Abstract: Since January 2008 when the next-generation DNA sequencing platforms were developed, the sequencing throughput has been significantly improved. However, this technology has been challenged by the large amount of sequencing data which grows dramatically even over the Moore's Law. As an emerging data-intensive workload, the high-throughput re-sequencing tools like Hash-based programs shows different characteristics from traditional computational applications. Both low arithmetic intensity and irregular memory access pattern are major sources of inefficiency on commodity multi-core platforms. In this paper, we propose co-processor architecture for accelerating a short reads mapping algorithm. The complete mapping flow in one processing element (PE) is integrated to an exclusive memory port to improve the parallel performance. This proposed architecture is then implemented on a Convey HC-1ex reconfigurable computer. The design includes 64 parallel PEs on 4 Xilinx Virtex-6 LX760 that operate at 150MHz. Compared with an Intel Xeon 8-cores CPU, the speedup achieves 28.5 times, and the average memory read bandwidth achieves 22.59GBps. Therefore, this proposed design can potentially supply a solution to the large-amount data challenge and be applied in high throughput re-sequencing.

Key words: high-throughput sequencing, short reads mapping, Hash-index, field programmable gate array (FPGA), heterogeneous architecture