengs et. al. 2012). The workload of the system is imbalanced when one route is way more congested than others routes. This can be seen by the availability of cards for the different routes. Once the imbalance is signaled to the release mechanism, the orders will be released prioritizing less congested routes, which will balance the workload in the system. As stated above, this will reduce the queue lengths and by this the STT and TTT.

In conclusion, in addition to the setting of the right card count the appropriate implementation of control loops is necessary to achieve effective workload balance.

2.2 Production Planning and Control Systems

PPCS systems can help to reduce cycle times, improve quality, reduce costs and holding due dates. Those aspects are especial important for MTO companies, as products cannot be produced in advance or get stored (Stevenson et. al., 2005). The purpose of such a PPCS is to plan capacity and materiel requirements, demand management and to schedule the orders and by this reduce cycle and lead times (Stevenson et. al., 2005).

(17)

Figure 3: Production Planning and Control (adapted from Fredendall et. al. 2010)

The first step order entry starts when an order arrives. Based on properties of the order like the production route, processing times, revenue and due date, a decision about the order is made. The order can either be accepted, rejected or new negotiated. Different strategies like total acceptance or due date negotiation can be used. This stage also sets the due dates and takes the workload capacity into account. At the end, accepted order get released into the order pool (Fredendall et. al., 2010; Moreira & Alves, 2009). More detailed descriptions can be found in Fredendall et. al. (2010) and Moreira & Alves (2009).

(18)

different routes and stations of a production system (Moreira & Alves, 2009). For further descriptions of these stage and the different used rules look at Moreira & Alves (2009).

The last stage is the priority dispatching, which controls the route of the order on the shop floor and the prioritization of orders in the shop floor queues. Different rules need different information to be performed, like processing time or routing of the order (Fredendall et. al., 2010; Moreira & Alves, 2009). When the queues are small, order dispatching is less important as orders do not have to wait long and the possibilities to prioritize orders are limited due to the low count of orders in the queues. Long queues indicate that the workload is not balanced and should be prevented by a correct order release (Land & Gaalmann, 1996). This means, if the order release fails to do this, order dispatching is needed to balance the workload and gets more important (Fredendall et. al., 2010).

2.3 Pull Systems

(19)

Pull systems are production control systems and can perform the above described functions of order release and order dispatching. As they do not plan the production, the order entry stage is not handled by pull systems. This stage is done before the pull systems starts to control the production. The pull system does not decide which orders get into the order pool.

The order release stage is handled by cards. Orders or jobs in the order pool need a certain number of cards to be released onto the shop floor, as long as the cards are not available the order has to wait. Once the cards are available, they get attached and the order gets released onto the shop floor. When the order got processed, the cards are detached and send back to get attached to new waiting orders. With this system, the number of cards displays and controls the maximal amount of workload in the system (Hopp & Spearman, 2000).

For pull systems, priority dispatching is also handled by cards and works the same as order release. Order dispatching is performed at different positions on the shop floor by so called loops. This loops divide the shop floor into smaller sub systems and control the production in these subsystems with the above described card mechanism. An example of a system with such loops is the Kanban pull system. If only one loop per route is installed, the production of the order cannot be changed once it entered the shop floor (González-R et. al., 2011).

(20)

2.3.1 Structure of pull systems

The structure of a pull systems is defined by the location and amount of the used control loops. Finding the best specific pattern of loops for the specific environment of the system is crucial for its performance (Gaury, 2000). For this research the structural aspect of route specific and non-route specific loops are important.

(1) Route specific vs non-route specific

In a route specific structure the cards belong to a specific production route or a part of it. Incoming orders need to take a specific route and therefore need specific cards which displays free or soon available capacity for this specific route. One or multiple loops per route are necessary for a route specific structure in order to differentiate between the routes. The in this research used m-CONWIP system is an example for a route specific structure (Germs & Riezebos, 2010).

Non-route specific structures do not have route specific loops. As an example, the CONWIP system has just one loop for all routes. Orders are released on to the shop floor regardless which route they need to take and cards just display available capacity for the whole system but not for a specific route (Germs & Riezebos, 2010, Ziengs et. al., 2012).

(21)

2.3.2 Configuration of pull systems

The configuration of a pull systems describes the characteristics of the used cards and their amount. For this research the below described aspects of the configuration are important.

(1) Number of cards

The number of cards that are used in a loop represents the amount of work which is allowed in the loop (Liberopoulos & Dallery, 2000). The optimal number depends on the characteristics of the system. If too many cards are in the systems, the queues are getting long and a lot of WIP is on the shop floor. This reduces capability of the system to reduce STT, especially if order dispatching is not possible. If too less cards are used, the idle times of the stations increase as not enough WIP is in the system. A lot of research focus on ways to find the optimal card count (Gaury, 2000).

(2) Product anonymous vs product specific

(22)

shown that product specific cards are not suitable for an MTO environment because of the routing and product variety, although product specific cards can theoretically be route specific.

(3) Unit-based vs load-based

Load-based cards represent a certain amount of processing times, whereas unit-based cards just represent one order. This means the number of unit-based cards equals the maximal amount of orders that can be simultaneously on the shop floor. As unit-based cards display only an order without the exact processing time, the needed capacity for the order is just a rough approximation (Ziengs et. Al. 2012). This makes unit-based pull systems less suitable for environments with a high variability in processing times, as the needed time to process the orders on the shop floor can only be estimated.

Load-based cards signal a certain maximal amount of processing time or capacity. By this, these can differentiate orders based on the required processing time or processing capacity and assign a needed required number of cards depending on the processing time of the order. For every order the processing time needs to be known. If an order needs more capacity than displayed by one card, more than one card needs to be attached to release the order. This way buckets of processing times are created and each bucket is represented by a certain number of cards.

(23)

In summary, a system that is used in an environment with high process and route variety, should use product anonymous load-based cards and have a route specific structure to be able to reduce TTT.

2.4 Release mechanisms

In this research two different release mechanisms will be simulated. This release mechanisms differ in the way they prioritize orders. For the understanding of the mechanisms and the prioritization the following is important to know.

As already described load-based cards display a certain amount of processing time. Orders that need this amount or less processing time, need one card to be released and will be called small orders. While orders that need more processing times than displayed by one card, need 2 or more cards to be released and will be called large orders. Hence, the ratio of large to small orders is the ratio of orders that need 2 cards to the amount of orders that need only one card.

(1) Release mechanism 1: Balancing workload between different routes

The first mechanism uses route specific cards and routes specific control loops to get information about the workload status in the different routes and to control the WIP and balance the workload. With this information, orders for less congested routes can be prioritized and released first and thus the workload can be balanced between the routes (Ziengs et. al., 2012).

(24)

Figure 4: Flow chart release mechanism 1

(25)

Figure 5: Start situation release mechanism 1

The release mechanism checks the orders in the order they wait in the order pool. This means the release mechanism first checks if enough available cards for the first order are available, but as only one card is available, this order cannot be released. This card gets reserved for this order. The mechanism then checks if the required cards for the second order are available. When that is the case, the order gets released onto the shop floor. Finally, the last order gets checked, but as the originally available card is reserved and hence not available, the order cannot be released.

This leads to the situation shown in the following figure.

Figure 6: Situation after card allocation release mechanism 1

(26)

enough capacity to release an order for route 2 is available, this order will be released and thus starving of stations in route 2 is prevented.

This mechanism can also be used with unit-based cards, where every order needs just one card to be released. In this case, the first order to be released would be the order for route 1, followed by the order for route 2. The unit-based cards system also balances workload between the routes, as the order for route 2 would also be released if no cards for route 1 were available. Routes without free capacity cannot block less congested routes.

In the unit-based pull systems researched by Germs & Riezebos (2010) and Ziengs et. al. (2012) this release mechanism was used. For constant processing times, it showed a good performance but it failed to balance the workload efficiently under processing time variety as no improvement in form of TTT reduction could be achieved.

The load-based version tested by Bodnar (2016) was able to balance the workload effectively with constant processing times. But with random processing times it was only able to reduce TTT if a specific ratio of 20% large orders to 80% small orders was simulated. With higher proportions of large orders the system failed to do so. Bodnar (2016) implies, that as the system prioritizes orders by their arrival rate and after the FIFO rule, large orders can block routes while they are waiting for the second card to be attached. If this happens only to one route, orders for other not blocked routes will be send in first.

(27)

reduces the utilization and the output rate of this stations and the whole system. Bodnar (2016) implies that this blocking is the reason why the system fails to balance the workload effectively.

(2) Release mechanism 2: Balancing workload between and within different routes

Along with his explanation for the lack of workload balancing capability of the first release mechanism, Bodnar (2016) implies that prioritizing small orders over large orders improves the performance of pull systems. He stated that large orders might block routes and thus cause starving of downstream stations. By prioritizing small orders this blocking is prevented, as capacity in form of cards is not reserved for large orders. This means that small orders, which are in the rear of the queue, will be released with the available card and the large orders have to wait until 2 cards at a time are available.

The following figure shows the flow chart for this release mechanism.

Figure 7: Flow chart release mechanism 1

(28)

and, as long as enough orders are in the order pool, closer to the target workload level. This improves the utilization of the stations without increasing the WIP and hence reduces the TTT.

In the following it will be described, how this release mechanism handles the same situation described for release mechanism 1.

Figure 8: Start situation release mechanism 2

In an example for this mechanism, the sequence of orders in the order pool is: an order that requires 2 cards for route 1, an order that requires one card for route 2 and at third position an order that requires one card for route 1. The available cards are one card for route 1, one card for route 2 and one card for route 3.

(29)

Figure 9: Situation after card allocation release mechanism 2

As it can be seen, this release mechanism releases the order for route 1 that needs only one card. This outcome is contrary to release mechanism 1, which did not release this order.

Once the order for route 1 with 2 cards got checked, the release mechanism still has to check if the required cards are available. The checking can either be done constantly or as an event whenever a card gets detached. This means it can happen that more than one order needs to be checked simultaneously.

In summary, as the needed capacity for the large order is not available, a small order that is waiting for this route will be released before the large order. This will prevent stations in this route from starving. With this prioritization rules the workload gets balanced within the route.

(30)

different release periods and workload limits for the optimal performance. They found that prioritizing orders by their size and releasing large orders first, can lead to better workload balance and a performance improvements of the system. For further information of the used COBACABANA system and about the different kind of cards, look at Thürer et. al. (2010), Thürer et. al. (2014) and Thürer et. al. (2016).

This research of Thürer et. al. (2010; 2014; 2016) show that also complex control systems struggle to balance the workload under processing time variability. They researched and stated that balancing workload within the route through prioritizing the orders by size helps to improve performance. However, this also shows that other prioritization approaches than the in this study researched one are possible and might lead to good results. Further, the best prioritization approach may depend on the used release mechanism and the control system itself.

2.5 CONWIP and m-CONWIP

In the following the pull systems CONWIP and m-CONWIP will be explained in more detail, as these are the systems which will be used in the simulation.

(1) Description CONWIP

(31)

maximal amount of WIP in the system. CONWIP is often used as a reference systems to the performance of other systems as it provides good performance in a lot of environments (Gonzalez et. al., 2012).

Figure 10: CONWIP ( adapted from Germs & Riezebos, 2010)

(2) Description m-CONWIP

(32)

Figure 11: m-CONWIP system (adapted from Germs & Riezebos, 2010)

2.6 Performance predictions and research questions

As explained above it is expected that the advanced release mechanism, with the ability to balance the workload not only between but also within the route, performs the best. This means the m-CONWIP system with the second release mechanism is supposed to achieve higher STT and TTT reduction compared to the simple CONWIP with unit-based cards.

This leads us to the following research question:

(33)

3 Methodology

In the following chapter, first the chosen methodology is described. Afterwards the model gets explained in detail before the different experiments are listed.

3.1 Research method: simulation

Based on the research questions a quantitative deductive research method is needed, that is able to model the pull systems and release mechanisms. Many different modeling approaches are used in literature and practice, including simple calculations, use of spreadsheets, linear programming, complex heuristics and simulations. It is needed to measure the outcome of the model under a variety of different pull systems and input parameters. These different settings together with the model described below result in a variable, interconnected and complex system. Simulations work very well to predict the outcome of such systems, to compare these systems and describe the effects of variables tested in different experiments (Robinson, 2004, p 3-5).

(34)

The following table shows and compares the most common simulation modelling approaches.

Figure 12: Simulation modelling approaches (Robinson, 2004, 42).

Together with other researchers, who worked on the model and used it for related studies, the use of a specialist simulation software was chosen. More specific the program Plant

Simulation from Siemens was used (Plant Simulation, 2017).

3.2 Model description

Description of the used systems and the shop floor topology

(35)

Figure 13: Schematic representation m-CONWIP and CONWIP system

As it can be seen, the systems share the same amount and positions of work stations. The systems consist of 7 work stations, starting with one in the first stage and doubling at every consecutive stage. This results in 4 different stations at the last stage, which equals the 4 possible different routes which are all each likely to occur. A product needs only one of this routes to get produced and every station only needs to be visited once and in the displayed order. This shop floor topology displays a MTO environment, where products need different production steps to be completed.

The 7 workstations of this system all have the same capacity and only one product can be processed at a time. The time a product needs to be processed doubles at every stage. For

(36)

The same configuration as in the research from Bodnar (2016), Ziengs et. Al. (2012) and Germs & Riezebos (2010) is used, as the results of the simulation should be comparable.

The m-CONWIP system will be used as it has a simple and clear loop structure, but still is able to use the proposed release mechanisms. The simple structure makes it easier to research the effects of the release mechanisms and explore the performance under processing time and route variety. Further, the outcome of the m-CONWIP system can be compared to the results of Bodnar (2016) and Germs & Riezebos (2010), who also used this system.

The CONWIP system will be used as a reference system and to simulate the push system by using such a high number of cards, that the system is practically unrestricted. As CONWIP is not route specific, the two described release mechanisms cannot be used with the CONWIP system.

(1) Structure of the systems

For the CONWIP system the structure is one loop that covers all stations. Starting at station 1 and ending at one of the end stations, this loop is referred as a 1:[4,5,6,7] loop. For the m-CONWIP systems the 4 end stations result in 4 different loops. The four loops are 1:4, 1:5, 1:6 and 1:7. For each incoming order the connected product needs to follow a specific route, which is known as soon as the order arrives. For the m-CONWIP system this means the order is assigned to a specific loop, whereas for the CONWIP system all routes are in the same loop.

(37)

The loops are important for allocating a card to an order. As two different card types are used, first the process of card allocation in the loops is described for the unit-based cards and afterwards for the load-based cards. The card attachment is done by the order release mechanism in front of the first workstation.

For the CONWIP system orders wait until one unit-based card or the needed amount of load-based cards are free, the cards get attached, then the order gets processed and the cards get released once the order leaves the system. The cards are not route specific, which is why this system is not able to balance the workload (Ziengs et. al., 2012).

As already mentioned, the route of the order is known as soon as it arrives. This information is used in the m-CONWIP system, which can be unit-based or load-based. In the unit-based card system every card represents one unit and a specific route, e.g. a card with the route 1:4 is in the correspondingly loop and only can be used for the route starting at station 1 and ending 4. Once the order arrives it has to wait until a card for its specific route is free. The card gets attached to the order and the order gets into line for station 1, to be processed. It then follows its specific route. Once it leaves the station at the last stage, the card gets detached and moved to the order release mechanism were it either gets attached to the next waiting specific order or waits until a corresponding order arrives.

(38)

available. Once these 2 cards are available, they get attached and the order gets processed. The cards get detached once the order leaves the system at the last stage.

(3) Release mechanisms

The order release mechanism of the pull system is responsible for attaching the cards to the orders, following the prioritization procedures. This rules influence the workload balancing capability of the system. For the CONWIP system the only rule is just to start one order after another depending on which came in first. For the m-CONWIP systems more complex

mechanisms are used. The for the m-CONWIP systems used release mechanism were described in chapter 2.4

For the comparisons of the performance of this release mechanisms, different

experiments will be conducted. This experiments and the parameters of the simulation will be explained in the next chapter

3.3 System parameter and experiment designs

(39)

In the experiments the CONWIP and the m-CONWIP systems will be tested with unit- and load-based cards and the different explained release mechanisms.

Every experiment will be conducted with different card counts to find the optimal number. This results in N different versions per experiment, where N is higher than the found optimum. According to figure 3, TTT time will be rising on both sides of the optimal card count and thus the optimal should be easy to identify over the TTT measurement. An unlimited number of cards will be used to simulate the unrestricted system.

For the simulation of different processing times, the Erlang distribution will be used. This distribution is often used in simulations to model the time for a task to be completed (Robinson,

Experimental factor Reserached settings

Systems CONWIP & m-CONWIP

Release mechanisms Release mechanisms „one“ & „two“

Type of cards Unit-based & load-based

Number of cards 1-N & ∞

Inter-arrival times Constant & random (negative exponential)

Utilization 80% & 90%

Batch size 1

Processing times Constant & random (Erlang-2 distributed)

Order ratios 20/80, 50/50 & 80/20

05/95, 10/90, 15/85, 25/75 & 30/70

(40)

2004). The Erlang-2 distribution will be divided into 2 groups of processing times, where the small needs 1 load-based card to be processed and the large 2 cards. The threshold between these two groups is the maximal amount of processing times displayed by one card. The research of Bodnar (2016) showed that no differences in TTT reduction could be measured when 3 instead of 2 groups have been used.

The size of the groups will differ to research the effect of different ratios of large to small orders on the performance. To alternate the ratio of large to small orders, the threshold between the groups will be changed for different experiments. With alternating the amount of processing time displayed by one card but keeping the distribution of processing times the same, the

utilization and arrival times stay the same while the ratio of large to small orders changes. Initially 3 different ratios with will be used. For the first ratio, a threshold of 23000 seconds will be used which results in a ratio of 50/50. For the second ratio, the amount of

processing time displayed by a card is 33800 resulting in a 20/80 ratio and last 15300 for a 80/20 ratio.

The inter-arrival time will be constant and negative exponential with an arrival rate that is fitted to cause a utilization rate of 80% or 90% for the stations. The batch size is set to 1.

(41)

length of the experiment runs, ten times the value of the warmup period was used (Robinson, 2004 p.143 - 149). For every experiment 50 runs were performed to gain a good estimate of the mean performance (Robinson, 2004 p.151 - 152).

Sets of experiments

(42)

Experimental factor

Set 1 Set 2 Set 3 Set 4

Systems CONWIP & m-CONWIP CONWIP & m-CONWIP CONWIP & m-CONWIP m-CONWIP Release mechanisms

„one“ & „two“ „one“ & „two“ „one“ & „two“ „one“ & „two“

Type of cards Unit-based & load-based Unit-based & load-based Unit-based & load-based Load-based

Number of cards 1-N & ∞ 1-N & ∞ 1-N & ∞ 1-N & ∞ Inter-arrival

times

Random Random Constant Constant

Utilization 80% 80% 80% 80% & 90%

Batch size 1 1 1 1

Processing times Constant Random Random Random

Order ratios Not applicable with constant processing times 50/50, 20/80 & 80/20 50/50, 20/80 & 80/20 30/70, 25/75, 20/80, 15/85, 10/90, 05/95 & 01,5/98,5

Table 2: Sets of experiments

(43)

The outcomes of this set will be compared to the paper of Germs & Riezebos (2010) to validate the simulation.

The second set is simulated to investigate the workload balancing capabilities of the different systems and release mechanisms in an environment with random processing times.

As Germs & Riezebos (2010) stated, random inter-arrival times increase the workload balancing capabilities of systems. Set 3 is done to investigate how the change from random to constant inter-arrival times influences the workload balancing capabilities of the systems.

Last, set 4 simulates a lot of different order ratios. This is done with the purpose to

further investigate how the release mechanisms react to the different ratios. In this set interarrival rates were set to result in 80% and 90% utilization in order to see how much the different

utilizations effect the workload balancing capabilities. Further, the utilization of 90% was necessary to compare the results to the results of Bodnar (2016).

(44)

4. Results

In the following chapter the results of the different experiment sets will be displayed and elucidated.

4.1 Set 1: Negative exponential inter arrival rates and constant processing times

The first set of experiments is used to show the workload balancing capabilities of the different tested systems. For the load-based systems only one card is necessary for every order, as the processing times are constant. The table 4.1 shows the total throughput times for every system compared to the unrestricted push system.

TTT in % STT in % No. of cards

Push 100% 100% ∞

CONWIP 100,23% 97,90% 24

mCONWIP unit. 89,69% 61,98% 3 per route

mCONWIP load. one 89,69% 61,98% 3 per route

mCONWIP load. two 89,69% 61,98% 3 per route

Table 3: Results random arrival rates and constant processing times

(45)

For the m-CONWIP system all different configurations achieve exactly the same STT and TTT reduction. With constant processing times, every order needs only one card in the load-based system and by this the unit- and load-load-based systems function similar. The workload balancing capability is effective and reaches a decent TTT reduction of 10,31%.

The following image shows an extract of a graph. In this graph the total throughput time and the shop floor throughput time of the different systems is displayed for different card counts. The card count is low at the right side of the graph and rises to the left.

As the m-CONWIP system all have the same results, they are all represented by one line. The push system is used as a reference and thus is at 100% STT and 100% TTT.

Figure 14: Graph of the different systems: random arrival rates and constant processing times 85,00% 90,00% 95,00% 100,00% 105,00% 110,00% 115,00% 120,00% 45,00% 55,00% 65,00% 75,00% 85,00% 95,00% 105,00% TTT STT

(46)

The m-CONWIP line shows the typical form (compared to figure 3) of a system that is able to balance the workload effectively and has a low point of the TTT at the card count of 3 card per route which equals 12 cards in total. The CONWIP line does not show this typical form as it is not able to balance the workload effectively.

In summary, the CONWIP system does not achieve any TTT reduction. The m-CONWIP systems all have the same results and show a TTT reduction of 10,31% compared to the push system. This provides evidence that those systems are able to balance the workload effectively. The observed reductions in TTT and STT of 10,31% and 38,02% are similar to the findings of Germs & Riezebos (2010). They measured a TTT reduction of 9,27% and 36,32% for the STT. In addition, the graph in figure 14 shows the same form compared to the graph of Germs & Riezebos (2010), which can be found in Appendix B.

This comparison verifies the outcomes and validity of the built simulation.

4.2 Set 2: Negative exponential inter arrival rates and random processing times

(47)

Figure 15:Graph of the different systems: random arrival rates and processing times

As it can be seen, the systems are able to reduce the TTT but only slightly. This already indicates that systems do not manage the workload very effectively under the influence of random processing times. This is also evident from the numbers in the following table.

80000 100000 120000 140000 160000 180000 200000 40000 50000 60000 70000 80000 90000 100000 T O T A L T H RP U G H P U T T IME

SHOPFLOOR THROUGHPUT TIME

(48)

50 % large / 50 % small TTT in % STT in % No. of cards Push 100% 100,00% ∞ CONWIP unit. 100,01% 99,89% 50 CONWIP load. 101,01 97,38% 50

m-CONWIP unit. 98,45% 85,60% 8 per route

m-CONWIP load. one 98,18% 86,16% 13 per route

m-CONWIP load. two 97,20% 83,90% 12 per route

Table 4: Results random arrival rates and processing times

The random processing times decreases the TTT reduction from over 10% in set 1 to about 1.5% for the unit-based m-CONWIP system, 1.8% for load-based “one” and 2.8% for load-based “two”. The STT reduction of the systems are way lower than in the experiment set 1.

The card-count of every system increases heavily, especially for the load-based systems. This indicates that the systems need higher WIP to achieve a TTT reduction and the load-based systems of course need more cards, as with random processing times a part of the orders need 2 cards to be processed.

Comparing the TTT reductions of the different m-CONWIP systems shows that the load-based systems achieve higher reductions than the unit-load-based system. This is as expected as those systems use the processing time information of each order and are supposed to use this

information to balance the workload more effectively. The load-based system “two” outperforms the load-based system “one”. This provides evidence for the statements of Bodnar (2016)

(49)

not occurring in the load-based system “two”, the performance of the system was assumed to be better.

Overall, the TTT reductions of the m-CONWIP systems are small but the systems achieve a better performance than the push system. This demonstrates that effective workload balancing under processing time variability and negative exponential arrival rates is supported.

Further different amounts of workload represented by one card were simulated. This means that the ratio of small to large orders, which was 50/50 in the previous experiments, changes. The initial amount of workload represented by one card were 23000 seconds, in the following tables its 33800 which results in a ratio of 20/80 and then 15300 which results in a ratio of 80/20.

20 % large / 80% small

Table 5: Results random arrival rates and processing times. 20/80 ratio

80 % large / 20% small

(50)

Again, the card counts for both systems change and depend on the ratio of small to large orders. Because a lower amount of processing time is displayed by one card, and so more cards are needed to send an order with a certain amount of processing time onto the shop floor and more orders are classified as large orders that need 2 cards.

The m-CONWIP system “one” shows small differences for the TTT reductions with the different ratios. Unexpectedly, the extremes of 20/80 as well as the 80/20 ratio have both slightly lower TTTs. For the 20/80 ratio, this might be explained as with less large orders, the percentage of time the routes are blocked is lower. The time, the system is able to balance the workload effectively, increases and thus the TTT reduction increases as well.

The reduction of the 80/20 ratio compared to the 50/50 ratio suggests that also for this ratio the workload balancing capability is increased. With a higher proportion of large orders, more often large orders will have to wait for a second card to be attached before they can get released. But as the amount of small orders is lower, there might be no small orders in the order pool to be released. If this is the case, the large order is not blocking the route, as no order is there that could be released instead of the large order. Subsequently the amount of time the routes are really blocked, is less than with a 50/50 ratio and the workload balancing capability is improved.

(51)

one card is available, but this card cannot be used for a small order as there is none with this specific route in the order pool. In this time the free capacity displayed by one card is not used, which decreases the performance of the system. However, at a certain amount of large orders, this does not seem to influence the performance of the system anymore, as the performance of the 50/50 and the 80/20 ratio are the same.

In summary, the workload balancing capabilities of the load based systems depend on the card values and thus the amount of orders which are rated as small or large orders. For the

system “one”, the extreme ratios improve the workload balancing capability. Whereas for system “two”, a lower amount of larger orders yields better performance. Logically the number of cards of the optimal solution of each systems increases when the amount of large orders gets higher.

Next, the outcome of experiments with constant arrival rate will be presented to analyze how the arrival rate effects these capabilities to reduce the TTT.

4.3 Set 3: Constant inter arrival rate and random processing time

(52)

50 % large / 50 % small TTT in % STT in % No. of cards Push 100,00% 100,00% ∞ CONWIP unit. 100,00% 100,00% 38 CONWIP load. 100,03% 99,98% 50

m-CONWIP unit. 100,02% 100,02% 20 per route

Table 7: Results constant arrival rates and random processing times.

The comparisons of the TTT and STT shows that the systems are not able to reduce those significantly. The capabilities to reduce the TTT were diminished. These outcomes are in line with the findings of Germs & Riezebos (2010). They stated that with constant arrival times the length of the order pool is shorter. With less orders in the order pool, the systems capability to balance the workload effectively gets diminished as the amount of decisions and the variety of orders to choose from, when releasing an order, are getting lower (Germs & Riezebos, 2010). Only the m-CONWIP “two” system is able to achieve a slight reduction of TTT.

(53)

A comparison of the card counts to the outcomes of chapter 4.2 shows that also the card numbers are increased. The m-CONWIP “two” system needs less cards than the m-CONWIP “one” system to achieve the optimal solution. However, similar to the graph in picture 15 the TTT outcome of the different card counts are similar. For example, for the m-CONWIP “one” system the difference in TTT between 22 and 29 cards is only 0,1%.

In the following two tables the outcomes of the different ratio experiments are displayed. 20 % large / 80%

small

Table 8: Results constant arrival rates and random processing times 20/80 ratio.

80 % large / 20% small

Table 9: Results constant arrival rates and random processing times 80/20 ratio.

(54)

Looking at all the outcomes with constant arrival times, it can be said that the tested systems were not able to balance the workload very effectively. Only the load-based system “two” showed a small workload balancing capability. Besides that, more cards were needed to achieve the optimal results. This indicates that the systems need a higher WIP level to guarantee a certain amount of utilization for the workstations when the orders have random processing times.

4.4 Set 4: Experiments with different ratios of large to small orders and 80% & 90% utilization

(55)

TTT in % No. of cards TTT in % No. of cards

Ratio 30/70 80% utilization 90% utilization

m-CONWIP load. one 99,99% 25 per route 99,93% 46 per route

m-CONWIP load. two 99,81% 15 per route 99,30% 18 per route

Ratio 25/75

Ratio 20/80

m-CONWIP load. one 100% 28 per route 99,91% 34 per route

Ratio 15/85

Ratio 10/90

Ratio 05/95

(56)

m-CONWIP load. two 99,31% 7 per route 97,93% 11 per route Ratio 1,5/88,5

m-CONWIP load. one 100% 23 per route 100% 37 per route

m-CONWIP load. two 100% 23 per route 99,35% 12 per route

Table 10: Results constant arrival rates and random processing times for different ratios.

Results load-based system “one”

(57)

Figure 16: TTT reductions of different ratios for load-based system “one”.

For the experiments with negative exponential arrival rates in set 2, the system achieved better results for extreme ratios. Contrary to this, the system does not perform well with high proportions of large orders and constant inter arrival times.

Results load-based system “two”

Again, the load-based system “two” achieves better results in TTT reduction and needs less cards for the optimal solution than the system “one”. But the capability to reduce TTT is still lower than 1% for a utilization of 80%. As expected, the reduction of TTT increases when the utilization increases and more cards are needed to achieve this reduction. The highest TTT reduction (for constant interarrival rates) of 2,07% was achieved with a ratio of 05/95 and 90% utilization. 99,75% 99,80% 99,85% 99,90% 99,95% 100,00% 100,05% 100,10% 80,00% 50,00% 30,00% 25,00% 20,00% 15,00% 10,00% 5,00% 1,50% T T T r ed u ctio n

percentage of larger orders

(58)

As it can be seen in the following graph the capability of the system “two” to reduce the TTT increases with less percentage of large orders. Or the other way around, if the proportion of small orders rises, the system balances the workload more effectively. But at a certain very small proportion of large orders, e.g. a ratio of 1,5/87,5, the systems capability to reduce TTT

diminishes.

Figure 17: TTT reductions of different ratios for load-based system “two”.

For 80% utilization, the highest TTT reduction is measured under a 10/90 ratio. With lower proportions of large orders the TTT increases again. For the 90% utilization, the reduction increases with less large orders and shows the highest reduction at a 5/95 ratio.

The experiments show that the optimal ratio of orders depends on the utilization of the system and with this on the length of the queues. Apparently when the queues are longer (90% utilization) the optimal ratio has less large orders then when the queues are shorter (80% utilization). 96,50% 97,00% 97,50% 98,00% 98,50% 99,00% 99,50% 100,00% 100,50% 80,00% 50,00% 30,00% 25,00% 20,00% 15,00% 10,00% 5,00% 1,50% T T T r ed u ctio n

percentage of larger orders

(59)

Both systems do not perform well when the proportion of large orders is extremely small. This is logical as with almost no large orders, nearly all orders need only one card, which then assembles a unit-based system. Unit-based systems are not able to balance the workload effectively with random processing times (Germs & Riezebos, 2010).

5. Discussion

In this section the results will be discussed.

A comparison of the TTT and STT values of the two load-based systems shows that the system “two”, which prioritizes small orders over large orders, always achieves a better

performance than the load-based system “one”. This means that the release mechanism “two” balances the workload more effectively. The load-based system “two” achieves 2,07% TTT reduction whereas the load-based system “one” only achieves 0,14% as best result under constant inter arrival rates.

This improved performance compare to the system “one” indicates that this release mechanism is able to balance the workload within and between the routes. With the ability to balance the workload not only between but also within the routes, the load-based system “two” is able to achieve superior TTT reductions than the system researched by Bodnar (2016), which only balances the workload between the routes. In addition, the system “two” needs less cards and with this less WIP to achieve the TTT reductions.

(60)

the TTT up to 5,18% under a 20/80 ratio and 90% utilization. This capability to balance the workload effectively could not be confirmed as this particular system did only achieve 0,09% TTT reduction under this specific conditions. Reasons for this divergence can be different implementations of the system or methodological differences. As Bodnar (2016) measured 2,6% TTT reduction for the CONWIP system and no reduction capability for the CONWIP system could be measured in this research, different implementations of the pull systems seems likely.

Conducting the different sets of experiments showed that the workload balancing capability of the systems are influenced by the randomness of the inter arrival times, utilization and the different order ratios.

For the inter arrival times and the utilization these dependencies were already stated by Germs & Riezebos (2010). With random inter arrival rates and higher utilizations the queues in the order pool and on the shop floor are getting longer which improves the workload balancing capabilities of the systems.

The results of the different experiment set and especially set 4 show that the TTT

reduction capability of the systems depends on the order ratio. For the load-based system “two”, the reduction under the 05/95 ratio was 10 times larger than the reduction under the 80/20 ratio. The graphs in chapter 4.4 show that fewer small orders improve the workload balancing

(61)

based system “one” this turning point is at a higher proportion of large orders than for the load-based system “two”. In general, it can be said that the optimal ratio depends on the utilization rate and the system itself. Further, it might depend on the queue lengths.

The exact connections between those parameters and the performance of the system could not be researched. Further research is suggested into how other pull systems react to different order ratios or product mixes.

All the simulated m-CONWIP systems achieve a way higher TTT reduction under

constant processing times than under random processing times, even for the best order ratio. This demonstrates that also for the best performing load-based system “two”, random processing times influence the effective workload balancing capability negatively. Contrary to this,

randomness in arrival rates increases the capability to balance the workload effectively (Germs & Riezebos, 2010).

This means further research is needed to find the reason why the m-CONWIP systems are not able to balance the workload as effectively under the influence of random processing times as they are under constant processing times.

It might be interesting to research at which “amount of randomness” the systems start to lose their capability to reduce the TTT. A research about how the systems react to certain

(62)

known how much variability needs to be eliminated to use a certain system or to reach a certain TTT reduction.

Pull systems have more benefits than only reducing TTT and STT. With limited WIP on the shop floor the number of unfinished products on the shop floor is decreased which reduces the capital costs of the production. With a more stable amount of WIP and shorter queues the production flow gets smoother and enables an improved quality of the manufacturing process (Hopp & Spearman, 2004).

As the m-CONWIP system “two” needs less cards to balance the workload and thus less WIP, the described benefits of this systems are superior to those of system “one” and the other systems. Although it cannot be said how high the benefits are in a practical case, they may have a high influence of the production costs.

(63)

For the theoretical implications, it can be said that balancing the workload within the routes through prioritization of small orders effects the workload balancing capability of load-based m-CONWIP systems positively. This capability strongly depends on the ratio of large to small orders. This shows that simple card-based pull systems can overcome the problem with processing time variability. For further development of this systems the proposed further research should be conducted. This research should include the effect of the prioritization rules on pull systems with a more complex loop structure like POLCA.

The managerial implications are that some load-based pull systems can be used in the MTO environment and yield positive results for the TTT reduction. Important for the

implementation are load-based cards and the release mechanism with prioritization rule. For an improved performance, the processing time displayed by one card should have a certain value that results in the optimal ratio of large to small orders.

This research and experiments have some limitations as they only focused on the

(64)

6. Conclusion

In the modern MTO sector throughput time performance is very important to gain advantages. But most pull-systems struggle to reduce TTT under the influence of random processing times which is typically for MTO companies (Germs & Riezebos, 2010). In order to reduce TTT the pull-system needs to be able to balance the workload effectively (Land, 2004). In this research the performance of a load-based pull system that is able to balance the workload not only between routes but also within routes was investigated. The ability to balance the workload within the routes was implemented through order prioritization based on their size.

The goal of this research was to answer the research question: To what degree does balancing workload between and within routes improve the effective workload balancing capability of pull systems?

The results show that prioritizing small orders over large orders and thus balancing the workload within the routes has a positive effect on the effective workload balancing capability of the system. Compared to the workload balancing capability of only between route balancing systems, the advanced system strongly improves the capability to reduce TTT. Besides reducing the TTT this prioritization rule reduces the needed number of cards. What is more, the results have shown that the capability to reduce TTT depends on the ratio of large to small orders and that ratios with less large orders yield higher TTT reductions.

(65)

(66)

References

Bodnar, F. (2016). Card based production planning and control: Investigating workload balancing capability of load-based pull systems. Master thesis University of Groningen.

Fredendall, L. D., Ojha, D., & Wayne Patterson, J. (2010). Concerning the theory of workload control. European Journal of Operational Research, 201(1), 99–111.

Gaury, E. G. A. (2000). Designing pull production control systems: Customization and robustness. Tilburg University.

Gaury, A. E. G. A., Kleijnen, J. P. C., Pierreval, H., Gaury, E. G. A., Kleijnen, J. P. C., & Pierreval, H. (2001). A Methodology to Customize Pull Control Systems.: Palgrave Macmillan

Journals on behalf of the Operational Research Society Stable.

Germs, R., & Riezebos, J. (2010). Workload balancing capability of pull systems in MTO production. International Journal of Production Research, 48(8), 2345–2360.

González-R, P. L., Framinan, J. M., & Pierreval, H. (2012). Token-based pull production control systems: An introductory overview. Journal of Intelligent Manufacturing, 23(1), 5–22.

Hopp, W. J., & Spearman, M. L. (2000). Factory Physics. Mcgraw-Hill Higher

Education. Boston.

Hopp, W. J., & Spearman, M. L. (2004). To Pull or Not to Pull: What Is the Question?

Manufacturing & Service Operations Management, 6(2), 133–148.

Land, M.J., & Gaalman, G.J.C., (1996). Workload control concepts in job shops: A critical assessment. International Journal of Production Economics 46, 535 – 548.

Land, M. J., & Gaalman, G. J. C. (1998). The performance of workload control concepts in job shops: Improving the release method. International Journal of Production Economics, 56–

(67)

Land, M.J., 2004. Workload control concepts in job shops, grasping the tap. Dissertation,

University of Groningen, The Netherlands.

Liberopoulos, G., & Dallery, Y. (2000). A unified framework for pull control mechanisms in multi-stage manufacturing systems, 93, 325–355.

Moreira, M. R. a., & Alves, R. a. F. S. (2009). A methodology for planning and

controlling workload in a job-shop: a four-way decision-making problem. International Journal

of Production Research, 47(10), 2805–2821.

Plant Simulation. (2017). Plant Simulation: Siemens PLM Software. Retrieved January 15, 2017, from

https://www.plm.automation.siemens.com/en_us/products/tecnomatix/manufacturing-simulation/material-flow/plant-simulation.shtml

Riezebos, J. (2010). Design of POLCA Material Control Systems. Journal, International

Production, O F, 27.

Robinson, S. (2004). Simulation: The Practice of Model. John Wiley & Sons Ltd.

Chichester.

Stevenson, M; Hendry, L. C.; & Kingsman, B. G. (2005). A review of production planning and control: the applicability of key concepts to the make-to-order industry, 54(5).

Thurer, M., Silva, C., & Stevenson, M. (2010). Workload control release mechanisms: from practice back to theory building. International Journal of Production Research, 48(12), 3593–3617.

(68)

Thürer, M., Land, M. J., Stevenson, M., & Fredendall, L. D. (2016). Card-based delivery date promising in high-variety manufacturing with order release control. International Journal of

Production Economics, 172, 19–30. http://doi.org/10.1016/j.ijpe.2015.11.003

Ziengs, N., Riezebos, J., & Germs, R. (2012). Placement of Effective Work-In-Progress Limits in Route-Specific Unit-Based Pull Systems. International Journal of Production

(69)

Appendix A

Graph of the Time Series Method:

Graph of the Welch Method: 0 50000 100000 150000 200000 250000 0,055033919 2,255446564 4,509762945 6,755285137 8,990700977 11,49300879 M ean of ob se rvation Periods in Days Time Series Method

(70)

Appendix B

[Workload balancing capabilities of load based pull production control systems] 1

Table of Contents