High performance reconfigurable architectures for biological sequence alignment

Abstract

Bioinformatics and computational biology (BCB) is a rapidly developing multidisciplinary field which encompasses a wide range of domains, including genomic sequence alignments. It is a fundamental tool in molecular biology in searching for homology between sequences. Sequence alignments are currently gaining close attention due to their great impact on the quality aspects of life such as facilitating early disease diagnosis, identifying the characteristics of a newly discovered sequence, and drug engineering. With the vast growth of genomic data, searching for a sequence homology over huge databases (often measured in gigabytes) is unable to produce results within a realistic time, hence the need for acceleration. Since the exponential increase of biological databases as a result of the human genome project (HGP), supercomputers and other parallel architectures such as the special purpose Very Large Scale Integration (VLSI) chip, Graphic Processing Unit (GPUs) and Field Programmable Gate Arrays (FPGAs) have become popular acceleration platforms. Nevertheless, there are always trade-off between area, speed, power, cost, development time and reusability when selecting an acceleration platform. FPGAs generally offer more flexibility, higher performance and lower overheads. However, they suffer from a relatively low level programming model as compared with off-the-shelf microprocessors such as standard microprocessors and GPUs. Due to the aforementioned limitations, the need has arisen for optimized FPGA core implementations which are crucial for this technology to become viable in high performance computing (HPC).

This research proposes the use of state-of-the-art reprogrammable system-on-chip technology on FPGAs to accelerate three widely-used sequence alignment algorithms; the Smith-Waterman with affine gap penalty algorithm, the profile hidden Markov model (HMM) algorithm and the Basic Local Alignment Search Tool (BLAST) algorithm. The three novel aspects of this research are firstly that the algorithms are designed and implemented in hardware, with each core achieving the highest performance compared to the state-of-the-art. Secondly, an efficient scheduling strategy based on the double buffering technique is adopted into the hardware architectures. Here, when the alignment matrix computation task is overlapped with the PE configuration in a folded systolic array, the overall throughput of the core is significantly increased. This is due to the bound PE configuration time and the parallel PE configuration approach irrespective of the number of PEs in a systolic array. In addition, the use of only two configuration elements in the PE optimizes hardware resources and enables the scalability of PE systolic arrays without relying on restricted onboard memory resources. Finally, a new performance metric is devised, which facilitates the effective comparison of design performance between different FPGA devices and families. The normalized performance indicator (speed-up per area per process technology) takes out advantages of the area and lithography technology of any FPGA resulting in fairer comparisons.

The cores have been designed using Verilog HDL and prototyped on the Alpha Data ADM-XRC-5LX card with the Virtex-5 XC5VLX110-3FF1153 FPGA. The implementation results show that the proposed architectures achieved giga cell updates per second (GCUPS) performances of 26.8, 29.5 and 24.2 respectively for the acceleration of the Smith-Waterman with affine gap penalty algorithm, the profile HMM algorithm and the BLAST algorithm. In terms of speed-up improvements, comparisons were made on performance of the designed cores against their corresponding software and the reported FPGA implementations. In the case of comparison with equivalent software execution, acceleration of the optimal alignment algorithm in hardware yielded an average speed-up of 269x as compared to the SSEARCH 35 software. For the profile HMM-based sequence alignment, the designed core achieved speed-up of 103x and 8.3x against the HMMER 2.0 and the latest version of HMMER (version 3.0) respectively. On the other hand, the implementation of the gapped BLAST with the two-hit method in hardware achieved a greater than tenfold speed-up compared to the latest NCBI BLAST software. In terms of comparison against other reported FPGA implementations, the proposed normalized performance indicator was used to evaluate the designed architectures fairly. The results showed that the first architecture achieved more than 50 percent improvement, while acceleration of the profile HMM sequence alignment in hardware gained a normalized speed-up of 1.34. In the case of the gapped BLAST with the two-hit method, the designed core achieved 11x speed-up after taking out advantages of the Virtex-5 FPGA. In addition, further analysis was conducted in terms of cost and power performances; it was noted that, the core achieved 0.46 MCUPS per dollar spent and 958.1 MCUPS per watt. This shows that FPGAs can be an attractive platform for high performance computation with advantages of smaller area footprint as well as represent economic ‘green’ solution compared to the other acceleration platforms. Higher throughput can be achieved by redeploying the cores on newer, bigger and faster FPGAs with minimal design effort.