"... In PAGE 19: ... In addition, we show the average performance degradation over optimal selection. The data in Table4 shows the average difference between our GC switching system and the best performing system at each heap size (column 2) and between our system and the worst performing system at each heap size (col- umn 3). In parentheses, we show the average absolute difference in milliseconds; for SPECjbb the value in parenthesis is the dif- ference in inverse throughput.... ..."

"... In PAGE 7: ... To alleviate this problem, we adapt the bucket size for the Golomb run-length coder from layer to layer. As shown in Table3 , block GC3 results in about 10 to 15 % lower compression efficiency than block C4 over dif- ferent process layers of layouts, assuming decoder buffer size of 1.7 kB.... In PAGE 7: ...7 kB. The test images in Table3 are 1024 H110031024 five-bit grayscale rasterized, flattened layouts, ex- amples of which are shown in Figs. 4H20849aH20850 and 4H20849bH20850.... ..."

"... In PAGE 8: ... In addition, we show the average performance degradation over optimal selection. The data in Table4 shows the average dif- ference between our GC switching system and the best perform- ing system at each heap size (column 2) and between our system and the worst performing system at each heap size (column 3). In... ..."

"... In PAGE 8: ... In addition, we show the average performance degradation over optimal selection. The data in Table4 shows the average dif- ference between our GC switching system and the best perform- ing system at each heap size (column 2) and between our system and the worst performing system at each heap size (column 3). In... ..."

"... In PAGE 8: ...esult using the K-means clustering algorithm. It can be seen that only four of them can be correctly segmented. However, after removing the center part which represents the font Comic Sans MS (figure 6(d)) and using the same parameters to extract the features, the remaining four parts can be correctly segmented using the same K-means clustering algorithm (figure 6(f)), which implicitly means the incorrect segmentation was caused by the feature similarity of different fonts. Comparing this result with the results shown in Table3 , we can draw the conclusion that the BPNN classifier did improve the classification accuracy. (a) (b) (c) (d) (e) (f) Figure 6.... ..."

"... In PAGE 6: ... 4.2 Effect of garbage collector accuracy In which region of memory is accuracy most important for each of our three hardware platforms? Table8 presents data that com- pares different levels of accuracy to conservative garbage collec- tion. Unlike Tables 6 and 7 which presented data only for pro- grams that use explicit deallocation, Table 8, presents data for our entire benchmark suite.... In PAGE 6: ...2 Effect of garbage collector accuracy In which region of memory is accuracy most important for each of our three hardware platforms? Table 8 presents data that com- pares different levels of accuracy to conservative garbage collec- tion. Unlike Tables 6 and 7 which presented data only for pro- grams that use explicit deallocation, Table8 , presents data for our entire benchmark suite. The Average gcA0gcD7CWCV gc column presents the average of the excess bytes that gc retains over gcD7CWCV as a fraction of the bytes retained by gc.... In PAGE 9: ...21 Table 9: Effect of changing heap layout for SPARC Table 9 presents the results of this experiment. The Original column of Table 9 is identical to the Average gcA0gcD7CWCV gc column in Table8 for the SPARC except that it presents numbers only for the 5 garbage collection runs. The Lower heap layout column is computed similarly to the Original column except that it uses a lower starting address for the heap.... ..."

"... In PAGE 6: ... From this figure we see that even though ijpeg allocates heavily, the number of bytes that are live at any given point is very small. Thus even though gcD7CWCV appears to do significantly worse than free on ijpeg according to Table6 , in absolute terms, the number of excess bytes retained by gcD7CWCV is rather small. From Table 6 we see that the performance of gcD7CWCV also varies with hardware platform.... In PAGE 6: ... Thus even though gcD7CWCV appears to do significantly worse than free on ijpeg according to Table 6, in absolute terms, the number of excess bytes retained by gcD7CWCV is rather small. From Table6 we see that the performance of gcD7CWCV also varies with hardware platform. This is because gcD7CWCV is not fully accu- rate for C programs (Section 3).... In PAGE 6: ... Even for our suite of benchmarks, we detected four programs where accurate garbage collection reclaimed more objects than explicit deallocation: bc, ft, yacr2, ijpeg. Only one of these leaks (ft) shows up in Table6 , since Table 6 presents the average differ- ence between explicit deallocation and garbage collection. In our experiments, conservative garbage collection reclaimed more ob-... ..."