"... In PAGE 30: ... One third of these severs were 2U servers. While the data in Table5 give a sense of the types of equipment in this space, I did not try to estimate power consumption based on this information because the energy demands vary depending on the internal configuration of the equipment. While servers generally use less power per unit area than routers, one 4U server may require significantly more power ... ..."

"... In PAGE 5: ...ixed mode of communications in these dimensions (Dix et al., 1998; Tang et al., 2005). Thus far, most work had been done to develop groupware to support distributed collaboration as indicated in the survey conducted by Schmidt et al. (1998) on the contributions made to the Computer-Supported Cooperative Work (CSCW) conferences from 1986 to 1996 (Table 1) as well as our own survey on the same conference publications from 1998 to 2004 ( Table2 ). These surveys, despite not being exhaustive, do cover the primary conference and give evidence of the under-exploration in the general area of co-located collaboration and the specific area of asynchronous co-located collaboration.... ..."

"... In PAGE 14: ... Additionally, the high overhead associated with diff storage and garbage collection limits the performance of Tmk for this application [19]. Table3 indicates that by making use of available hardware coherence mechanisms, the number of messages sent drops by 177%, and there is a 182% reduction in the number of diff requests sent. Reduction in Number Msgs Reduction in Number Bytes Reduction in Diff Requests Reduction in Cold Misses Change in Segv Time Useless Faults in Tmk-Mt Barnes Hut 177% 82.... In PAGE 14: ... Making use of the available hardware coherence mechanisms allows the effects of false sharing between threads in the same process to be restricted to cache line movement, as opposed to network traffic. Table3... In PAGE 16: ...which both threads wait for a single page of data, limiting the computation/communication overlap and reducing potential performance improvements. The data under Useless Faults in Tmk-Mt in Table3 indicates how many times a thread was forced to suspend inside the access violation handler because another thread was already waiting for the same page of shared data. For Tmk-Mt, 43% of all access violations were useless in CG, as shown in Column 7 in Table 3.... In PAGE 17: ... In Tmk, where every thread is running in its own process address space, there is no such restriction. As a result, the total time spent in the access violation handler increases by 18% under Tmk-Mt, as indicated in Table3 . FFT-3D is the only application studied for which this occurs.... In PAGE 19: ...44 for Tmk-Mt. Table3 shows that 33% of the misses are useless, as another thread in the same process is waiting for the page when the second thread faults on it. This reduces parallelism and contributes to a load-balancing problem for the Tmk-Mt implementation.... ..."

"... In PAGE 14: ... Additionally, the high overhead associated with diff storage and garbage collection limits the performance of Tmk for this application [19]. Table3 indicates that by making use of available hardware coherence mechanisms, the number of messages sent drops by 177%, and there is a 182% reduction in the number of diff requests sent. Reduction in Number Msgs Reduction in Number Bytes Reduction in Diff Requests Reduction in Cold Misses Change in Segv Time Useless Faults in Tmk-Mt Barnes Hut 177% 82.... In PAGE 14: ... Making use of the available hardware coherence mechanisms allows the effects of false sharing between threads in the same process to be restricted to cache line movement, as opposed to network traffic. Table3... In PAGE 16: ...which both threads wait for a single page of data, limiting the computation/communication overlap and reducing potential performance improvements. The data under Useless Faults in Tmk-Mt in Table3 indicates how many times a thread was forced to suspend inside the access violation handler because another thread was already waiting for the same page of shared data. For Tmk-Mt, 43% of all access violations were useless in CG, as shown in Column 7 in Table 3.... In PAGE 17: ... In Tmk, where every thread is running in its own process address space, there is no such restriction. As a result, the total time spent in the access violation handler increases by 18% under Tmk-Mt, as indicated in Table3 . FFT-3D is the only application studied for which this occurs.... In PAGE 19: ...44 for Tmk-Mt. Table3 shows that 33% of the misses are useless, as another thread in the same process is waiting for the page when the second thread faults on it. This reduces parallelism and contributes to a load-balancing problem for the Tmk-Mt implementation.... ..."

"... In PAGE 4: ... Zone 1 and its buffer has no pattern. Table4 gives the patterns of the root node which is the unions of the patterns and instances of the child nodes of the root that are given in Table 3. Table 2.... ..."

"... In PAGE 1: ... Displays are typically placed at such a distance from each other that a user is only able to interact with one display at each point in time. Table1 shows a classification of distributed dis- play systems which differentiates between tiled displays, co- located distributed displays and dispersed displays. 3.... ..."