Parallel versus distributed storage and control in parallel mixed integer programming branch and bound
This talk describes implementation of classical general MIP branch-and-bound on the CM-5 family of MIMD systems. Computational results indicate that efficient parallel solution of many {open_quotes}real-world{close_quotes} MIP`s is now a reality. The most interesting issues relate to the degree of centralization in search control and data storage; the code is capable of centralized {open_quotes}master-slave{close_quotes} fully decentralized, or partially centralized decision making. A {open_quotes}quasidistributed{close_quotes} active-message storage scheme for subproblem nodes makes fully or partially centralized control of the search process practical even on fairly large configurations. For distributed search control, the implementation asynchronously layers a traditionally {open_quotes}SIMD{close_quotes} synchronous-style load balancing algorithm over the simple randomized scheme of Karp and Zhang. On average, this combination appears to work about as well as central control, but is {open_quotes}scalable{close_quotes}. A distributed scheme for storing global heuristic {open_quotes}pseudo-cost{close_quotes} information is also presented.
- OSTI ID:
- 35976
- Report Number(s):
- CONF-9408161--
- Country of Publication:
- United States
- Language:
- English
Similar Records
What have we learnt from using real parallel machines to solve real problems
The PMESC parallel programming paradigm and library