high performance computing on graphics processing units: hgpu.org

In this work we focus on tailoring and optimizing the computational Private Information Retrieval (cPIR) scheme proposed in WAHC 2014 for efficient execution on graphics processing units (GPUs). Exploiting the mass parallelism in GPUs is a commonly used approach in speeding up cPIRs. Our goal is to eliminate the efficiency bottleneck of the Dor"{o}z et al construction which would allow us to take advantage of its excellent bandwidth performance. To this end, we develop custom code to support polynomial ring operations and extend them to realize the evaluation functions in an optimized manner on high end GPUs. Specifically, we develop optimized CUDA code to support large degree/large coefficient polynomial arithmetic operations such as modular multiplication/reduction, and modulus switching. Moreover, we choose same prime numbers for both the CRT domain representation of the polynomials and for the modulus switching implementation of the somewhat homomorphic encryption scheme. This allows us to combine two arithmetic domains, which reduces the number of domain conversions and permits us to perform faster arithmetic. Our implementation achieves 14-34 times speedup for index comparison and 4-18 times speedup for data aggregation compared to a pure CPU software implementation. tion compared to a pure CPU software implementation.