Harmony: an execution model for heterogeneous systems

s away much detail about the actual execution of a program from application developers: developers only specify logical relationships between data; the Datalog compiler constructs a suitable execution plan by extracting data and control flow from the Datalogprogram. Compared to emerging distributed programming languages for GPUs such as Map-Reduce [64], Datalog expresses a richer set of relational and database operations that are cumbersome to implement in Map-Reduce. The abstraction and declarative nature that helps programmers reason about Datalog programs are exactly what can potentially help create high performaning evaluation plans for Datalog programs, as well. A Datalog program’s data flow is made explicit through the logical relations used to define data. Furthermore, well-defined properties associated with logical relations, such as commutativity and associativity of conjunctions, readily expose data parallelism in Datalog programs. Lastly, the limited number of control operators (e.g., one sequencing operator, no unlimited recursion), makes it easy to construct a finite data and control flow graph. The current Datalog compiler decomposes a program into such a data-control flow graph, which then serves as an evaluation plan for the LogicBlox runtime. At the time of this writing, the Datalog compiler and the runtime do not, however, take advantage of the readily exposed data, task, and pipeline parallelism in a Datalog program. Furthermore, the compiler/runtime do not take advantage of the different hardware available on different deployment environments. Lastly, the Datalog compiler does not have the ability to determine which part of the computation can be best executed where, given the available hardware. Since LogicBlox applications target the cloud where hardware per instance of deployment can vary, it is crucial for LogicBlox to be able to take advantage of the hardware available. The Redfox compiler replaces the existing LogicBlox compiler and decomposes the program into a graph of Harmony kernels. In this representation the graph exposes task and pipeline parallelism that can be exploited by the Harmony