A novel fast online placement algorithm on 2D partially reconfigurable devices

Multi-Shape Task Placement Algorithm Based on Low Fragmentation Resource Management on 2D Heterogeneous Dynamic Partial Reconfigurable Devices

“…Similar to previous studies [10]- [17], the heterogeneous multi-shape task is able to be placed at arbitrary locations on the DPR device as long as there are enough unoccupied programmable resources to be assigned. And it is assumed that there are no dependency constraints between different tasks.…”

“…In order to avoid the DPR scheme degrading the performance of the reconfigurable device, the DPR control system has to decide a better task placement location for each task and update the programmable resource status efficiently. Various approaches have been put forward to solve the task placement and resource management problems for a 2D or 3D DPR device [10]- [17]. However, most existing algorithms assume that the task is a regular rectangle shape that results in unnecessary resource waste.…”

“…The problem of searching for the optimal execution locations on a DPR device for a series of tasks at run-time is known to be an NP-hard problem [18]. Based on different resource management methods, various task placement algorithms have been proposed for the 2D DPR device [10]- [15].…”

“…For homogeneous DPR device architectures, to keep the maximum unoccupied area in the center of the 2D DPR device, Marconi et al [10] presented a quad-corner (QC) task placement strategy that tries to place new tasks in the four corners of the device as much as possible according to the task size as follows: upper-left corner first (for very large tasks), upper-right corner first (for large tasks), lower-right corner first (for medium size tasks), and lower-left corner first (for small tasks). The QC strategy is highly efficient since the strategy searches candidates in the corner corresponding to the size of the new task to find the available location for the task, rather than traversing the entire resource matrix.…”

“…Programmable resources in the unoccupied internal area get wasted because they are treated as occupied resources according to the rectangular task model so that blocking the possibility of the space being assigned to other tasks. On the other hand, most existing approaches [10], [12]- [16], [19] consider that the DPR device only contains one kind of programmable resource: CLB. However, to improve the computing performance, the recent DPR device architectures (e.g., Xilinx 7 Series FPGAs, Zynq-7000 SoC, etc.)…”