"... In PAGE 5: ... We validate the traffic model in two aspects: 1) we generate sample data from the traffic model, and compare the statistics between the generated video traf- fic and the video trace; 2) we inject the video trace data and the generated video data into the network and compare the queueing behavior of the network transmission buffer. The comparison of average video statistics, including the mean and the standard deviation to average ratio (SDA), for the trace data and generated video data samples are shown in Table1 . It can be seen from the table that the two set of data share very similar statistics.... ..."

"... In PAGE 6: ... Table 2 shows the configuration of the baseline model. The indirect branch prediction accuracy of the BTB in the baseline model is compared with the base trace cache model whose simulator parameters are shown in Table3 . The base trace cache... In PAGE 9: ... 5.2 Trace Cache with Update Policy In the base trace cache model in Table3 , we use the keep-longest policy for writing to the trace cache since it is shown to be best policy by Patel [15] for maximizing fetch bandwidth. However, if the always overwrite policy, which always overwrites the trace cache lines, was used, then updating the indirect branch target addresses would be handled automatically.... ..."

"... In PAGE 16: ... We also reduced the number of states in the Star Wars trace and the Ingalls trace by half. Table14 displays the size of the model for every test case trace. In most cases the model is approx- imately 40% the trace size.... ..."

"... In PAGE 11: ...______________________________________________________________________________________ Borella, Source Models of Network Game Traffic A summary of our models of traces Quake-1 and Quake-2 appear in Table4 and Table 5. In order to save space, we do not show the distributions and Q-Q plots for the Quake-2 traffic.... In PAGE 11: ... We performed this tail removal for the interarrivals of several hosts in trace Quake-2. In the tail column of Table4 and Table 5, we use the format x/y to denote the number of items removed from the upper tail, and the quality of the tail estimation after these items, if any, have been removed. The quality of the tail estimation metric takes on one of three values: - if the model underestimated the tail, + if the model overestimated the tail, and 0 if the model did a reasonable job of tail estimation.... ..."

"... In PAGE 11: ... We performed this tail removal for the interarrivals of several hosts in trace Quake-2. In the tail column of Table 4 and Table5 , we use the format x/y to denote the number of items removed from the upper tail, and the quality of the tail estimation after these items, if any, have been removed. The quality of the tail estimation metric takes on one of three values: - if the model underestimated the tail, + if the model overestimated the tail, and 0 if the model did a reasonable job of tail estimation.... ..."

"... In PAGE 7: ... The solidlines denote the results obtained from the synthetic trace model and the points denote the data obtained from the real traces. The parameters used by the synthetic model is listed in Table2 . Figure 4 showed that the synthetic model accurately captured the behavior of the real trace for various block sizes and Eqn (9) gives reasonably good estimation of the spatial locality property.... In PAGE 7: ... The y-axis is the accumulated miss ratio which is the total number of misses in a 8KByte cache over the total number of line references accessed. In Table2 , all of the traces exhibit a certain degree of spatial locality. As a result, increasing the block size almost always improves the cache miss ratio, provided that the cache size is large enough to Trace (1 ? p1) p2 p3 p4 m El LISP 0:290 0:775 0:776 0:786 536 4.... ..."