"... In PAGE 3: ...isMadeOf + hasLayer+ ProcessPartitiveAttribute StructuralPartitiveAttribute+ hasSubprocess+ PartitiveAttribute+ Figure 1: A fraction of the Galen role hierarchy node x0, called the root node, with L(x0) = fDg. T is then expanded by repeatedly applying the rules from Table1 until either the root node is marked satis able or none of the rules is applicable. If the root node is marked satis able then the algorithm returns satis able; otherwise it returns unsatis able.... ..."

"... In PAGE 3: ...isMadeOf + hasLayer+ ProcessPartitiveAttribute StructuralPartitiveAttribute+ hasSubprocess+ PartitiveAttribute+ Figure 1: A fraction of the Galen role hierarchy node x0, called the root node, with L(x0) = fDg. T is then expanded by repeatedly applying the rules from Table1 until either the root node is marked satis able or none of the rules is applicable. If the root node is marked satis able then the algorithm returns satis able; otherwise it returns unsatis able.... ..."

"... In PAGE 11: ... The results for the following benchmarks are initially reported: Matrix Multiplication, Quicksort, Takeuchi, Tower of Hanoi, Boyer (a reduced version of the Boyer-Moore theorem prover), Compiler (the Aquar- ius Prolog compiler from UC Berkeley that is approxi- mately 2,200 lines of Prolog code), Poccur (A list pro- cessing program), BT cluster (A clustering program from British Telecom, UK), Annotator (the annotator part of the amp;-Prolog parallelizing compiler that is about 1,000 lines) and, Simulator (a simulator for simulating parallel Prolog execution that is about 1,100 lines). Table1 indicates, for each benchmark, the execu- tion time (in ms.) and the speed-up obtained.... In PAGE 11: ... The speed-up gures are with respect to the optimized execution. Table1 illustrates the speedups obtained for the var- ious benchmarks. The gures clearly indicate that the current implementation, even though not completely op- timized, is quite e ective.... In PAGE 11: ... These optimizations yield, depending on the program, an improvement of 5% to 25% over the unoptimized version. Some improve- ment data is shown in Table1 (each entry in Table 1 shows the time in milliseconds before the optimization and after the optimization; the number in parenthesis gives the speed-up obtained; for compiler and simulator benchmarks the unoptimized gure is not shown). As is obvious, improvements due to optimization can be sub- stantial; in some cases superlinear speedup is obtained.... ..."