"... In PAGE 8: ... The first set of experiments compares the result of us- ing the original CME framework and that of our extended framework. Table2 shows the number of misses and hits produced by the original framework and our framework, re- spectively. For all except the last benchmark, we assume that arrays are aligned on cache line boundaries for sim- plicity.... In PAGE 9: ... Comparison: Orig. CME / Our Frame- work / Trace-Driven for 4KB Data Cache two benchmarks shown in Table2 . Hence, for these bench- marks, the very fact that our framework can consider them is an advantage that the original framework does not pos- sess.... In PAGE 9: ... For the sake of comparison on equal ground, we created a synthetic benchmark with a loop structure that is analyz- able by Coyote. This is the last row in Table2 . Here, we see that our framework produces tighter estimates than the original Coyote framework even for programs analyzable by Coyote.... ..."

"... In PAGE 7: ... 5.2 Significance Tests In Table2 we show the values for significance tests be- tween the scores of the different retrieval models, once again at a window size of 20 shots. Significance testing was done using the Wilcoxon matched-pairs signed-ranks test, at the 0.... In PAGE 7: ... Firstly, we find that our retrieval framework produces consistently higher MAP scores for topics than it does for concepts. As shown in Table2 , at window size 20 the highest score for topics is almost 6 times larger than for concepts, with MAP scores of 0.0790 and 0.... In PAGE 8: ... For the concepts a mixed message emerges: the empirical model has a slight edge over the power law model in one case (model 2 vs model 4) and vice versa in the other case (model 3 vs model 5). In sum, the significance tests in Table2 show that the retrieval models based on power law approximations do not produce significantly lower MAP scores than those based on empirical data.4 Hence, it seems safe to recommend the use of power law based models for incorporating redundancy within retrieval models.... ..."

"... In PAGE 7: ... 5.2 Significance Tests In Table2 we show the values for significance test between the scores of the different retrieval models, once again at a window size of 20 shots. Significance testing was done using the Wilcoxon matched-pairs signed-ranks test, at the 0.... In PAGE 7: ... Firstly, we find that our retrieval framework produces consistently higher MAP scores for topics than it does for concepts. As shown in Table2 , at window size 20 the highest score for topics is almost 6 times larger than for concepts, with MAP scores of 0.0790 and 0.... In PAGE 8: ... For the concepts a mixed message emerges: the empirical model has a slight edge over the power law model in one case (model 2 vs model 4) and vice versa in the other case (model 3 vs model 5). In sum, the significance tests in Table2 show that the retrieval models based on power law approximations do not produce significantly lower MAP scores than those based on empirical data.4 Hence, it seems safe to recommend the use of power law based models for incorporating redundancy within retrieval models.... ..."

"... In PAGE 35: ....2.3 Heterogeneous access networks There have been a wide variety of wireless access network technologies developed during the past decade. As described in Table1 , wireless data access can be provided over heterogeneous wireless networks [17]. Some of the wireless network standards have aimed at specific market regions such as North-America, Europe and Asia, especially in cellular networks.... In PAGE 37: ... However, the practical performance of these cellular data systems is usually much lower than the nominal rates, even down to ten times lower than the nominal rate [22]. As shown in Table1 , wireless local and personal area networking standards include IEEE 802.... In PAGE 39: ...ominated by 802.11 standards. In addition, the lack of end-to-end QoS support in the public Internet so far may have lowered the interest for providing QoS support in the local access network. As illustrated in Table1 , there are a wide variety of wireless networks, and many local and personal area standards have emerged. However, whatever the used WLAN system is, it is providing service only locally .... ..."

"... In PAGE 24: ... The data set comprised 26 756 measurement records with 6 traits. Table9 gives the number of levels for each e ect leading to 233 796 normal equations. The columns headed by \trait quot; represent the model matrix (cf.... ..."

"... In PAGE 7: ... The LRU algorithm out-performed the true transition probability in every application, and is much easier to implement. Table4 shows the reference fre- quency for each predictor address in a Markov prefetcher with four prediction addresses. The columns represent the relative age of the predictor reference and predictor 0 is always the most recently used address predictor.... ..."

"... In PAGE 9: ... from one article to another, they classify it as deleted from the source article. From Table1 , we note that the precision is low, by search standards. Our problem, however, is a prediction problem, not a retrieval problem, and thus it is intrinsically different.... ..."

"... In PAGE 9: ... from one article to another, they classify it as deleted from the source article. From Table1 , we note that the precision is low, by search standards. Our problem, however, is a prediction problem, not a retrieval problem, and thus it is intrinsically different.... ..."

"... In PAGE 9: ... from one article to another, they classify it as deleted from the source article. From Table1 , we note that the precision is low, by search standards. Our problem, however, is a prediction problem, not a retrieval problem, and thus it is intrinsically different.... ..."

"... In PAGE 6: ... The LRU algorithm out-performed the true transition probability in every application, and is much easier to implement. Table4... ..."