iefeb07 realities

8/6/2019 IEFeb07 Realities

1/6

36 Industrial Engineer

Realtes ofRskSpply chas mst scrtze the kowB y M a r k L . S p e a r M a n


2/6

February 2007 37

Companies are recognizing the severity of

danger facing them by not addressing sup-

ply chain risk. Indeed, more than two-thirds

of the companies surveyed by Accenture, as

recorded in Manufacturing Business Tech-

nology, said they had experienced a supply

chain disruption from which it took morethan one week to recover. Furthermore, the

study revealed that 73 percent of the ex-

ecutives surveyed had a major disruption

in the past five years. Of those, 36 percent

took more than one month to recover.

Modern supply chains are susceptible

to such disruption for two reasons: Supply

chains are not designed with risk in mind,

and modern planning and scheduling are

not robust enough to operate under condi-

tions significantly different from those for

which they are planned. Why does it take

companies so long to recover? Why are so

many companies susceptible to disrup-tion? Are sophisticated advanced planning

and optimization (APO) systems useful?

Last-cetry spply chaplag Although material requirement plan-

ning (MRP), manufacturing resources

planning (MRP II), and even enterprise

resource planning (ERP) are now so last-

century, to use my daughters favorite

phrase, it is useful to review what worked

and what didnt. The goal of these systems

has always been to reduce inventory, im-

prove customer service, and reduce costby increasing the utilization of expensive

equipment and labor. Unfortunately, these

processes do not explicitly consider risk

and have ignored key facts regarding the

systems they try to control.

The first attempt to confront supply

chain risk dates back to the 1960s with the


3/6

creation of MRP, which provided material

plans with no consideration of capacity.

This evolved into MRP II, which provided

some capacity checking modules around

the basic MRP functionality. Interestingly,

virtually all ERP systems today (including

high-end offerings) continue to have the

same basic MRP calculations as the core

of their production plan offering. Not sur-

prisingly, a 2006 survey in CA Magazine

showed that users gave low marks to these

high-end systems for performance in dis-

tribution and manufacturing (averaging

2.5 and 2.6, respectively, out of 4.0).

There are two basic problems with these

systems: Lot sizing is done without consid-

eration of capacity and planning lead-times

are assumed to be attributes of the part.

The first issue results in conservative (i.e.,

large) lot sizes that increase inventory andreduce responsiveness. The second issue

is similar. Because the planning lead-time

does not depend on current work-in-pro-

cess levels and because being late (result-

ing in poor customer service) is worse than

being early (resulting in extra inventory),

most systems employ long lead-times. The

result, again, is more inventory and less re-

sponsiveness and is especially aggravated

when the bill-of-materials is deep.

Efforts to address these problems have

gone on for many years. Almost 20 years

ago, manufacturing logistics expert J.J.

Kanet described problems with contem-

porary planning systems that remain

today. Consequently, most companies do

not run their plants using the output of

their ERP systems but instead massage

the output using ad hoc spreadsheets.

21st-cetry spply chaplagObviously, with this kind of work-around

there exists an opportunity to sell more

sophisticated software, and for some time

now there have been numerous offerings.

These applications are typically deter-

ministic simulations of the process that

assume the demand, inventory, WIP, run

rates, and setup times are all known andthen seek to generate a schedule that is op-

timal under some specified criteria.

Three factors, however, prevent the

approach of using deterministic simula-

tion from effectively managing supply

chain risk:

The supply chain and plant have in-

herent randomness that do not al-

low for the complete specification of

a time for each job at each process

center with a given labor component.

Such detailed schedules are often

quickly out of date because of the

intrinsic variability in the system.

Moreover, they do not manage risk,

which involves random events that

may or may not happen. In the past,the unknown has been addressed us-

ing ever more detailed models requir-

ing ever more computer power. This

misses the point. Variability and risk

are facts of life and are the result of

not only process variation (some-

thing that one attempts to control)

but also unforeseen events and vari-

ability in demand (things that cannot

be controlled). The result is the sameregardless of the variability source.

Detailed scheduling can provide only

a short-term solution, and in practice,

the solution often becomes invalid

between the time it is generated and

the time the schedule is distributed

and reviewed as part of production

planning meetings.

The detailed scheduling system must

be re-run often because of the short-

term nature of the solution. More-

over, without a method for deter-

mining whether a significant change

has occurred, the schedule is often

regenerated in response to random

noise a temporary lull in demand

which is then fed back into the sys-

tem. Unfortunately, feeding back ran-

dom noise increases the variability in

the system being controlled. Because

of these problems, many companieshave turned off their advanced plan-

ning and scheduling systems after

spending a great deal of money to

install them. One of our own clients,

a biopharmaceutical company, shut

down its advanced planning and opti-

mization module after having trouble

keeping it fed.

It is impossible to find an optimal

schedule. Problems addressed are


realtes of rsk

Fgre 1. Ths otpt of a APO applcato shows a prodcto schedle ad vetory.

apoapplicationoutput


4/6

February 2007 39

mathematically characterized as NP-

hard, which means that no algorithm

exists to work in polynomial time to

solve scheduling problems. For real-

istic problems faced in modern fac-

tories and in the supply chain, there

has not been enough time since thebeginning of the universe to find an

optimal schedule regardless of the

speed of the computer. Consequently,

heuristics must be applied to gener-

ate a near-optimal schedule although

the effectiveness of these heuristics is

typically unknown for a broad range

of applications.

The result is that a great deal of comput-

er power is used to create a detailed sched-ule for a single instance that will never hap-

pen and that becomes obsolete as soon as

something unanticipated occurs.

Advanced systems today offer two

methods of planning for manufactur-

ing supply chains: What-if analysis uses

a deterministic simulation of the sup-

ply chain, and optimization of a set of

penalties (again, using a deterministic

simulation) associated with inventory,

on-time delivery, setups, and wasted ca-

pacity. This combined approach may be

tedious and less intuitive. For instance,

the monotony of what-if analysis comes

from all the detail that must be con-

sidered. Figure 1 shows the output of a

typical APO application. Visible is the

scheduled production on six machines

in one factory along with projected in-

ventory plot of one of the items pro-

duced. The planner can move jobs in theschedule and drill down on other items

to view inventory projections.

While this level of integration is im-

pressive, it is not particularly useful when

there are hundreds of machines (not to

mention all the associated labor) to con-

sider, along with thousands of individual

items each with its own demand.

Likewise, the use of penalties to deter-

mine an optimal schedule is not intuitive.

What should the penalty be for carrying

additional inventory? What is the cost of

a late order? What savings is generated by

reducing the number of setups, particular-

ly if there is no reduction in head count?

What is the cost of having idle machines?

As evidenced, huge gaps remain betweenwhat is needed and what is offered. Nev-

ertheless, a combination of existing appli-

cations with new technology can help cir-

cumscribe risk without the tediousness or

the complexity of modern APO systems.

Lessos factory physcsInformation technology cannot serve as

the solution to every business problem.

Instead of faster computers and moreintegration, a fundamental change in

managing manufacturing supply chains

is necessary.

Consider some basic factory phys-

ics principles that describe how all

manufacturing supply chains work. The

first principle is that all manufactur-

ing supply chains consist of demand

and transformation. Without demand,

there would be no market for our goods.

Without transformation, the people de-

manding those goods could simply find

and use products without having to pur-

chase them. (For example, people have

demand for air every day, but there is no

business case for selling air because its

freely available to everyone.).

Likewise, in any manufacturing sup-

ply chain there are two primitive compo-

nents: stocks and flows. The flows repre-

sent both the transformation element (anoutflow) and the demand element (an

inflow). The stocks provide a means to

separate and coordinate these two flows.

In a perfect world, transformation

would exactly meet demand. This means

that we would have perfect processes to

produce exactly what is needed and per-

fect customers to purchase exactly what

is produced. Because there is variability,

this perfect world cannot exist. To syn-chronize transformation and demand,

buffers are required. And while a stock

is one way to match demand to transfor-

mation, there are two others: extra time

and extra capacity. Another law of factory

physics is that there are only three possi-

ble buffers to interface between demand

and transformation: inventory, time, and

capacity. In other words, the transforma-

tion can satisfy demand in several ways:

immediately from stock, by making the

demand wait for the product, or by main-

taining extra capacity in order to provide

timely production upon demand.

Fgre 2. A make-to-order costat work--process le

work-in-processwithbuffers


5/6


Effective risk management is essen-

tially a matter of effective buffer deploy-

ment. It is how we determine the amount

of risk to take on and how to buffer that

risk, whether with extra inventory, extra

time to satisfy the customer, or extra ca-

pacity to have on hand to cover disrup-tions. Unfortunately, there is no single

solution. The optimal configuration of

risk and buffers will be different for vari-

ous business situations.

The aspirin operations of Bayer Corp.

and the desktop factory of Dell Corp.

have very different configurations. Bayer

is likely to have a considerable amount of

inventory with little excess capacity. Con-

versely, Dell has no finished goods inven-tory and a significant amount of extra

installed capacity. How else could Dell re-

spond in a timely way to peaks in demand

at Christmas and graduation? While nei-

ther of these situations use a significant

time buffer, Moog Corp., which supplies

custom servo valves to high-tech and

aerospace customers, requires time to

create a customized product.

Note that a flexible buffer is less cost-

ly than a permanent one. This is why

postponement the ability to delay

committing a common component into

the final product until the last possible

moment is effective in reducing in-

ventory. Likewise, having a flexible ca-

pacity buffer such as the makeup shift

used by Toyota or temporary workers

that can be called upon when demand

is higher than normal is less costly thanhaving all the workers all the time. Like-

wise, quoting due dates with variable

lead-times is more effective than stating

a constant lead-time.

So how do we use this understanding to

create a system that can accommodate risk,

provide a method for effective planning

and control, and retain the ability to see

the unknown if the need arises? Smooth

and continuous flow has been the key tothe Toyota Production System; although

achieving flow does not require that one try

to duplicate exactly what Toyota has done

in automotive assembly. If we achieve flow,

then planning and controlling produc-

tion become easy. Instead of attempting

to schedule hundreds of individual jobs

on a Gantt chart or trying to specify arti-

ficial penalties in an APO, a planner needs

simple tools to manage flow.

Figure 2 shows a supply chain seg-

ment and indicates all three buffers:

time, inventory, and capacity. Figure 2 il-

lustrates a simple linear flow, but it could

contain several levels and many routings.

The stock contains either jobs that com-

plete before their due dates (in the case

of make-to-order) or of finished goods

inventory (in the case of make-to-stock).

Coming from the left are two arrowsindicating make-to-order demand and

raw materials. These materials would

be supplied by a previous supply chain

segment. From the right comes an arrow

showing make-to-stock demand.

This simple framework can create a

new and more effective way to manage

manufacturing supply chains. The key is

to reduce the number of values that must

be monitored as well as control variablesin the problem. This means that instead

of providing a detailed schedule indicating

where each job is scheduled on a machine

during a given time slice, we determine a

set of control variables that will dynamical-

ly and automatically determine the sched-

ule as the system evolves. This approach is

called dynamic risk-based planning and

scheduling (DRPS). Under DRPS we need

only monitor projected inventory and pro-

jected service levels and then control a few

key parameters reorder points or lead-

times, production quantities (lot sizes),

installed capacity, makeup capacity, WIP

level, and the virtual queue.

A brief reflection on the design of

the system will show that if optimized

planning parameters listed above are in

place, the entire system will run well so

long as projected service and inventory

levels stay within acceptable ranges. Lotsizes have been established that mini-

mize WIP and finished goods subject to

the available capacity. Lead-times (or re-

order points) are set by considering the

trade-offs between service and inventory

levels. Therefore, jobs can move through

the flow in first-in, first-out order with-

out the need for a detailed schedule.

Safety stocks are set considering the

time in the virtual queue plus the time

realtes of rsk

Fgre 3. Statoal comparso of APO ad DRPS

SiTuATiOn

Capacty greatertha expected

Capacty less thaexpected

Demad greatertha expected

Demad less thaexpected

ERP APO

Caot work ahead ofgeerated schedle

Schedle becomescreasgly feasble,orders become late lesshgh safety lead-tme

Shortages less very hghsafety stock

ueeded orders released;vetory rses

DRPS

Ca work ahead pto defed earlestrelease dates

Shortfall trgger dcateseed for addtoalcapacty. Do ot eedhgh safety lead-tme.

Shortfall trgger dcateseed for addtoalcapacty; o shortages,mmal safety stock.

no eeded ordersreleased; vetory stays plaed rage

complexapo


6/6

February 2007 41

through the flow and with regard to the er-

ror in the forecast. Once these parameters

are set (reviewed periodically perhaps once

per month), the planner need only moni-

tor the projected inventory and service

levels. The only time action is required is

when these levels exceed set trigger points.

These trigger points indicate both insuf-

ficient capacity and more capacity than is

needed. If the virtual queue grows beyond

what is planned for in the lead-times, there

is insufficient capacity and so the extra

capacity (overtime or make-up shifts) is

applied. If the virtual queue becomes too

low, one can either reduce capacity or pull

in some jobs early. The result is not only

a much simpler supply chain to manage

but one that can automatically respond to

random changes in demand and supply

without the need to reschedule. This abilityto automatically adjust is a tremendous im-

provement and makes DRPS more effective

than the more complex APO (Figure 3).

This is good news for industrial engi-

neers since these are the only engineers

trained in the art of setting the planning

parameters. Once these parameters are

known, the system can be implemented.

Fortunately, the task can be simplified

by using the existing MRP system and

a generalized pull system known as con-

stant work-in-process (CONWIP).

The CONWIP pull protocol is a meth-

od for achieving pull production in a wide

range of production environments and

is key to flow management. This mecha-

nism maintains a nearly constant level

of WIP that is needed to keep bottleneck

processes busy by starting a new job only

when the WIP level has fallen below a

given maximum level. The result is pre-

dictable cycle times, smoother flow, and

higher throughput than would be seen in

an equivalent push (e.g., MRP) system.

The CONWIP mechanism also isolates

the variability inherent in demand from

disturbing the flow. This is accomplished

using the virtual queue. If demand goes

up, WIP does not but the virtual queue

does and the projected service levels beginto fall. Likewise, if demand begins to fall,

the virtual queue falls but WIP stays con-

stant and projected inventory levels rise.

usg MRP wth pllIt is important to note that the plans gen-

erated by good old-fashioned MRP can

be quite good if several conditions hold:

Average demand does not exceed ca-

pacity.

Jobs are released using a pull system

that prevents WIP explosions.

Cycle times are short enough to avoid

expediting.

Safety stock (or safety lead-time) is

determined by considering demand

and capacity. Batch sizes are determined consider-

ing demand and capacity to ensure

minimal WIP and low inventories.

Due dates are checked by considering

the characteristics of the pull system

and the sequence of the jobs to be

pulled.

If the system is designed well, these

conditions will hold. MRP is used in the

usual way but instead of starting jobs ac-cording to MRPs planned order release

list, the CONWIP protocol pulls them.

The entire release list becomes the virtual

queue. Then, knowing the cycle time of

the line and the rate at which jobs are be-

ing pulled into the line, we can accurately

predict when each job will complete.

Again, neither traditional systems

such as MRP nor modern systems like

APO adequately address risk issues. It is

also apparent that more IT is not the so-

lution to the problem without complete-

ly rethinking the problem. Supply chain

managers benefit greatly from a system

that is dynamic to avoid rescheduling,

risk-based to accommodate risk factors

and randomness, and that links plan-

ning and execution.

Mark L. Spearman is president and chief ex-

ecutive officer of Factory Physics Inc. Prior to founding Factory Physics, Spearman was a

professor of industrial and systems engineer-

ing in the industrial engineering department

at Georgia Tech. For over 15 years, his research

and teaching dealt with manufacturing enter-

prise integration and supply chain manage-

ment. Spearman holds a Ph.D. in industrial

engineering, is a senior member of IIE, a full

member of INFORMS, and is certified in pro-

duction and inventory management by APICS.

Fgre 4. WiP, demad, throghpt, ad cycle tme

conwipflow

CourtesyFactoryPhyscsinc.

iefeb07 realities

Documents