|
|
|
|
|
|
|
Stability. Data should represent a conservative estimate of miss rate. Small programs especially may perform significantly better than DTMR. |
|
|
|
|
|
|
|
|
Data Note 2: Empirical Data for Different Application Environments and Machine Traces (Figures 5.11 and 5.12). |
|
|
|
|
|
|
|
|
Extensive data for 370, VAX, 360/91, and 6400 traces are given in Smith [260]. Data for 32016-UNIX traces are taken from Alexander [11]. Data for 8200 traces are from Agarwal [4]. All these data are obtained by cache simulation on program traces. |
|
|
|
|
|
|
|
|
Use of the data. This data is presented as a validation for the DTMR. The DTMR data shown are for two-way associativity and 16-byte lines. The other data were selected to be as comparable as the sources allow. |
|
|
|
|
|
|
|
|
Data Note 3: Effect of Set Associativity (Figures 5.13 and 5.14). |
|
|
|
|
|
|
|
|
Data for different set associativities are taken from Hill [129]. Relative miss rates are computed directly from miss rates and the curves shown here have been smoothed. |
|
|
|
|
|
|
|
|
Use of the data. This data has been prepared for use with DTMR data in Figures 5.10, 5.28, and 5.29. |
|
|
|
|
|
|
|
|
Reliability. The data presented seems consistent with other reported results. It should be useful for conservative design estimates. |
|
|
|
|
|
|
|
|
Stability. Individual programs may vary as to the effect of set associativity. Variance is expected to be higher with direct-mapped and low degrees of association. |
|
|
|
|
|
|
|
|
Data Note 4: Effect of System Activities (Figures 5.21 and 5.22). |
|
|
|
|
|
|
|
|
Materials on effect of system activities are taken from Agarwal [4]. Since data on few program traces have been reported, the relative miss rates should be treated cautiously. |
|
|
|
|
|
|
|
|
Use of the data. Data (Figure 5.22) should be used as an adjustment to basic DTMRs. |
|
|
|
|
|
|
|
|
Reliability. Data presented is for a particular operating system running a relatively small number of benchmarks. While the data is certainly indicative of expected performance, the actual values may vary with environment. |
|
|
|
|
|
|
|
|
Stability. One would expect a significant variation with operating systems. There are only limited data sources available for this data. |
|
|
|
|
|
|
|
|
Data Note 5: Cold-Start Caches and Effect of Multiprogramming (Figure 5.26). |
|
|
|
|
|
|
|
|
These curves are computed from DTMR using the Markovian cache model by Haikala [116]. The warm-start cache data include the effect of system activities. Due to the accuracy of the cache model used and the estimation of the input data, miss rates shown in these figures are pessimistic. (See Appendix C for a more complete discussion.) |
|
|
|
|
|
|
|
|
Use of the data. This data should be used with care. Cold-start (transaction) data should not be adjusted for I/O, as these effects are included in the transaction model. |
|
|
|
|
|