AN APPROXIMATION ALGORITHM FOR THE GENERALIZED ASSIGNMENT PROBLEM

Mathematical Programming 62 (1993) 461-474
North-Holland
461
An approximation algorithm for the generalized
assignment problem
David B. Shmoys* and Eva Tardos**
School of Operations Research and Engineering. Cornell University, 232 ET&C Building, Ithaca, NY, USA
Received 10 April 1991
Revised manuscript received 14 January 1993
The generalized assignment problem can be viewed as the following problem of scheduling parallel machines
with costs. Each job is to be processed by exactly one machine; processing job j on machine i requires time pif
and incurs a cost of c,f, each machine / is available for 7", time units,and the objective is.t»minimize the total cost
incurred. Our main result is as follows. There is a polynomial-time algorithm that, given a value C, either proves
that no feasible schedule of cost C exists, or else finds a schedule of cost at most C where each machine / is used
for at most 27", time units.
We also extend this result to a variant of the problem where, instead of a fixed processing time p,r there is a
range of possible processing times for each machine-job pair, and the cost linearly increases as (he processing
time decreases. We show that these results imply a polynomial-time 2-approximation algorithm to minimize a
weighted sum of the cost and the makespan, i.e., the maximum job completion time. We also consider the objective
of minimizing the mean job completion time. We show that there is a polynomial-timealgorithm that, given values
M and 7", either proves that no schedule of mean job completion time M and makespan /"exists, or else finds a
schedule of mean job completion time at most M and makespan at most 27".
Key words: Approximation algorithms, generalized assignment problem, scheduling unrelated parallel machines.
1. Introduction
The generalized assignment problem can be viewed as the followingproblem of scheduling
parallel machines with costs. Each of n independent jobs is to be processed by exactly one
of/n unrelated parallel machines; job j takes p^ time units when processed by machine /,
and incurs a cost c,y, /= ,...,m,j= 1,...,/?. For notational simplicity, we shall assume that
n^rn. Given values Cand T,, /= ,...,m, we wish to decide if there exists a schedule of total
cost at most C such that for each machine i, its load, the total processing time required for
Correspondence to: Prof. David B. Shmoys, School of Operations Research and Engineering, Cornell Univer-
sity, 232 ET&C Building, Ithaca, NY 14853, USA.
'Research partially supported by an NSF PYI award CCR-89-96272 with matching support from UPS, and Sun
Microsystems, and by the National Science Foundation, the Air Force Office of Scientific Research, and the Office
of Naval Research, through NSF grant DMS-8920550.
**Research supported in part by a Packard Fellowship, a Sloan Fellowship, an NSF PYI award, and by the
National Science Foundation, the Air Force Office of Scientific Research, and the Office of Naval Research,
through NSF grant DMS-8920550.

462 D. B. Shniovx, E. Turdos / Generalized assignment problem
for the jobs assigned to it, is at most T/,i= 1, ..., m. By scaling the processing times we can
assume, without loss of generality, that the machine load bounds satisfy
7, = T2= • • • =Tm = T. In other words, the generalized assignment problem is to find a
schedule of minimum cost subject to the constraint that the makespan, the maximum
machine load, is at most T. Minimizing the makespan is also an important optimization
criterion, and so we shall study this problem as a bicriteriaoptimization problem.
Lenstra, Shmoys and Tardos [4] give a polynomial-time 2-approximation algorithm for
the single criterion problem of minimizing the makespan, where a p-asproximation algo-
rithm is one that is guaranteed to produce a solution with objective function value at most
p times the optimum. In this paper, we generalize that result to the bicriteria problem
mentioned above.
Trick [8,9] and Lin and Vitter [5] consider variants of this bicriteria problem. Lin and
Vitter [5] give a polynomial-time algorithm that, given cost C, makespan 7", and £->0,
finds a solution of cost at most (1 + s)C and makespan at most (2+ /e)T, if there exists
a schedule of cost at most C and makespan at most T. In the variant considered by Trick
[8, 9], there is an interval of possible processing times, rather than a fixed timep,-,-, and the
cost of processing job j on machine / linearly increases as the processing time decreases.
Trick [9 ] focuses on the single criterion problem of minimizing a linear objective function
that is a weighted sum of the cost and the makespan, and gives a polynomial-time2.618-
approximation algorithm.
The main result of our paper is as follows.We present a polynomial-timealgorithm that,
given values C and 7, finds a schedule of cost at most C and makespan at most 27. if a
schedule of cost C and makespan T exists. As a corollary, we show a similar result for the
bicriteria problem of minimizingthe makespan and the mean job completion time; we also
give a polynomial-time 2-approximation algorithm for the variant considered by Trick.
All of the above algorithmsare based on solving linear relaxationsof a particularinteger
programming formulation, and then rounding the fractional solution to a nearby integer
solution. Whereas the results of Trick [8, 9] and Lin and Vitter [5] invoke the rounding
theorem of Lenstra, Shmoys and Tardos [4], the main contribution of this paper is the
introduction of a new rounding technique. The technique used in [4] requires that the
solution to be rounded must be a vertex of the linear relaxation. One interesting aspect of
the new technique is that it does not have this restriction.
The most time-consuming part of our approximation algorithms is the solution of the
linear relaxations. For our results that separately treat the two criteria, we observe that these
linear programs fall into the class of fractional packing problems considered in [7], and
therefore a slightly further relaxed schedule can be found by a randomized algorithm in
O(Ai2
log w) expected time, ordeterministically, in O(mn2
log n) time.
Approximation algorithms for special cases of the scheduling problems considered in
this paper have been studied over the last twenty-five years, and for a survey of this literature,
the reader is referred to [3]. Finally, we note that it is likely that our results cannot be too
substantially improved upon, since Lenstra, Shmoys and Tardos [4] have shown the fol-
lowing result for the single criterion problem of minimizing the makespan: for any e< i,
no polynomial-time (1 -he)-approximation algorithmexists, unless P = NP.

D.B. Shmoys. E. Tardos / Generalized assignment problem
2. The main result
463
We first consider the simplest version of our scheduling problem, when there is a fixed
processing time p(j and a cost cy- associated with each machine i= 1, ..., m, and each job
j=, ..., n. For any t^-T, integer solutions to the following linear program, LP(f), arein
one-to-one correspondence with schedules of cost at most C and makespan at most T.
£/,•
/= ! fory=l ..... n,
YjPijXij^T fori-l
y-i
Xy >0 for / = 1, ..., m, y:
= I , ..., n,
xtj = 0 if PJJ > t, i = 1, ..., w, j = 1 ..... /i.
Theorem 2.1. //"LP(0 /las afeasible solution, then there exists a schedule that has mak-
espan at most T+1 and cost at most C.
We will prove the theorem by providing an algorithm that converts a feasible solution .v
of LP(0 to the required schedule. We will construct a bipartite graph B(x) = (V, W, E)
and a value x'(t w) for each edge (r, w) e£. One side of the bipartite graph consists of
job nodes
The other side consists of machine nodes
V={c,(: /= 1 ..... m, s= 1 ..... A-,},
where k,- = fE/r,:/l; the k, nodes {r,v: s = 1, ...,&,•) correspond to machine /',/ = 1, ..., m.
Edges of the graph B( x) will correspond to machine-job pairs ( i,j) such that xtj > 0. For
each positive coordinate of .r, there will be one or two corresponding edges in B ( x ) . The
vector .v' defined on the edges of B(x) will have the property that, for each /'= 1 ..... m.
j= ..... n,
The cost of each edge (<.>,-.„ Wj) eE is c,j.
The graph B(x) and the vector x' are constructed in the following way. To construct the
edges incident to the nodes corresponding to machine/, sort thejobs in order of nonincreas-
ing processing time pu; for simplicity of notation, assume for the moment that

464 D.B. Shmoys. E. Tardos / Generalized assignment problem
Pi ^P,2> •
•
• >Pi,,-
If E"= i^,j< 1, then there is only one node c,, e V corresponding to machine/: in this case,
for each jfy->0, include (un, w;) e£, and set X'(L>,,, Wj) --=x(j.
Otherwise, find the minimum index j, such that Ej'_iJt</> 1. Let £ contain those edges
U'/i. "•
>) • j= 1, • •• > 7i ~ 1. f°r
which *,7>0, and for each of these set jf'(i>,-,, w,) — Ay,
Furthermore, add edge (u,-,, w;i) to £ and set jr'd',-,, vc;i) :
= 1-E^_~|V(f/i. w;). This
ensures that the sum of the components of x' for edges incident to vn is exactly 1. If
EjL i *y > 1, then a fraction of the value xij{ is still unassigned, and so create an edge
(t',2, Wj,), and set
•
(£*«)-> •
^,=i '
We then proceed with jobsj>jt, i.e., those with smaller processing times on machine /,
and construct edges incident to r,2, until a total of exactly one job'is assigned to r,2, and so
forth. More precisely, for each 5= 2, ...,k / — , find the minimum index /, such that
Ej'_ iJfy > 5. Let £ contain those edges (t',v, vv,),_/' =j,_ , + 1 _/', — 1, for whichjry > 0, and
for each of these set x'(c,v, Wj) -=x{j. Furthermore, add edge (t'lv, wit) to E and set .v'
U',-,, v^j ;= 1-Ej'J'y^.-uJf^t',., ^)- If EjL|Jfy>5, then also put edge (i1
/.,+ ,, H^ ) e£,
and set
y' be the last job assigned in this way; that is,/ =jki- •For each; >/ for which xtj >0,
create an edge (i'rti., w^) andsetjr'(t)rti, v^) :
=Xy. For each machine node <,',„ let p,™x
denote
the maximum of the processing times p/j corresponding to edges (v/,, w}) e£; let p™"
denote the analogous minimum.
We shall use the following instance to give an example of this construction: /n = 3;
n = m(m— 1) + l;p,, =m, i = 1, ...,m;ptj= 1,/= 1, ..., m,j =2, ..., n; Cy = 0,' = 1 m,
j= 1, ...,/?; C = 0; T=m. Figure 1 gives a feasible solution X=(x^) to LP(7~), and the
corresponding graph B(x).
l>13
1 / 3 1 1 0 0 0 0
1 / 3 0 0 1 1 0 0
1 / 3 0 0 0 0 1 1
For solid edges, x'(vu,,Wj) = 2/3
For dashed edges, x'(vu,,Wj) = 1/3
Fig. I. Constructing BU

D.B. Shmoyx, £. Tardos / Generalized assignment problem 465
A non-negatfve vector z on the edges of a graph is afractional matching if, for each node
u, the sum of the components of z corresponding to the edges incident to u is at most 1. The
fractional matching exactly matches a node u if the corresponding sum is exactly 1. A
fractional matching z is a matching if each component of z is 0 or 1. The following lemma
summarizes some simple properties of the above construction.
Lemma 2.2. The vector x' is a fractional matching in B(x) of cost at most C. It exactly
matches each node Wj,j= 1, ...,n, and each node vis, for each i = 1 m, s= 1, .... k,— 1.
Finally, p™n
>p™ijor each / = 1 m, s= 1, ...,*,.- 1. D
The algorithm to construct a schedule from a feasible solutionx of LP( r) is as follows.
The algorithm.
Step 1. Form the bipartite graph B ( x ) with costs on its edges.
Step 2. Find a minimum-cost(integer) matchingM that exactly matches all job nodes
inflU).
Step 3. For each edge (i1
,,, w,) eM, schedule job) on machine /.
Proof of Theorem 2.1. We shall prove that the schedule produced by the algorithm satisfies
the requirementsof the theorem. By Lemma 2.2, x' is a fractional matchingin B(x) of cost
at most C, which matches all job nodes exactly. This implies there exists an (integral)
matching M in B(x) of cost at most Cthat exactly matches all job nodes (see,for example,
[6]). Therefore, the matchingrequired in Step 2 exists and has cost at most C. The cost of
the matching is the same as the cost of the schedule constructed. Therefore, the cost of the
schedule constructed is at most C.
Next we show that the makespan of the schedule constructed is at most T+1. Consider
the load of machine /',/'= 1, ..., m. There are k/ nodes corresponding to machine / in B(x),
and for each of these, there will be at most one job scheduled on machine / corresponding
to some incident edge. Therefore, the length of the time required by machine / is at most
£*'- i pJ?ax
- Clearly,/?,™11
</. Lemma 2.2 implies that the sum of the remaining terms.
.t-2 . v = l
<E E Pvx'(i>i.*wj)-
.'- I j' ('•
„ . wj> ef j—
which proves the theorem. D
Observe that the algorithm ensures that if x,j= 1, then job j is assigned to be scheduled
on machine /, since each edge incident to Wj in B(x) is of the form (r,,, Wj) for some s. Also
note that the obvious w-machine generalization of the example given in Figure 1shows that
the analysisof this algorithm is asymptoticallytight.

466 D.B. Shmovx, E. Tardos / Generalized assignment problem
Remark. The"roundingtechniqueused to obtain Theorem 2.1 can also be used for integer
programs of a slightly more general form with varying right hand sides. Consider the
extension of the problem of scheduling unrelated parallel machines in which there are «y
jobs of type y, 7= 1 n, and each machine / is available for processing for only 7, time
units, i'=l, ..., m. Feasible schedules of cost at most C are in one-to-one correspondence
with integer solutions to the following linear program for any f,>7", for /= 1, ..., m.
El
i-lV-
^pijXij^Ti for/=! m,
;-i
Xjj^Q for / = 1, ..., m, j = I n,
x/j = 0 if p,j> 1^, i = 1, ..., m, j = 1, ..., n,
The natural extension of the proof of Theorem 2.1 shows that if this linear program has a
feasible solution, then there exists a schedule of cost at most C where the load on machine
/ is at most T, + /,, for / = 1,..., m. The immediateextensionof the construction of Theorem
2.1 is pseudopolynomial in the job multiplicities nf. there are at least £,-«, machine nodes
i'i, in the bipartite graph B ( x ) . In order to find a solutionin polynomialtime we first schedule
the integer part of the fractional schedule*; that is, we assign [.vj jobs of type7 to machine
/' for each7= 1, ..., n and /'= 1 m, which leaves fewer than mn jobs unscheduled. We
then used the fractional part of the vector x in the construction of Theorem 2.1 to schedule
the remainingjobs.
Corollary 2.3. In the problem with fixed processing times p^, i= I, ..., m,j= 1, ..., n,for
any given cost C and makespan T, we can find, in polynomial time, a schedule of cost C
and makespan at most 2T, if one of cost C and makespan T exists.
Proof. If there exists a schedule of cost at most C and makespan at most T, then LP( 7")
must have a feasible solution. We can use any polynomial-time linear programming algo-
rithm to find a feasible solution to this linear program. The algorithm used to prove Theorem
2.1 can be implementedto run in polynomialtime, which impliesthe claim. D
Corollary 2.4. In the problem with fixed processing times ptj, /=!,..., m,j= 1 n, and
non-negative costs,for any given cost C and makespan T, and for anyfixed e>0, we can
find a schedule of cost at most (1 + £•) C and makespan at most (2 + e)T, if a schedule of
cost C and makespan T exist, using a randomized algorithm that runs in expected O(n2
log
n) time.
Proof. Plotkin, Shmoys and Tardos [7] developed an algorithm that efficiently finds approx-

D.B. Shmovs, E. Tardos / Generalized assignment problem 467
imate solutions to a wide class of linearprogramming problems, known as fractional packing
problems. If the costs in LP(7) are nonnegative, then this linear program is a fractional
packing problem of the form considered in [7]. The techniques of [7] can be used to
determine that LP(T) is infeasible, or else produce a solution that is nearly feasible, in the
sense that it is feasible if the right-hand sides of the cost constraint and machine load
constraints are relaxed by a factor of 1+ e. Thus, if this algorithm produces such afractional
solution, we can then use Theorem 2.1 to find the claimed schedule.
To view the linear program LP(7) as a fractional packing problem, we partition the
constraints into two categories: the m machine load constraints and the cost constraint are
the packing constraints, and the remaining constraints are the job assignment constraints.
The algorithms of [7 ] work by maintaininga solution that satisfies the latter, anditeratively
moving towards a solution that also satisfies the former.
An important parameter of a fractional packing problem is its width, which is the maxi-
mum ratio of the right-hand side to the left-hand side of any packing constraint for any
solution x that satisfies the remaining constraints. The width of the above formulation can
be as high as E/ max, c,//C. This can be improved as follows: add constraints that set xit = 0
if Cjj > C. As a consequence, the width is reduced to at most /;. If there exists a schedule of
makespan at most T and cost at most C then this modified linear program has a feasible
solution. Furthermore, we can use the algorithm of Theorem 2.1 to round a feasible solution
to this linear program to a schedule.
Since the width is at most n and there are m -I- I packing constraints, the packing algorithm
of [7] finds an approximate solution in O(« log n) iterations (see Theorem 2.7 of [7] ).
In each iteration, the algorithm first computes a dual variable corresponding to each packing
constraint, which is completely determined by the current primal solution; let v denote the
dual variable corresponding to the cost constraint, and let^correspond to the load constraint
for machine/, i= 1, ...,m. The algorithm then selects ajob) uniformlyat random, andfinds
the machine i on which job; may be scheduled (i.e., p(i < t and c,y < C) for which vc,, + v,ptl
is minimum. A small fraction of job7 is rescheduled on this machine. Each iteration takes
O(m) time. Therefore, the packing algorithm terminates in O(mn log n) expected time.
The resulting vector* has O(n log n) nonzero coordinates. Therefore, the graph B(.)
has at most 2n + m = O(n) nodes and Q(n ogn) edges. The minimum-cost matching that
exactly matches the n job nodes can be found via n shortest path computations; using
the Fredman-Tarjan implementation of Dijkstra's algorithm, the algorithm runs in
O(n2
log n) time. D
There is also a deterministic version of the algorithm of [7] that yields a running time
of 0(mn2
log n).
Next consider the version of the problem where job) can be processed by machine / in
t,j time units, where /,y < ^< MI:> and the cost linearly increases as the processing time
decreases. In this model, we are given the minimum cost cj) and the maximum cost c'u of

468 D.B. Shmoys, £. Tardos / Generalized assignment problem
assigning job./' to machine /, for each /'= 1 m,j= 1, ..., n. The cost associated with
processing job j on machine / in time ry- is /icy- + (1 —/i)c}J if the time can be written as
tij = //,/y + (1 - /A)My, where 0 < /j.< 1.
In the linear programming relaxation LPspecd(r) of this problem there are two variables
x'jj andx"j associated with each machine-job pair (i,j), /=!,..., m,j= 1, .... n. A. feasible
solution to the linear program directly corresponds to a feasible schedule if x'^+x"; is
integral for each machine-job pair ( i , j ) . /= I >n,j= 1, ..., n. Job j is assigned to
machine / if x"j
j
!-x.j= 1, where the assigned time is tll = x"juu+xl
,jlu at a cost of CyjcJJ• +
44-
In the linear program LP(/), we forced the variable *,-,• to zero if py->r. Analogously, we
want to make sure that no job is processed on a machine at a speed on which it would
require more than t time units to process the whole job. To do this, we revise the upper
bound of the processing times to w,y =minj/, «,-,}. The revised cost c|) associated with the
revised upper bound is the cost of processing job j on machine / in time M,-,-, i.e., if
My = fJMij+ ( 1 - /!)/,>• then we set cy- = /ACJ) + ( 1 - /A)CV-. The resulting linear program
LPspeed(0 1sas
follows.
LPspeed(t): £ £ (
^^5+C
U
us+4)-' for;-!.....«,
(K^ + ^Xr f o r / = l , ....«,
l
.^y-^O f o r / = l w, 7=1, ...,
J =Xy- =0 if ltj > t, i= m, j =
Theorem 2.5. If the linear program LPspecd(/) has a feasible solution, then there exists a
schedule with makespan at most T+1 and cost at most C.
Proof. We will prove this theorem by constructing a feasible solution to a related linear
program LP(0 and then applying Theorem 2.1. Consider a feasible solution x' and x" to
LPspee.d(r). Define the corresponding feasible solutionx, schedulingtimes p, and costs c as
follows. Let
x =JT" +x' •
(I <J ' •
* ( / »
that is, Xjj is the fraction of job j that is scheduled on machine /. For any machine-job pair
(i,j) such thatjfyX), define its processing time as
that is, Pfj is the time it would take to process all of job; on machine i at the speed used in
the fractional schedule; the corresponding cost is defined to be

D.B. Shmov.t. E. Tardos / Generalized assignment problem
= ( f " u
* c
469
For machine-job pairs ( i , j ) such that jc,y = 0, set p,y= +<» (where any value greater than
T+t will suffice) and c,y = 0.
Observe that x,y > 0 implies that /?,-,- ^ t, lu ^ p{J ^ utj, and ctj is the cost of assigningjob; to
machine / for time ptj. Notice that x is a solution to the linear program LP( t) defined by T,
C, and p/j and c,y for /= 1 ..... m, _/'= 1, ..., n. Therefore, Theorem 2.1 implies that the
claimed schedule exists. D
Notice that the algorithm ensures that integral assignments (i.e., pairs (/, j) with
x'ii + x"j integral) are used in the schedule constructed.
Corollary 2.6. In the problem with variable processing times,for any given cost C and
makespan T, we can find, in polynomial time, a schedule of cost C and makespan at most
2T, if one of cost C and makespan T exists.
Proof. If there exists a schedule of cost at most C and makespan at most 7, then LPspecd( T)
must have a feasible solution. We can use any polynomial-time linear programming algo-
rithm to find a feasible solution to this linearprogram. By applying Theorem 2.5, we obtain
the corollary. D
Corollary 2.7. In the problem with variable processing times ptj, i= 1, ..., m,j= 1, .... n,
and non-negative costs,for any given cost C and makespan T, and for any fixed s>0, we
canfind a schedule of cost at most ( 1 + e) C and makespan at most (2 + s)T,ifa schedule
of cost C and makespan T exists, using a randomized algorithm that runs in expected
Q(n2
logn) time.
Proof. This proof of this result relies on techniques from [7] in a way analogous to the
proof of Corollary 2.4. To make the width of the corresponding packing problem small, we
must further restrict the allowed speeds for the assignments. We increase the lower bounds
l,j to a modified lower bound /„, if necessary, to ensure that the corresponding cost c'y is at
most C. D
3. Mean job completion time and makespan
In additionto the makespan of the schedule, another important objective is to minimizethe
meanjob completiontime M. In this section we consider the bicriteriaproblem of minimizing
the makespan and the mean job completion time with fixed processing times.
Horn [2] and Bruno, Coffmann and Sethi [ 1 ] showed that a schedule with minimum
mean job completion time can be found by reducing the problem to the minimum-cost
bipartite matchingproblem. The bipartite graph formed by the reduction is quite similar to
the graph B(x) used in the proof of Theorem 2.1. There is a node w, corresponding to each

470 D.B. Shmtiyx. £. Tardos / Generalized assigpmeni problem
job;' = 1 n, and there aren nodes r,v, s'= I n,corresponding to each machine/, /= 1,
..., m; there is an edge between every job node and machine node, and the cost of the edge
(t',v, Wj) is S'pij. Schedules correspond to matchings in this bipartite graph: a job node wt
matched to machine node c,v is interpreted as job; being scheduledsth-to-last on machine
/. Ifjob; is the Jth-to-last job on machine /', then it contributes p:j to the completiontime of
s jobs; hence, there is a one-to-one correspondence between schedules of minimum mean
job completion time and matchings that exactly match all job nodes of minimum cost.
Observe that in an optimal schedule,the jobs processed on each machine i, /= 1 m,
are sequenced in order of nondecreasing processingtime. Thus, in the constructionused in
Theorem 2.1, we can similarly view r,-, as the ^th-to-last position on machine i.
Theorem 3.1. In the problem with fixed processing times p^, i= m,;'= 1 n.for
any given mean job completion time M and makespan T, we can find, in polynomial time,
a schedule of mean job completion time M and makespan at most 2T, if one of mean job
completion time at most M and makespan at most T exists.
Proof. We shall give a polynomial-time algorithm analogousto the one used in the proof
of Theorem 2.1. If there exists a schedule with mean job completion time at most M and
makespan at most T, then the following linear program has a feasible solution by setting
x']j = I ifjob; is scheduled as the 5th-to-last job on machine /, and setting x*j=Q otherwise.
E E
£w —
/ - I ;- I
E
for /= I m,
for
'=
I. «,
xj, >0 for / = 1, ..., m, s = 1 n, j= , ..., n,
x]j=Q if pfj>T, i' = I /n. 5=1 «,;'=!, ..., n.
If there exists a schedule of mean job completion time at most M and makespan at most T.
then we can use any polynomial-time linear programming algorithm to find a feasible
solution x*j, i = I, ..., m,j= 1, ..., n, s= 1, ..., n, to this linear program. Consider a vector
x with coordinates^ = £"_, ;cj,-, / = 1,...,m,j — l,...,n.Notice thatA: satisfies all constraints
of the linear program LP(7"), except for the cost constraint. Use the construction of the
proof of Theorem 2.1 to construct the graph B(x), but let the cost of the edge (r,t, w}) e £
be s-ptj. The vector x' is a fractional matching of B(x) that matches all job nodes exactly.
We will argue that the total cost of x' is at most nM. If a set of jobs are (integrally)
assigned to be processed on machine /, then the mean job completion time on this machine

D.R. Shuiov.f. E. Tardos/Generalized assignment problem 471
is minimized by scheduling these jobs in order of noridecreasing processing time. The
following fractional analogue will suffice to bound the cost of x': each job,/' has a fraction
Xjj assigned to machine /',and the sum of these fractions, £"_,;t,y, 's at mos
t k/; we view
machine / as having k/ slots to process jobs, and each fraction xtj is to be subdivided among
these A:, job slots; each of the k,Job slots can be assigned fractions that sum to at most one:
if a fraction v of job) is assigned to the .yth-to-last job slot, it incurs a cost of ys-p(l the
minimum cost assignment can be found by working from the last job slot to the first,
scheduling the fractions ,v,:/ in nonincreasing order of /?,,, and filling the .rth-to-last slot to
capacity before starting the (s+ 1)th-to-last slot. A simple perturbation argument shows
that this procedure constructs an optimal assignment. This procedure is exactly the one used
to construct x from x. However, since the solution xs
u is a feasible assignment of cost at
most nM, the total cost of x' is at most nM.
Since x has cost at most nM, there exists a matching in B(x) that exactly matches all job
nodes and has cost at most nM, and such a matching can be found in polynomial time.
Schedule jobj as the .rth-to-last job on machine i if the job node w/ is matched to machine
node I'/, in the matching. The sum of the job completion times for this schedule is at most
nM, and therefore the schedule has mean job completion time at most M. The proof of
Theorem 2.1 implies that the makespan of the schedule is at most 27. D
By applying Theorem 3.1. to the minimum mean job completion time M*, we get the
following: if T is the minimum makespan among all schedules with mean job completion
time M*, then we can efficiently find a schedule with mean job completion time M* and
makespan at most 27. A much stronger result would state that it is possible to find a schedule
with mean job completion time M* and makespan at most 27*, where 7* denotes the
minimum makespan. This result is known for the special case when the machines are
identical: a schedule of minimum mean job completion time can be found by sequencing
the jobs in nondecreasing processing time order, and iteratively scheduling the next job on
the machine on which it would finish earliest; the same algorithm is known to be a 2-
approximation algorithm for the minimum makespan problem. We leave the following
question as an interesting open problem: is the makespan of any minimum mean completion
time schedule at most 27*?
4. Minimizing a combined objective function
In this section, we return to the setting with variable processing times, and consider the
problem of minimizing the objective function consisting of a weighted sum of the makespan
and the operating cost,
for some parameter /x > 0, where the desired makespan is no longer a part of the input. We
assume that the operating costs

D.B. Shmovs, £ Tardos / Generalized assignment problem
and for /'= 1, ..., m, j - n.
We shall also assume that the lower and upper bounds on the processing times, /,-,- and w,;,
respectively, are integral, for each machine-job pair (i,j). Let P denote the maximum of
the upper bounds on the processing times. Trick [9] gave a p-approximation algorithm for
this problem, where p is roughly 2.618. Here we use Theorem 2.5 to give a 2-approximation
algorithm.
For each value t > 0, consider the following linear program LPop,(/), where c£, as in the
previous section, is the cost of scheduling job j on machine /' so that it is processed in
My = min{r, «,-,-} time units.
subject to £^(4+4) =
' fory= 1,..., n,
i- i
£ (My-4 + /y-4) < T for i = 1 m,
x"j, for /= 1, ..., m, ;' = 1 ..... «,
4=4=0 if/y->r, /= y = l
To find a schedule with objective function value at most twice optimal, we will perform
a bisection search on the range of possible makespan values, and maintain the following
invariant: all scheduleswith objective function value less than half the objective function
value of the best schedulefound thus far must have makespanwithin the current range.The
number of iterations of this bisectionsearch can be boundedby using the following lemma
of Trick [9], which follows from a simple perturbation argument.
Lemma 4.1. Among all schedules with minimum objective function value, consider one
with minimum makespan: the makespan of this schedule is integral. D
Since the makespan of any plausible schedule is at most nP, it follows that we can
initialize the search by setting the interval to [0, n P ] . The core of the bisection search is
given by the following lemma.
Lemma 4.2. For each value t>0,one can, inpolynomial time, find a schedule with objective
function value f, and conclude that one of the following holds: (i) each schedule with
objective function value less than {/has makespan less than t, or (ii) each schedule with
objective function value less than {fhas makespan greater than t.
Proof. The algorithm works as follows. First find an optimal solution x to LPopIU) and
compute/(/). Let C and T denote the cost and makespan of this fractional solution; that is.

D.B. Shmoys, E. Tardos/Generalized assignment problem 473
f ( t ) = C+ /J,T. Since A: is a feasible solution to LPspeed(r) for these values C and T, we can
apply Theorem 2.5 to obtain a schedule. The objective function value of this schedule is at
most C+ p(T+t) =f(t) + /j,t.
We consider two cases: /(/) < /j,t or/(r) > /jit. Suppose that /(?) < fj.t. The schedule
obtained has objective function value /</(/) + ^,/< 2n,t. Any schedule with objective
function value less than j/< p,/ must clearly have makespan below /. Hence, we can
conclude that alternative (i) holds. Suppose instead that/(f) > /A?. We will show that each
schedule with makespan at most t has objective function value at least %f. The schedule
constructed has objective functionvalue/</(0 + pU<2/(0-Since/(0 isthe optimal value
of LPopt(0, each integral schedule with makespan at most t must have objective function
value at least/(f) > f. Hence, we can conclude that alternative (ii) applies. D
Observe that if/(f) = /J.t, then the algorithm can halt, since (i) and (ii) together imply
that there does not exist a schedule with objective function value less than if. By combining
Lemma 4. 1 and Lemma 4.2, we obtain the following theorem.
Theorem 4.3. For the problem of minimizing a weighted sum of the cost and the makespan
in scheduling with variable speeds, there exists a 2-approximation algorithm, which needs
to solce the linear program LPopl(f) at most log nP times. D
Note added in proof
The question at the end of Section 3 was answered by G. Rote, who provided an instance
for which all minimum mean completion time schedules have more than twice the optimal
makespan. Based on this construction, C. Hurkens showed that the performance guarantee
is, in fact, no better than log w/log log n.
Acknowledgments
We would like to thank Leslie Hall for prompting this research. Her discovery of an error
in our earlier, trivial, but fallacious proof of Corollary 2.3 led us to find this more interesting,
correct solution.
References
[ I ] J.L. Bruno. E.G. Coffmann Jr. and R. Sethi. "Scheduling independent tasks lo reduce mean finishing time,"
Communications of the ACM 17 (1974) 382-387.
|2] W.A. Horn. "Minimizing average flow time with parallel machines," Operations Research 21 ( 1973) 846-
847.

474 D.B. Shmtiys, E. Tardos / Generalized assignment problem
[3] E.L. Lawler, J.K. Lenstra, A.H.G. Rinnooy Kan and D.B. Shmoys, "Sequencing and scheduling: algorithms
and complexity," in: A.H.G. Rinnooy Kan and P. Zipkin, eds.. Handbooks in Operations Research and
Management Science, Vol. 4: Logistics of Production and Inventory (North-Holland, Amsterdam, 1993)
pp. 445-522.
[4] J.K. Lenstra, D.B. Shmoys and E. Tardos, "Approximation algorithms tor scheduling unrelated parallel
machines," Mathematical Programming 46 ( 1990) 259-271.
[5] J.-H. Lin and J.S. Vitter, "^-approximations with minimum packing constraint violation," in: Proceedings
of the 24th Annual ACM Symposium on the Theory of Computing (1992) pp. 771-782.
[6] L. Lovaszand M. Plummet,Matching Theory (Akademiai Kiado, Budapest, and North-Holland, Amsterdam,
1986).
17] S.A. Plotkin, D.B. Shmoys and E. Tardos, "Fast approximation algorithms for fractional packing and
covering problems" Technical Report 999, School of Operations Research and Industrial Engineering.
Cornell University (Ithaca, NY. 1992).
[8] M.A. Trick, "Scheduling multiple variable-speed machines," in: Proceedings of the 1st Conference on
Integer Programming and Combinatorial Optimization ( 1990) pp. 485-494.
[9] M.A. Trick, "Scheduling multiple variable-speed machines," unpublished manuscript ( 1991).

AN APPROXIMATION ALGORITHM FOR THE GENERALIZED ASSIGNMENT PROBLEM

More Related Content

Similar to AN APPROXIMATION ALGORITHM FOR THE GENERALIZED ASSIGNMENT PROBLEM (20)

More from Monica Gero (20)

Recently uploaded (20)

AN APPROXIMATION ALGORITHM FOR THE GENERALIZED ASSIGNMENT PROBLEM