FLEA - FIT-Aware Heuristic for Application Allocation in Many-Cores based on Q-Learning

Reinforcement Learning for Thermal and Reliability Management in Manycore Systems

“…This paper extends the work presented in SBESC'23 [28]. The SBESC paper introduced a Q-Learning-based heuristic for application allocation in manycore systems.…”

“…If stability persists over this duration, rewards are allocated based on FIT variation, leading to the computation of Q new (p, s), where s represents the PBS and p denotes the TPC executing in the PE (Equation 5). The diagram is divided into three main processes; the blue path presents the task-to-core selection process; the orange path is the Q-table update process; and the red presents the temperature and FIT estimation process [28].…”

“…Subsequent nested loops (lines [16][17] analyze each PE within the window, adding temperature data (line 18) and quantifying available PEs (lines [19][20][21]. If the analyzed window meets the spatial requirements and has a lower temperature than any previously identified candidates, it becomes the new candidate cluster (lines [24][25][26][27][28][29][30][31].…”

“…Fig.5: Training flow diagram. The diagram is divided into three main processes; the blue path presents the task-to-core selection process; the orange path is the Q-table update process; and the red presents the temperature and FIT estimation process[28].…”

“…the heat Average power consumption of tasks (τ m,n | m ∈[1,7], n ∈ [1, N m ]). Tasks are classified into three categories: p 0 , p 1 , and p 2[28].influence of a task executing at north, south, east, or west is the same. One example of equivalent arrangements is {p 0 , λ, λ, λ}, {λ, p 0 , λ, λ}, {λ, λ, p 0 , λ}, and {λ, λ, λ, p 0 }, (where λ means a idle PE) therefore, in any of those situations the PBS will be considered the same.3.…”