"... In PAGE 2: ... Section 2.3 discusses the lookup and ac- tion overheads of the systems in Table1 . Section 2.... ..."

"... In PAGE 15: ... First the sending/receiving of user/control data between the flows and the A-modules, and then the sending of both events and noti- fications. These data types have different properties and are illustrated in Table2 on page 15. The most important differences between the two channel types is the packet size (either fixed or varia- ble) and the number of concurrent processes (threads in our context) that simultaneously try to write data.... In PAGE 68: ... Each class that implements the functionality of one dedicated File Client can be named a File Client type class. The base class for all these classes is called ClCtrlComp and has the interface according to Table2 0, whereas all methods are virtual and abstract. All File Client type classes are inherited from this class and have to implement above mentioned three... In PAGE 80: ... Those functions have to be pure C . As the Video Viewer was designed and implemented following an object-oriented approach to allow for the overloading of functions, it was decided to imple- ment the attach and detach function for the sharing as methods (see also Table2 3). Pure C functions were implemented and their pointers are used as parameters in the GM_Register function.... ..."

"... In PAGE 15: ... First the sending/receiving of user/control data between the flows and the A-modules, and then the sending of both events and noti- fications. These data types have different properties and are illustrated in Table2 on page 15. The most important differences between the two channel types is the packet size (either fixed or varia- ble) and the number of concurrent processes (threads in our context) that simultaneously try to write data.... In PAGE 68: ... Each class that implements the functionality of one dedicated File Client can be named a File Client type class. The base class for all these classes is called ClCtrlComp and has the interface according to Table2 0, whereas all methods are virtual and abstract. All File Client type classes are inherited from this class and have to implement above mentioned three... In PAGE 80: ... Those functions have to be pure C . As the Video Viewer was designed and implemented following an object-oriented approach to allow for the overloading of functions, it was decided to imple- ment the attach and detach function for the sharing as methods (see also Table2 3). Pure C functions were implemented and their pointers are used as parameters in the GM_Register function.... ..."

"... In PAGE 4: ... The sort rsite denotes the sites that read, or input, values; the sort wsite denotes the ones the write, or output, values. The signature for the shared-memory style is Holds: variable val ! bool CallSite: site function ! bool Writes: wsite variable ! bool Puts: wsite val ! bool Reads: rsite variable ! bool Gets: rsite val ! bool Table1 contains (partial) theories associated with the two architectures in Figures 1 and 3. D denotes the data ow theory and M the shared-memory the- ory.... ..."

"... In PAGE 8: ... The first five applications shown in the table are from the SPLASH suite[30], the three following ones are from the NAS parallel benchmarks[8], the next four are engineering kernels, and the last solves a numerical problem. Table8 presents the running time and speedup performance of these applications on Alewife. The table includes results for Mod MP3D , which is a version of the original MP3D application that eliminates some useless code and improves locality by modifying the mapping of particles to processors.... In PAGE 8: ...he mapping of particles to processors. Section 4.2 discusses both the original and modified versions of MP3D in detail. All the speedups presented in Table8 are based on the parallel implementation of each program running on one processor except those that are marked in the table with asterisks and the different ver-... In PAGE 10: ...Table 8: Performance of shared-memory applications on Alewife. MP3D (see Table8 ). One possible explanation for this effect is that multithreading can tolerate the latency of replacement cache misses, which are difficult to predict when implementing software prefetch- ing.... ..."

"... In PAGE 8: ... The first five applications shown in the table are from the SPLASH suite[30], the three following ones are from the NAS parallel benchmarks[8], the next four are engineering kernels, and the last solves a numerical problem. Table8 presents the running time and speedup performance of these applications on Alewife. The table includes results for Mod MP3D , which is a version of the original MP3D application that eliminates some useless code and improves locality by modifying the mapping of particles to processors.... In PAGE 8: ...he mapping of particles to processors. Section 4.2 discusses both the original and modified versions of MP3D in detail. All the speedups presented in Table8 are based on the parallel implementation of each program running on one processor except those that are marked in the table with asterisks and the different ver-... In PAGE 10: ...Table 8: Performance of shared-memory applications on Alewife. MP3D (see Table8 ). One possible explanation for this effect is that multithreading can tolerate the latency of replacement cache misses, which are difficult to predict when implementing software prefetch- ing.... ..."