|
|
|
|
|
|
|
Data Note 1. The data in Tables 3.2, 3.3, 3.4, and 3.5 are primarily from Huck and Flynn [136] and Rossmann [248, 247], although other published sources have been considered. The Rossmann data are generally based on extensive studies of IBM mainframes. This seems to be one of the few architectures that provide a reasonable body of available data for the commercial environment. The systems data were taken from studies at DEC Western Research Lab [245] and from OLTP data from HP labs. The studies target L/S and (in the DEC case) R+M architectures; our R/M data is interpolated. |
|
|
|
|
|
|
|
|
Reliability. There is a significant variability in the distribution of operations by application. The data presented assume an optimizing compiler (when such data are available). These data are an attempt to create a conservative representation of the operation profiles by emphasizing the expected occurrence of more difficult operations, such as floating point and decimal. The data provided for the L/S architecture in the commercial environment is completely syntheticsimply mapped-over data from R/M and R+M studies. The L/S data also make assumptions concerning the availability of certain instructions to interpret operations that occur in R/M architectures. The overall reliability for the scientific environment and for the R/M and R+M commercial environment should be reasonably good if somewhat conservative. The reliability of the L/S commercial data is low. |
|
|
|
|
|
|
|
|
Variance. The variance in Tables 3.2 and 3.3 is expected to be relatively low, while the variance in Table 3.4 within each architecture is much higher. The occurrence of floating point, e.g., in scientific applications, has been variously reported between 4 and 20 percent. Similarly, the occurrence of branch can be much higher than that indicated here. There is less data available for the classification profile of commercial and systems applications (Tables 3.5 and 3.6). Similarly, the variance across classes is expected to be relatively high. |
|
|
|
|
|
|
|
|
Data Note 2. The data in Tables 3.10 and 3.11 are taken from Rossmann [248, 247] and Huck and Flynn [136] (for Table 3.10) and Peuto and Shustek [231] (for Table 3.11), based upon monitor and trace studies done largely for the System/370 architecture. |
|
|
|
|
|
|
|
|
Reliability. The data presented in these tables should be interpreted as typical data. |
|
|
|
|
|
|
|
|
Stability. Significant differences may arise from program to program, especially in areas such as branch distances (Table 3.11). |
|
|
|
|
|
|
|
|
Data Note 3. The data in Tables 3.14, 3.15, and 3.16 are derived from Huck and Flynn [136] and Rossmann [248, 247]. |
|
|
|
|
|
|
|
|
Reliability. The data are generally consistent with other reported results in the area. |
|
|
|
|
|
|
|
|
Variance. As the categorization is refined in these tables, the variance for individual classes is expected to increase (compared to, e.g., Tables 3.4 and 3.5). |
|
|
|
|
|