Figure 5 shows the results of the detection by the DEAR scheme for the cscope and cpp applications. In each graph, the x-axis is the virtual time and the y-axis is the logical block numbers of those referenced at the given time. The detection results are given at the top of the graph assuming a detection period of 500 references. For cscope, the DEAR scheme initially detects a sequential reference pattern but changes its detection to a looping reference pattern after the sequentially referenced blocks are re-accessed. This results from cscope always reading the file cscope.out sequentially whenever it receives a query about the C source. For cpp, the DEAR scheme detects a probabilistic reference pattern throughout the execution since as we can see from the graph, some blocks are more frequently accessed than others. This reference pattern results from the characteristic of cpp that header files are more frequently referenced than C files.
Figure 5: Block reference patterns and detection results for cscope and cpp.
Figure 6 shows the detection results of the other applications. Although the result shows that the DEAR scheme performs reasonably well for the other applications, it also reveals the limitation of the current DEAR scheme, notably for the sort and postgres2 applications. They have either parallel or nested reference streams, which indicates a need for the proposed DEAR scheme to address more general reference patterns with arbitrary control structures such as parallel, sequence, and nested.
Figure 6: Block reference patterns and detection
results for the other applications.
