Citation Context

...it until anothers.'~rrival before revising the decision, possibly missing largesidle times, such as a lunch break. CTMDP model was fur-sther generalized with semi-Markov decision process models(SMDP) =-=[12]-=-. SMDP model can treat a general distributionsoccuring at the same time with an exponential distribution.s]In [12] the hard disk transition to and from the sleep stateswas modeled using uniform distri...

Citation Context

...any of the cores that are of interest in NOC design already have multiple power and performance states which can be exploited. Dynamic power management (DPM) and DVS algorithms presented to date [10]–=-=[13]-=-, [15]–[25] have been designed with portable systems in mind, not NOCs with very fast response times. In addition, they are targeted devices such as hard disks or wireless LAN that in our case is equi...

Citation Context

...ervice processes of requests and power state transition processes are random. If these processes are stationary, the DPM problem can be formulated as a Markov decision process (MDP) problem [1], [8], =-=[13]-=- for which a stationary optimal DPM policy exists and can be computed. Typically, the objective is to optimize performance (e.g., waiting time . Z. Ren is with the Signal Electronics and Embedded Syst...

Citation Context

...-exponential transitions occurring at the same time when the card transitions from doze mode into o state. Thus we could not apply policy optimization algorithms based on exponential models, such as =-=[9, 10, 13]-=-. Large errors result if exponential distribution is used for all transitions, as was shown in [11]. Another approach to handle non-exponential transitions is to use adaptive method, such as in [12]. ...

Citation Context

...sion processes (CTMDP) was proposed in [4]. The policyschange under this model is asynchronous and thus more suitable forsimplementation as part of a real-time operating system environment.sReference =-=[5]-=- also improved on the modeling technique of [3] bysusing time-indexed semi-Markovian decision processes.sAlthough the abovementioned DPM techniques may successfullysreduce the system power consumption...

Citation Context

...d on the typical access speeds of DRAM described in Section II, the time to read or write a checkpoint of size 5 kB is assumed to be 0.4 ms. We choose a power consumption value of 400 mW for the DRAM =-=[47]-=-. Hence, the energy consumption for saving or retrieving a checkpoint is 160 µJ. In addition, based on data provided in the literature in [37]–[40], it was assumed that a single DVS transition takes 1...

Citation Context

...then time indexing is needed, as shown for idle and sleep with empty queue states. In that case, the time-indexed semi-Markov decision processes (TISMDP) model is used, as described in more detail in =-=[18]-=-. Each core in the reliability network is modeled with a power and reliability state machine (PRSM) as shown for the ARM core in Fig. 4. PSRM is a state diagram relating service levels to the allowabl...

Citation Context

...nch [17] are selected and run on datasets provided by [18]. Moreover, we have also used a set of programs from Mediabench [40] benchmark suite and run them on SPARC cores. Pareto distribution is used =-=[19]-=- to introduce idle intervals between the MiBench tasks. A timeout-based dynamic power management policy is applied in order to determine the active and low power states which each core experiences dur...