Accelerated high throughput parallel search on binary search trees via field programmable gate arrays

Field Programmable Gate Arrays are reprogrammable semiconductors that can be programmed according to the desired application after production, they can execute different instructions at the same time without performance loss. Binary Searh Tree is a data structure that stores the data as a sorted tree order. The aim is to increase the throughput by taking advantage of the Field Programmable Gate Array's parallel and pipeline capable architecture. The tree is split into partitions by different techniques to create an environment that lets more keys to be searched at the same time. The approach where each level of the tree is placed to a different block ram therefore increasing the number of ports that can be used, is called horizontal partitioning. To increase the throughput achieved by horizontal partitioning, duplication method is proposed. Duplication method creates duplicates of the tree to make it possible to have more than two keys searching the same level of the tree. Another duplication approach where some but not all of the tree levels are duplicated is also explained. By storing the first levels of the tree inside of registers instead of block memories, the block memories are fully utilized and the number of duplicated nodes are decreased. The hybrid approach whose aim is to have similar throughput to duplication however by not duplicating the tree, is proposed. Although hybrid approach can reach the high throughput levels of the duplication approach, due to not having enough ports to be able to fetch the same amount of keys the search may stall. The horizontal and duplicate methods do not have stalls since they have one port for each key to be searched. As a solution to the stalls, buffers were added to hybrid implementations and two ways of mapping the keys to the buffers were suggested: direct and queue based. The queue based approach decreases the stalls more efficiently than the direct approach however since it has a more complex structure it needs more clock period to function and more resources. With the duplication method the throughput has increased 8X, if there is not enough resources the hybrid method can be used. The throughput of the hybrid approach is not constant unlike the other approachs, its throughput converges to duplication in the best case and horizontal partitioning in the worst case.