Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
configurations until one works. The fact that such a large number of tests are required<br />
for such a simple problem (only two tuning and two uncertainty parameters) indicates<br />
that this approach is fairly ill-suited for this application. The barrier steepest-descent<br />
real-time optimization performs much better than SA, requiring an average of 46<br />
hardware tests to tune successfully. The maximum value of tests is still pretty high at<br />
231, however, and is expected to increase <strong>with</strong> the number of tuning and uncertainty<br />
parameters in the problem.<br />
The isoperformance tuning algorithm, a hybrid of model and hardware tuning<br />
techniques, stands out as the clear winner. Like the hardware-only methods it suc-<br />
cessfully tunes each and every one of the hardware realizations, but does so <strong>with</strong> only<br />
a few hardware tests. The maximum number of tests required by the isoperformance<br />
tuning method is 4 and the average across the sample space is 2.5. These statistics<br />
are a factor of ten less than BSD and a hundred times less than SA. It is true that<br />
the iso-lines in this problem are particularly well-suited for this tuning methodology<br />
as discussed previously, but the additional example considered shows that even if<br />
the isoperformance lines are less favorable the method performs very well. As the<br />
complexity of the model and tuning problem increases, it is expected that the model-<br />
only methods will fail more often and that the hardware tuning methods will require<br />
even more tests to find a successful tuning configuration. Therefore, although the<br />
model-only methods are attractive for their simplicity and low-cost, and the hard-<br />
ware methods are attractive for their success rate, neither method is really a practical<br />
solution as there is always the chance of failure or prohibitive expense. In contrast,<br />
the isoperformance tuning method is able to consistently provide successful tuning<br />
solutions <strong>with</strong> only a small number of hardware tests.<br />
4.4 Summary<br />
In this chapter, the concept of dynamic tuning is defined as the process of adjusting<br />
hardware to bring the system performance <strong>with</strong>in requirements. The tuning process is<br />
formalized as an optimization, and guidelines for choosing appropriate tuning param-<br />
148