Tietojenkäsittelytieteen laitos

Perf. Eval., Homework 6 (due 24.4.2002)

1. Consider the multiple class memory constrained model and its iterative solution (slides 13-16 in Lecture 11).
   1. What would be better initial population vector than the one used (slide 14)? When is the one used not so good? When is it good?
   2. In each round, when computing solution for any given class (e.g., class 2), would it be better to use class average population for earlier classes (e.g., class 1) that same round, or from the previous round? Why?
   3. What would be good criteria for convergence? Why?
   4. In each step, would exact solution method be better than approximate MVA? Why?
   5. In step 3 (slide 15), why do we solve the model for class c populations 1 through J_c, even though we know that n_c might be much less than J_c?
2. Consider a two-class priority model which uses a shadow server for lower class CPU (see slide 21 in Lecture 11). How would compute overall utilization and queuelength for CPU in this system?
   
      
3. [2 htp] Calibrate the example model given in Lecture 12 (see Table 9.1 [Men 94], rights updated 21.4.02 ) further so that utilizations and response time would be closer match to measurements, as compared to the original PMVA model output given in fig.9.1.out. (Original PMVA model data is given in fig.9.1.pmva)
   1. Would a ghost server help?
   2. How about max mpl? Original data was for system with max mpl, but that did not mean that the mpl would be at its maximum all the time.
   3. What about adding a new class to count for "other workload"?
   4. How about some combination of those above?

   What is your best calibrated model?

    

4. [1 htp] Fill in the course feedback form for this course. Place special emphasis on question 14 ("How the course could be developed further?"), and bring a copy of your answer to that question to the practice session. Consider especially the text book problem for this course. Thank you.