"... In PAGE 19: ...pplication is given in Section 6.1. Each timing was measured starting when the user selected the door from the resource list on her phone (a Nokia 6620), and ended when the door unlocked. As shown in Table1 , this delay was approximately six seconds excluding any user interaction (more on this below), with a small variance resulting from background work on the phone, such as alarms, housekeeping, and other applications. The second macrobenchmark is the time required for a user to log into a 2GHz Windows XP workstation.... In PAGE 19: ... Her capture-resilient key is then used at these intervals to create a short-lived certificate for a non-capture-resilient public key (a step which does require PIN entry) that is used to sign access requests. As such, the common case incurs only the latency of a signature with this non-capture-resilient key; the measurements in Table1 reflect this. Third, the network address for each of the computers regulating access was already stored in the resource list of the phone and so, e.... In PAGE 20: ... Among the most significant sources of delay are RSA signatures and accessing the Record Management Store (RMS), which is a sim- plified file system and the only long-term storage available to the Java VM. The left columns of Table 2 describe these costs for the Nokia 6620, the device used in the experiments described in Table1 . The 30KB RMS read and write benchmarks measure the time it takes to read and write a standard address-book picture, each about 30KB in size.... ..."

"... In PAGE 9: ...1. Table2 shows function prototypes to get an access right to a device and to register a dedicated handler thread. uio request(ioa start, ioa end, [type]): This function grants a user process a right to access I/O addresses from ioa start to ioa end.... ..."

"... In PAGE 9: ... We note that in both cases the user-level page migration engine re- duces drastically the number of remote accesses after the rst invocations of the runtime system and that the bench- marks reach very soon an optimal execution point at which the number of remote accesses is practically minimized. Table1 gives some more statistics for the two page migra- tion engines. The statistics in the table were collected from executions of BT and SP on 32 processors.... ..."

"... In PAGE 9: ... We note that in both cases the user-level page migration engine re- duces drastically the number of remote accesses after the rst invocations of the runtime system and that the bench- marks reach very soon an optimal execution point at which the number of remote accesses is practically minimized. Table1 gives some more statistics for the two page migra- tion engines. The statistics in the table were collected from executions of BT and SP on 32 processors.... ..."

"... In PAGE 6: ... Table10 shows an anomaly similar to that in the microbenchmark, verifying that the effects of L2 caching are easily visible in places other than simple file system microbenchmarks. However, these numbers have peculiarities of their own.... ..."

"... In PAGE 6: ... Table10 shows an anomaly similar to that in the microbenchmark, verifying that the effects of L2 caching are easily visible in places other than simple file system microbenchmarks. However, these numbers have peculiarities of their own.... ..."

"... In PAGE 6: ... Once this counter is approximately equal to N, the process performs the more expensive call to get the state of the last N pages on the inactive queue. As shown in the top half of Table2 , the victimList abstraction can be implemented in only 109 C statements; in fact, more than half of the code is needed to setup the memory-mapped counter. User-Level Policies: With the victimList abstraction, the user-level infoReplace library can frequently poll the OS and when new pages are near eviction, obtain the list of those pages; if any of these pages should not be evicted according to the target policy, infoReplace accesses them to move them to the active list.... In PAGE 7: ... Following these basic steps, we have implemented FIFO, LRU, MRU, and LFU on top of the Linux 2Q-based re- placement algorithm. The bottom half of Table2 shows the amount of C code needed to implement infoReplace. Although more than one thousand statements are required, most of the code is straightforward, with the bulk for simu- lation of different replacement policies.... ..."

"... In PAGE 6: ... Once this counter is approximately equal to N, the process performs the more expensive call to get the state of the last N pages on the inactive queue. As shown in the top half of Table2 , the victimList abstraction can be implemented in only 109 C statements; in fact, more than half of the code is needed to setup the memory-mapped counter. User-Level Policies: With the victimList abstraction, the user-level infoReplace library can frequently poll the OS and when new pages are near eviction, obtain the list of those pages; if any of these pages should not be evicted according to the target policy, infoReplace accesses them to move them to the active list.... In PAGE 7: ... Following these basic steps, we have implemented FIFO, LRU, MRU, and LFU on top of the Linux 2Q-based re- placement algorithm. The bottom half of Table2 shows the amount of C code needed to implement infoReplace. Although more than one thousand statements are required, most of the code is straightforward, with the bulk for simu- lation of di erent replacement policies.... ..."

"... In PAGE 6: ... Once this counter is approximately equal to N, the process performs the more expensive call to get the state of the last N pages on the inactive queue. As shown in the top half of Table2 , the victimList abstraction can be implemented in only 109 C statements; in fact, more than half of the code is needed to setup the memory-mapped counter. User-Level Policies: With the victimList abstraction, the user-level infoReplace library can frequently poll the OS and when new pages are near eviction, obtain the list of those pages; if any of these pages should not be evicted according to the target policy, infoReplace accesses them to move them to the active list.... In PAGE 7: ... Following these basic steps, we have implemented FIFO, LRU, MRU, and LFU on top of the Linux 2Q-based re- placement algorithm. The bottom half of Table2 shows the amount of C code needed to implement infoReplace. Although more than one thousand statements are required, most of the code is straightforward, with the bulk for simu- lation of different replacement policies.... ..."