Stochastic modeling of a power-managed system-construction and optimization
read more
Citations
Stochastic model checking
Thermal Modeling, Analysis, and Management in VLSI Circuits: Principles and Methods
Evaluating the effectiveness of model-based power characterization
Autonomic power and performance management for computing systems
Logic soft errors in sub-65nm technologies design and CAD challenges
References
Stochastic models in operations research
Low Power Digital CMOS Design
Low Power Design Methodologies
Related Papers (5)
Frequently Asked Questions (13)
Q2. What is the purpose of the SQ model?
The SQ model can be used to model the commonly used priority queue in an operating system where two different priorities are assigned to tasks and highpriority tasks, when they come, are inserted into the front of the queue.
Q3. Why does the PM always buffer the incoming service requests?
Especially in a power-managed system, the PM always buffers the incoming service requests, that is, to achieve the best power-delay trade-off.
Q4. What is the simplest way to model a power-managed system?
The Power-Managed System (SYS) can be modeled as a continuous-time Markov process which is the composition of themodels of the SP, the SR and the SQ.
Q5. what is the value of jiji rrrr?
The value ofji rr , σ (the transition rate from state ri to state rj) canbe calculated as:jiji rrrr ,, τσ = , ri≠rj; ∑ ≠ −=ijjiii rrrrrr ,, σσ (3.4)The request rates λl(r) and λh(r) are associated with state r∈R.
Q6. What is the effect of the new model on the delay of high priority requests?
In addition, the original model always over-estimates the delay of highpriority requests, i.e., the simulated delay of high-priority requests is always smaller than the pre-set constraint.
Q7. What are the two different scenarios for the Optimal Policies for both models?
Optimal policies for both models are calculated under following two different scenarios: 1. Since the original model cannot distinguish between high-priority requests and low priority requests, to make sure that the delay of high priority requests meets the constraint C, it has to apply the constraint C on all requests.
Q8. What is the transition rate for a transition between active and inactive states?
It can be obtained as:),()(),( rsara SQ II SP II SQSP GGG ⊕=− (4.4)Notice that, after the operation, the parameter s in GSQ(s, r) has been removed by substituting the real state of the SP.),( raII SQSP −G is calculated directly by the ⊕ operation because transition between inactive SP states is not correlated with the transition of SQ state.
Q9. Why is the transition rate correlated between the SP and the SQ?
Because whenever the SP makes a transition from a busy state to an idle state (finishes the service for a request), the SQ must make a transition which decreases the number of requests in SQ by 1.
Q10. How long does the SP wait for a request?
2. Timeout policy (T=20): turn on the SP whenever a request comes and turn off the SP whenever the SP has been idle for 20 seconds and there is no request in the queue.
Q11. What is the switching energy ene(si, sj) matrix?
The switching energy ene(si, sj) matrix is: ∞ ∞∞∞=093066.004.4210000),(JJJJJJ ssene jiEntry of ∞ means that the SP cannot switch between the corresponding states.
Q12. What is the way to improve the performance of the dynamic power management method?
The authors also showed that the dynamic power management method out-perform the heuristic approaches in terms of better and more flexible power-delay trade-off.
Q13. What is the way to handle the workload?
The new model can adjust the optimal policy when workload characteristics change, while the greedy and timeout methods are not adaptive to the workload.