Fast query processing with parallel languages

Recent efforts have shown that compiling queries to machine code for a single-core can remove iterator and control overhead for significant performance gains. So far, systems that generate distributed programs only compile plans for single processors and stitch them together with messaging. The low-level compiler can only look at these fragments in isolation! We are investigating a different approach with Radish: take advantage of parallel language compilers and runtimes by generating parallel programs from queries.

Verified relational query optimizers

Query optimizers need to apply transformations to plans to find the fastest implementation. How do we know the resulting program is correct and our query will return the intended answer? In correctness-critical applications, testing may not be sufficient to ensure queries return the right answer. Crimp is a query transformer written in Coq that generates verified imperative (C-like) implementations of query specifications (e.g., programs written in SQL).

Productivity and performance for parallel irregular applications

A number of important, data intensive applications are irregular. These applications may involve irregular and unbalanced data structures (such as natural graphs) and unpredictable data accesses. A large effort is usually required to achieve good performance for such programs, the key being supporting enough concurrency to keep the machine (efficiently) busy. The programmer must write non-application logic like batching/sorting communication and managing fine-grained synchronization.

Our system, Grappa, is a C++11 library that addresses this problem by providing an intuitive and flexible PGAS programming model and runtime support for high-performance irregular data access.

Dissertation

Papers

Demos