< previous page page_377 next page >

Page 377
The selection of r» 0.5 is an empirical initial guessa conventional ''rule of thumb." It implicitly assumes that the processor cost and memory cost are balanced. If, for example, the memory cost is much greater than the processor cost and increasing its bandwidth to achieve r» 0.5 significantly contributed to this cost, it would be better to let r rise, even to the point where r» 1. At this point the Flores model (open-queue MB/D/1) is no longer a useful model and other modeling techniques must be used to determine bandwidth.
EXAMPLE 6.2
Suppose we wish to design a memory system for a processor with peak performance of 50 MIPS and one instruction decoded per cycle.
1. Assume the memory module has
d87111c01013bcda00bb8640fdff6754.gif
Ta = 200 nsec and Tc = 100 nsec,
d87111c01013bcda00bb8640fdff6754.gif
and assume a total of 1.5 references per instruction = 1.5 references per processor cycle (1.5 references each 20 ns).
2. Compute MAPS = 1.5 ´ 50 = 75 MAPS (ignoring I/O traffic).
3. Find r.
0377-01.gif
d87111c01013bcda00bb8640fdff6754.gif
If we use m = 16, then
0377-02.gif
d87111c01013bcda00bb8640fdff6754.gif
Or, for this case, we can simplify to:
0377-03.gif
0377-04.gif, and total mean queue size Q0-t = 16 × .18 = 3.

 
< previous page page_377 next page >