Mehdi Amirijoo1 Dynamic power management n Introduction n Implementation, levels of operation n Modeling n Power and performance issues regarding power management n Policies n Conclusions

Mehdi Amirijoo2 Introduction n To provide the requested services and performance levels with a minimum number of active components or a minimum load on such components. n Assume non-uniform workload. n Assume predictability of workload. n Low overhead of caused by power manager; performance and power.

Mehdi Amirijoo3 Introduction n The power manager (PM) implements a control procedure based on observations and assumptions about the workload. n The control procedure is called a policy. n Oracle power manager

Mehdi Amirijoo4 Implementation n Hardware –Frequency reduction –Supply voltage –Power shutdown n Software –Mostly used –Most flexible n Operative system power manager (OSPM) –Microsoft’s OnNow –ACPI

Mehdi Amirijoo5 Modeling n View the system as a set of interacting power- manageable components (PMCs), controlled by the power manager (PM).

Mehdi Amirijoo6 Modeling n Independent PMCs. n Model PMCs as FSMs; PSMs n Transition between states have a cost. n The cost is associated with delay, performance and power loss. n Service providers and service requesters.

Mehdi Amirijoo7 Modeling n Ex. StrongArm SA-1100 processor (Intel)

Mehdi Amirijoo8 Power and performance issues.. n Power management degrades performance.

Mehdi Amirijoo9 Power and performance issues.. n Break-even time T be - minimum length of an idle period to save power. Move to sleep state if T idle > T be T 0 : Transition delay (shutdown and wakeup) E 0 : Transition energy P s, P w : Power in sleeping and working states

Mehdi Amirijoo10 Policies n Different categories: –Predictive –Adaptive –Stochastic n Application dependent n Statistical properties n Resource requirements

Mehdi Amirijoo11 Policies - Predictive n Fixed time-out: –Static –Assume that if a device is idle for , it will remain idle for at least T be. –If device idle for , change state to sleep. –Time-out  is computed and set off-line. –Very simple to implement. Requires a timer. –Power is wasted in waiting for time-out. –Can cause many under-predictions. –Adaptive version where  is adjusted online.

Mehdi Amirijoo12 Policies - Predictive n Predictive shut-down [Golding 1996]: –Take decisions based observations of past idle and busy times. Take decision as soon as an idle time starts. –The equation f yields a predicted idle time T pred –Shut down if –Sample data and fit data to a non-linear regression equation f (off-line). –Computation and memory requirements.

Mehdi Amirijoo13 Policies - Predictive n Predictive shut-down [Srivastava 1996] –Take decision based on observing the last busy time. Take decision as soon as an idle time starts. –If change state. –Suitable for devices where short busy periods are followed by long idle periods. L-shape plot diagrams (idle period vs busy periods). n FSMs similar to multibit branch prediction in processors. n Predictive wake-up

Mehdi Amirijoo14 Policies - Adaptive n Static policies are ineffective when the workload is nonstationary or not known in advance. n Time-out revisited: 1. Adapt the time-out . 2. Keep a pool of time-outs and choose the one that will perform best in this context. 3. As above, but assign a weight to each time-out according to how well it will perform relative to an optimum strategy for the last requests.

Mehdi Amirijoo15 Policies - Adaptive n Low pass filter [Wu1997] :

Mehdi Amirijoo16 Policies - Stochastic n Predictive and adaptive policies lack some properties: –They are based on a two state system model. –Parameter tuning can be hard. n Stochastic policies provide a more general and optimal strategies. n Modeled by Markov chains, Pareto.

Mehdi Amirijoo17 Policies - Stochastic (Markov) n Stationary (or WSS). Statistical properties do not depend on the time shift, k. n A set of states. Probability associated with the transitions. n The solution of the LP produces stationary, randomized (nondeterministic) policy. n Finding the minimum power policy that meets a given performance constraint can be cast as a linear program (LP, solved in polynomial time).

Mehdi Amirijoo18 Policies - Stochastic (Markov)

Mehdi Amirijoo19 Policies - Stochastic (Markov) n The policy computed by LP is globally optimum [Puterman 1994]. n However, requires knowledge of the system and its workload statistics in advance. n An adaptive extension [Chung 1999]: –Policy precharacterization (PC) –Parameter learning (PL) –Policy interpolation (PI)

Mehdi Amirijoo20 Policies - Stochastic (Markov) n An adaptive…(cont.) –Two-parameters Markov. Parameters “describe” the current workload. –PC constructs a 2-dim table, addressed by the values of the two parameters. –The table elements contain the optimal policy, identified by the pair. –Parameter learning is performed during operation. –PI is performed to find a policy as a combination of the nearby policies given by the table and the parameters.

Mehdi Amirijoo21 Conclusions n The policies are application dependent and have to be adopted to devices. n Policies based on stochastic control and specially Markov allows a flexible and general design, where all requirements can be incorporated. n Current models are based on observing requests arrivals. A trend in power management is to include higher-level information, particularly software-based information from compilers and OSs.