Professional Documents
Culture Documents
Paper HP Case
Paper HP Case
January–February 1998 25
BURMAN, GERSHWIN, SUYEMATSU
to evaluate many designs very quickly. ment, which includes stable employment.
This empowered them by providing them HP set a target of 300,000 printers per
with the flexibility to experiment and test month for this product line at the Vancou-
their intuition about system behavior. It ver site. To maintain stable employment,
reduced the time to market, and it led to a HP constrained the size of the workforce.
more robust and optimized design. As a result, capacity was limited to ap-
Business Need proximately 200,000 printers per month
HP is a multinational manufacturer of using the existing manual methods. HP
electronic equipment with 1996 sales ex- needed new methods of production. It de-
ceeding $38 billion. The Vancouver Divi- cided to improve productivity through au-
sion (VCD) is part of HP’s Computer tomation, considering it the best method
Products Organization (the CPO group), of meeting the quality, production, cost,
which includes DeskJet products, LaserJet and management constraints. HP decided
products, and computer systems. VCD is to invest approximately $25 million in a
one of two divisions in Vancouver, Wash- new automated system for assembling the
ington, and is one of three manufacturing printer mechanism that would bring out-
hubs for the DeskJet series printers. Ink-jet put to 300,000 per month while satisfying
printers are a multibillion dollar industry the other constraints. This automated sys-
worldwide. This market, as well as the tem was code-named Eclipse.
competition for it, has grown rapidly since The Assembly System
HP introduced the DeskJet product a de- Raw materials enter the factory (Figure
cade ago. 1) through an automated material-handling
As the ink-jet printer market grew dur- system that includes an automated storage
ing the early 1990s, VCD faced the follow- and retrieval system (AS/RS). The AS/RS
ing conflicting objectives: routes materials to one of two mechanism-
(1) upholding the HP reputation for qual- assembly systems (Eclipse). Each Eclipse
ity and service; system is structured in a modular fashion
(2) meeting the increasing demand for with parts feeding into subassembly mod-
printers and improving HP’s market share ules and subassemblies feeding onto the
position; main assembly system. From the Eclipse
(3) achieving its targets for profit and rev- system, the products go to the mechanism
enue growth; and buffer and then to final assembly. The last
(4) sustaining the “HP way” of manage- step in final assembly is testing. After test-
Figure 1: Raw material enters the factory through an automated material handling system and
is then routed to the mechanism assembly area. Assembled mechanisms are temporarily stored
in a buffer. From the buffer they go to final assembly and then packaging and shipping.
INTERFACES 28:1 26
HEWLETT-PACKARD
Original System
100 Non-Buffered Processes Steps
Direction of
Material
Movement
Product
Base
Empty Subassembly Process Finished
Subassemblies Assembly
Pallet Cell Station
Figure 2: In the original system design, the base is assembled in the upper left subassembly
cell (black rectangle). It is attached to a pallet and it moves clockwise on the main loop as vari-
ous operations are performed. The first subassembly is added at the upper right. After addi-
tional operations take place and further subassemblies are added, the completed print mecha-
nism is separated from the pallet. The pallet stays in the system and the completed assemblies
are moved downstream. The main loop contains 30 automated work stations. Each subassembly
cell does approximately as much work as four main loop stations. Essentially no in-process in-
ventory space was designed in the system.
ing, the printers are automatically routed to and begin the assembly of the next mecha-
packaging and shipping. nism. There was essentially no in-process
In the original Eclipse design (Figure 2), inventory space within the Eclipse systems
a central automated pallet conveyor was themselves.
the primary method to move work from The efficiency of a machine or manufac-
station to station. Pallets received an turing system is the ratio of the working
empty product base from an off-line sub- time to the total time available. Alterna-
assembly station. Subsequently, this base tively, it is the ratio of the actual produc-
traveled on the pallet clockwise around tion during a time period to what the pro-
the conveyor, receiving processing at each duction would have been if none of the
station along the way. Every few stations, machines ever failed. The efficiency of a
another major subassembly was added to system is less than that of its least efficient
the mechanism. This continued until the machine because of the interactions among
mechanism was completed and stored in a machines. Because repairs and failures are
separate buffer for completed mechanisms random, production volume is random.
to await final assembly into a finished Estimates of efficiency predict the average
printer. This would free up the pallet to ratio over long periods. The only forms of
immediately receive another product base down-time that we consider here are ma-
January–February 1998 27
BURMAN, GERSHWIN, SUYEMATSU
chine failures and idleness due to machine time of nine seconds, or an uptime pro-
interactions. Some other interruptions, duction rate of 400 units per hour. For
such as worker absence, can be treated both systems, this comes to a total of 800
similarly. Others, such as setups (in multi- units per hour. We thought it reasonable to
product systems), must be treated assume that the actual capacity would ex-
differently. ceed 540 units per hour and that the plant
Machine interactions take the form of would achieve this if the main line assem-
starvation and blocking. If there is no buf- bly machines had efficiencies of 99 percent
fer between a pair of machines and the and yields of 99.5 percent and if the subas-
sembly systems had seven-second cycle
Machines can fail only while times and better efficiencies. Based on ap-
they are working. proximately 685 hours of available pro-
duction time per month, we estimated that
upstream machine fails, the down-stream the system could produce 369,900 units
machine is immediately starved and per month. This exceeded the target of
forced to be idle because it has nothing to 300,000 units per month.
work on. If the downstream machine fails, History of the Project
the upstream machine is idle because it is HP sponsored a simulation project to ei-
blocked. ther solidify confidence in the system de-
A manufacturing system may have sign or provide specific recommendations
none, one, or many buffers. If a produc- for improvements. However, the scope of
tion line has buffers and a machine fails, the simulation programming was much
the next buffer downstream loses material greater than the vendor had anticipated. It
while the machine downstream from it did not expect results until well after the
continues to operate. If this condition per- opportunity to affect design changes had
sists long enough, the buffer becomes passed.
empty and that machine is starved. The At this point, the HP development engi-
larger the buffer, the longer the time be- neers and managers sought help from aca-
fore it is starved. Conversely, the next buf- demia. They chose MIT because it had ap-
fer upstream of a failed machine gains ma- plicable analytic techniques and the ability
terial while the previous upstream to review the design and to propose
machine is still working. If this goes on changes to meet the objectives.
long enough, the buffer fills up and the Buzacott-Model Estimate of Eclipse
upstream machine is blocked. Thus, buf- Capacity
fers defer idleness and thereby increase Mitchell Burman, an MIT PhD candi-
production rate. Larger buffers more effec- date, worked on this project. He based his
tively decouple machines because they ab- first estimate of the efficiency of the
sorb larger disruptions; but they do so at Eclipse system on Buzacott’s [1967] zero-
the cost of increased inventory. buffer formula (appendix). He described
Each machine on the main Eclipse line his analysis in an HP report. It showed
was designed to have a constant cycle that the system could just barely produce
INTERFACES 28:1 28
HEWLETT-PACKARD
the targeted 300,000 units per month if it HP and MIT formed a team to develop
met the assumptions of isolated cycle recommendations for correcting the prob-
times and station efficiencies. lem while staying on schedule and within
The Buzacott formula is based on a set the current space allocations, and main-
of assumptions. The machines have equal, taining current labor levels.
constant operation times. The first ma- Need for Some Inventory
chine is never starved and the last is never Since the system was already under con-
blocked. Machines can fail only while they struction, the team could not change indi-
are working. There are no buffers in the vidual machines without disrupting the
system, so that when one machine fails, all system’s development cycle. Consequently,
other machines are forced to be idle until the team could not easily change the effi-
it is repaired. The efficiency is calculated ciencies of the components of the produc-
as a function of the mean time to fail and tion system. Gershwin [1994] describes the
the mean time to repair. relationship between buffer space and sys-
The basic design of the Eclipse system tem efficiency (Figure 3): when buffer space
was sound, but its performance depended is small, small increases in buffer space in-
on its meeting some design and parameter crease system efficiency dramatically. Sys-
assumptions. Some were questionable: tem efficiency asymptotically approaches
—That individual stations would achieve the efficiency of the least efficient machine
efficiencies of 99 percent, in the system as the buffer space ap-
—That station yields would be 99.5 proaches infinity. This is because buffers
percent, prevent the variability of each machine’s
—That stations would achieve constant production from blocking or starving the
cycle times of nine seconds, and others.
—That the system was tightly coupled
with little buffer space.
HP collected performance data as soon as
two of the subassembly cells had been in-
stalled and run as isolated operations. This
early data indicated that the isolated ma-
chine efficiencies were closer to 97 percent
than the needed 99 percent. In addition,
station cycle times varied much more than
anticipated and sometimes exceeded the
nine-second design target.
The Buzacott model predicted that, if all
stations exhibited this performance, the ca- Figure 3: The production rate increases as in-
pacity would be about 125,000 units per process inventory space increases. This in-
crease is rapid at first and then small. The up-
month. In this case, HP would have had to
per and lower limits are easy to calculate, but
add labor to meet the requirements, nulli- the rest of the curve requires the decomposi-
fying many of the benefits of automation. tion method.
January–February 1998 29
BURMAN, GERSHWIN, SUYEMATSU
This relationship suggested that the best one of Analytics’ main software develop-
way to improve the throughput of the sys- ment initiatives. Meanwhile, we accounted
tem would be to install limited buffers at for these limitations by using
strategic points. The goal was to dampen approximations.
the propagation of the effects of machine These approximations included break-
failures without expanding inventory ing up the problem into two parts: esti-
space excessively. We needed a method of mating the performance of the main loop,
determining what this curve actually and determining the sizes of the buffers
looked like for the Eclipse system and, needed between the subassembly systems
in turn, how much buffer space we and the main loop. In analyzing the main
needed to meet the 300,000 unit per month loop, we were able to ignore the subas-
target. sembly cells because they were faster and
Decomposition more reliable and would therefore only
Because of the magnitude and sensitivity rarely affect production once the buffers
of its business investments, HP wanted us were installed between them and the main
to be careful in determining how much line. We ignored observed differences in
buffer space it should install. Burman de- operation times and assumed the common
termined that the quick analytic models
found in the flow-line literature [Dallery The model’s estimate of
and Gershwin 1992] were appropriate for production rates were within
the task. He chose the decomposition equa- 10 percent of the actual
tions (appendix) developed by Gershwin
observations.
[1987] and the DDX decomposition algo-
rithm [Dallery, David, and Xie 1988]. operation time was that of the slowest ma-
Gershwin developed this decomposition chine. In addition, we approximated the
method under the assumption that ma- loop as a line.
chines have equal, deterministic operation We determined the sizes of buffers be-
times. He assumed failures and repairs to tween the subassembly systems and the
be random, independent of each other, main line by treating the material flow
and independent of the time since any into and out of each such buffer as if it
previous event. He assumed buffers to were in a two-machine continuous-
have finite capacities. Except for the as- material transfer line. (The first machine
sumption of equal operation times, these represented the subassembly system and
assumptions are reasonably accurate for the second represented the main line.) This
the Eclipse system. obviated the need for a long-line decom-
We used the method because it was the position and made it possible to treat the
only one available that could deal with different operation speeds in the different
finite buffers and unreliable machines. parts of Eclipse. We chose the buffers to be
Overcoming the limitation of equal opera- large enough so that the subassembly sys-
tion times was one of Burman’s main tems would, in fact, rarely affect main line
goals in his PhD work (appendix) and is production.
INTERFACES 28:1 30
HEWLETT-PACKARD
Figure 4: We recommended the addition of (1) an empty pallet buffer, (2) space for in-process
inventory between the subassembly systems and the main line, and (3) space on the main line.
January–February 1998 31
BURMAN, GERSHWIN, SUYEMATSU
because the subassembly systems had data for such items as cycle time, uptime,
shorter operation times than the main line. and downtime, and it was used exten-
Although this might appear to be a great sively by engineering and operations in
deal of inventory, 30 minutes was needed achieving the Eclipse objectives. Using the
to guarantee that subassembly failures decomposition model, we performed sen-
(that have a mean duration of several min- sitivity analysis to determine which action
utes) would rarely interrupt main-line would have the greatest impact on system
production. performance.
We suggested that HP install three other Simultaneously, Burman was asked to
buffers of approximately 12 units each one develop extensions to the Gershwin de-
quarter, half, and three quarters of the composition that
way through the line. (1) would handle different machine opera-
Based on the performance improve- tion times,
ments we estimated with the analytic (2) could handle very long lines,
models, HP made these modifications. The (3) would run in seconds,
total capital costs for the changes to both (4) had a better user interface,
Eclipse systems were $1,400,000. (5) had simple data requirements, and
System Improvement Prioritization (6) converged more reliably.
During Ramp-Up Burman completed this work successfully
By mid-1994, HP had finished the hard- by March of 1995 and described it in his
ware changes to the system. It set new tar- PhD thesis [Burman 1995].
get levels for station efficiencies based on Validation
more realistic performance levels. It still Using system data from May and June
had to achieve these efficiencies and sus- of 1995, Burman compared actual system
tain the modeled cycle times. performance to the performance predicted
In parallel to the hardware changes, we by the model he developed [Burman
developed and implemented a more accu- 1995]. The model’s estimates of production
rate and reliable information system on rates were within 10 percent of the actual
observations.
The increased throughput Actual system production was raised to
resulted in incremental between 250,000 and 300,000 per month
revenue of $280 million. with the expected productivity. When
called upon, the system has demonstrated
the Eclipse system. This system provided the ability to sustain production greater
the information engineering and opera- than 300,000 per month.
tions needed to know precisely how the HP Impact
system was performing as measured by its This technology contributed greatly to
critical parameters. This would pinpoint HP.
the exact location of any deviations and The technology we recommended for
would signal when repairs were adequate. system development helped the Vancou-
It provided both historical and real-time ver Division to meet its business commit-
INTERFACES 28:1 32
HEWLETT-PACKARD
ments during this period. The recommen- tion achieved in excess of the predicted
dations took into account the real-world manual assembly capacity of 200,000 per
constraints of schedule, space, and re- month. HP also realized additional reve-
sources. The approach to improvement nues from ancillary products, replacement
cartridges, media, and so forth.
Supply will exceed demand, Because productivity was increased by
prices will drop, and the focus up to 50 percent, assembly of the print en-
will shift to cost control. gine was cost competitive. This is a signifi-
cant benefit for the future, because the
had a great value to HP because the rec- business has left the period of explosive
ommendations were based on rigorous an- growth. Supply will exceed demand,
alytic methods. This increased HP’s will- prices will drop, and the focus will shift
ingness to take an enormous risk based on from revenue to cost control.
the recommendations. The changes re- Other Commercial Impacts
sulted in substantial incremental revenue. The success of this project led Burman
HP expects the technology to have an to found Analytics, Inc. Analytics provides
impact on its future business. The poten- manufacturing-systems design services
tial long-term benefit of this project at HP that are, in part, based on the principles of
is the establishment of a methodology for the work described here. Analytics has
designing manufacturing systems that used this methodology in a project for
have improved throughput, robustness, Johnson and Johnson, a project for Boeing,
and predictability. The Vancouver Division and an unrelated project for Hewlett-
has leveraged this design approach for Packard in Corvalis, Oregon.
next-generation manufacturing systems. It Analytics has continued to improve the
estimates that this method can increase the technology and has added features includ-
throughput of all its future manual and ing assembly modules (appendix) and a
automated systems. graphical user interface. The speed and
This project helped HP to design and im- flexibility of this technology make it an
prove the Eclipse automated print-engine ideal complement to simulation. By using
assembly system. It had an immediate im- this approach, one can make robust
pact because of the timeliness of the solu- strategic-design decisions and fine-tune
tions. It also helped HP to set a clear man- with simulation after solidifying the major
agement focus for system improvement. It system architecture.
made the issues comprehensible and quan- Impact on MIT
tifiable, and it provided a useful discipline The immediate impact at MIT was in
and language for discussing and evaluating the PhD research of Mitchell Burman,
manufacturing-system designs. which he completed in 1995. He based this
During the period when demand ex- research on the problems he observed di-
ceeded supply, the increased throughput rectly at the HP printer factory. Some of
resulted in incremental revenue of $280 these problems could be treated with
million. We base this figure on the produc- methodology already in the literature, but
January–February 1998 33
BURMAN, GERSHWIN, SUYEMATSU
INTERFACES 28:1 34
HEWLETT-PACKARD
Figure 5: In this representation of a production line, the squares represent machines and the
circles represent buffers. Material moves in the direction indicated from machine to temporary
storage buffer to machine. Machines fail at random times and stay down for random lengths of
time. The buffers are finite, so disruptions are propagated in both directions in the form of
starvation and blockage. Because of the complexity of its behavior, this system can be evalu-
ated only by simulation or by a decomposition approximation.
January–February 1998 35
BURMAN, GERSHWIN, SUYEMATSU
deterministic processing times appeared to Prentice Hall, Englewood Cliffs, New Jersey.
be prohibitively difficult, so Burman [1995] Dallery, Y.; David, R.; and Xie, X.-L. 1988, “An
used the continuous material two-machine efficient algorithm for analysis of transfer
line of Gershwin and Schick [1980] in a lines with unreliable machines and finite buf-
fers,” IIE Transactions, Vol. 20, No. 3 (Septem-
similar decomposition for long lines with
ber), pp. 280–283.
continuous material and different process- Dallery, Y. and Gershwin, S. B. 1992, “Manufac-
ing rates. (The differing processing rates turing flow line systems: A review of models
made the decomposition equations consid- and analytical results,” Queueing Systems The-
erably more difficult to derive than those ory and Applications, Special Issue on
of earlier systems.) He also adapted the Queueing Models of Manufacturing Systems,
DDX algorithm for this case (and called Volume 12, No. 1–2 (December), pp. 3–94.
the new algorithm the ADDX algorithm). Di Mascolo, M.; David, R.; and Dallery, Y. 1991,
It was also extremely fast. “Modeling and analysis of assembly systems
with unreliable machines and finite buffers,”
Decomposition of Continuous-Material
IIE Transactions, Vol. 23, No. 4 (December),
Model of Assembly-Disassembly System pp. 315–331.
Since the Hewlett-Packard production Gershwin, S. B. 1987, “An efficient decomposi-
system included assembly, we needed tion method for the approximate evaluation
more than the ADDX algorithm for pro- of tandem queues with finite storage space
duction lines. Consequently, Gershwin and and blocking,” Operations Research, Vol. 35,
Burman [1997] extended the ADDX algo- No. 2 (March–April), pp. 291–305.
rithm (as well as Gershwin’s [1991] earlier Gershwin, S. B. 1991, “Assembly/disassembly
systems: An efficient decomposition algo-
assembly-disassembly system work for
rithm for tree-structured networks,” IIE
synchronous systems and Di Mascolo, Transactions, Vol. 23, No. 4 (December), pp.
David, and Dallery’s [1991] work for 302–314.
continuous-material systems with equal Gershwin, S. B. 1994, Manufacturing Systems En-
processing rates) to assembly-disassembly gineering, Prentice Hall, Englewood Cliffs,
systems with continuous material and dif- New Jersey.
ferent processing rates. Again this algo- Gershwin, S. B. and Burman, M. H. 1997, “De-
rithm proved to be fast and practical. It composition equations and algorithm for
has been implemented with a graphical A/D systems—Derivation of equations,”
working paper.
user interface, and Hewlett-Packard is us-
Gershwin, S. B. and Schick, I. C. 1980, “Contin-
ing it as part of its standard methodology
uous model of an unreliable two-stage mate-
for designing new printer production rial flow system with a finite interstage buf-
systems. fer,” Massachusetts Institute of Technology
References Laboratory for Information and Decision Sys-
Bonvik, A. M. 1996, “Performance analysis of tems Report LIDS-R-1039, September.
manufacturing systems under hybrid control Schor, J. E. 1995, “Efficient algorithms for buffer
policies,” PhD thesis, Massachusetts Institute allocation,” MS thesis, Massachusetts Insti-
of Technology, Cambridge, Massachusetts. tute of Technology, Cambridge,
Burman, M. 1995, “New results in flow line Massachusetts.
analysis,” PhD thesis, Massachusetts Institute
of Technology, Cambridge, Massachusetts.
Buzacott, J. A. 1967, “Automatic transfer lines
with buffer stocks,” International Journal of
Production Research, Vol. 5, No. 3, pp. 183–
200.
Buzacott, J. A. and Shanthikumar, J. G. 1993,
Stochastic Models of Manufacturing Systems,
INTERFACES 28:1 36