Fixed-Parameter approximation schemes for weighted flowtime
Artículo

Open/ Download
Publication date
2018Metadata
Show full item record
Cómo citar
Wiese, Andreas
Cómo citar
Fixed-Parameter approximation schemes for weighted flowtime
Author
Abstract
Given a set of n jobs with integral release dates, processing times and weights, it is a natural and
important scheduling problem to compute a schedule that minimizes the sum of the weighted flow
times of the jobs. There are strong lower bounds for the possible approximation ratios. In the nonpreemptive case, even on a single machine the best known result is a O(
√
n)-approximation which
is best possible. In the preemptive case on m identical machines there is a O(log min{
n
m , P})-
approximation (where P denotes the maximum job size) which is also best possible.
We study the problem in the parametrized setting where our parameter k is an upper bound
on the maximum (integral) processing time and weight of a job, a standard parameter for
scheduling problems. We present a (1 + )-approximation algorithm for the preemptive and
the non-preemptive case of minimizing weighted flow time on m machines with a running time of
f(k, , m)·n
O(1), i.e., our combined parameters are k, , and m. Key to our results is to distinguish
time intervals according to whether in the optimal solution the pending jobs have large or small
total weight. Depending on this we employ dynamic programming, linear programming, greedy
routines, or combinations of the latter to compute the schedule for each respective interval.
Indexation
Artículo de publicación SCOPUS
Identifier
URI: https://repositorio.uchile.cl/handle/2250/169471
DOI: 10.4230/LIPIcs.APPROX-RANDOM.2018.28
ISSN: 18688969
Quote Item
Leibniz International Proceedings in Informatics, LIPIcs, Volumen 116, 2018
Collections