14.07.2013 Views

X - UWSpace - University of Waterloo

X - UWSpace - University of Waterloo

X - UWSpace - University of Waterloo

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

angular linear programs by fixing the dual coupling variables. Gnanendran and Ho [1993]<br />

investigated strategies for improving efficiency in distributed Dantzig-Wolfe decomposition by<br />

better balancing the load between master and subproblem processors because the parallel<br />

efficiency <strong>of</strong> the distributed approach is critically dependent on the duration <strong>of</strong> the inherently<br />

serid master phase relative to that <strong>of</strong> the bottleneck subproblem.<br />

Enviken [1996] presented the experimental results <strong>of</strong> using parallel Benden<br />

decomposition to solve staircase multistage problems on a shared memory multiprocessor<br />

cornputer which has 6 processors. He showed that paralle1 decomposition can solve a large<br />

problem with staircase structure faster than the simplex method even when serial decomposition<br />

is slower than the sirnplex method.<br />

Another use <strong>of</strong> parailel processon for decomposition alprithms is for the area <strong>of</strong><br />

stochastic pro*ming problems since the equivaient detenninistic problem <strong>of</strong> a stochastic<br />

model is typically very large. Dantzig and Glynn [1990] suggested the use <strong>of</strong> parallel processors<br />

to calculate the subproblems <strong>of</strong> Benden decomposition for stochastic models. Ruszczynski<br />

[1993] suggested parailelizing a variant <strong>of</strong> the nested decomposition algorithm by queuing<br />

subproblems for ide processors. and Birge, Donohue. Holmes and Svintsits ki [19%] tested and<br />

compared the parailel implementation <strong>of</strong> a nested decomposition algorithm for mupistage<br />

stochastic linev pro+nms over a serial implementation using PVM on a network <strong>of</strong> RS/6000<br />

model 32OH workstations connected by a local ethemet. Their computational experience on a<br />

large test set <strong>of</strong> pmcticai problems with up to 1.5 million constraïnts and almost 5 million<br />

variables showed that the paralle1 implementations worked very well but they require careful<br />

17

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!