|
|
|
|
|
|
|
Assuming the penalty due to contention is one cycle, the additional CPI is 0.118 ´ 1.47 = 0.173. |
|
|
|
|
|
|
|
|
Based on the preceding penalty, we need to compare the performance due to contention versus the performance lost due to smaller split cache and higher miss rate. |
|
|
|
|
|
|
|
|
Write-Through versus Copyback |
|
|
|
|
|
|
|
|
A choice must be made between WTNWA and CBWA for the write policy. The responses with respect to the different requests of these schemes is briefly described in Table 10.15. |
|
|
|
|
|
|
|
|
The characteristics of memory traffic generated by the two schemes are quite different. With the CBWA scheme, all the memory traffic generated is lines, whereas the WTNWA scheme generates line traffic during a read miss and a word write to memory during every write. |
|
|
|
|
|
|
|
|
To provide a more quantitative view of memory traffic difference between the implementation choices, the following simple model has been developed. We present two additional models for a Write Assembly Cache (WAC) and a WTNWA first-level cache with a CBWA secondary cache. The factors important in this model are: |
|
|
|
|
|
|
|
|
Tword is the word access time (Taccess). |
|
|
|
|
|
|
|
|
% of total lines that are dirty (W). |
|
|
|
|
|
|
|
|
Reference traffic patterns. |
|
|
|
|
|
|
|
|
To calculate Tline access, see equation 6.3. |
|
|
|
|
|
