Aligner-D: Leveraging In-DRAM Computing to Accelerate DNA Short Read Alignment
Aligner-D: Leveraging In-DRAM Computing to Accelerate DNA Short Read Alignment
复制标题
DOI：
10.1109/jetcas.2023.3241545
复制
发表时间：
2023-03
影响因子：
4.6
通讯作者：
Fan Zhang;Shaahin Angizi;Jiao-Jin Sun;W. Zhang;Deliang Fan
中科院分区：
文献类型：
--
作者：
Fan Zhang;Shaahin Angizi;Jiao-Jin Sun;W. Zhang;Deliang Fan
DNA短读比对任务已成为下一代测序平台产生大量数据的主要测序瓶颈。本文提出了一种基于DRAM的高效、高通量内存处理（PIM）加速器（Aligner-D），该加速器采用最先进的BWT对齐算法执行DNA短读对齐。我们首先提出了利用DRAM内部高并行性和吞吐量的PIM设计。它将每个DRAM阵列转换为执行对齐任务的有效处理单元。所提出的Aligner-D可以有效地执行校准任务所需的批量逐位xnor匹配操作，只有3个晶体管/冷开销。然后，我们介绍了一种基于BWT的高度并行和定制的读取对齐算法，该算法支持精确和非精确匹配任务。接下来，我们介绍了如何映射对齐任务的相关数据，以最大限度地利用新硬件和算法的并行性。实验结果表明，与其他内存计算平台：Ambit （Seshadri等人，2017）、drsa - 1t1c （Li等人，2017）、drsa - 3t1c （Li等人，2017）和ReDRAM （Angizi和Fan， 2019）相比，Aligner-D分别获得了$\sim 4倍、$\sim 2.45倍、$\sim 3.26倍和$\sim 1.65倍的改进。对于DNA短读比对，Aligner-D在ReCAM、CPU、GPU、FPGA、Ambit和DRISA上，每瓦比对吞吐量分别提高$\sim 20104\times $、$\sim 3522\times $、$\sim 927\times $、$\sim 88\times $、$\sim 5.28\times $和$\sim 2.34\times $。
DNA short read alignment task has become a major sequential bottleneck to humongous amounts of data generated by next-generation sequencing platforms. In this paper, an energy-efficient and high-throughput Processing-in-Memory (PIM) accelerator based on DRAM (named Aligner-D) is presented to execute DNA short-read alignment with the state-of-the-art BWT alignment algorithm. We first present the PIM design that utilizes DRAM’s internal high parallelism and throughput. It converts each DRAM array to a potent processing unit for alignment tasks. The proposed Aligner-D can efficiently execute the bulk bit-wise XNOR-based matching operation required by the alignment task with only 3-transistor/col overhead. We then introduce a highly parallel and customized read alignment algorithm based on BWT that supports both exact and inexact match tasks. Next, we present how to map the correlated data of the alignment task to utilize the parallelism from both new hardware and algorithm maximumly. The experimental results demonstrate that Aligner-D obtains $\sim 4\times $ , $\sim 2.45\times $ , $\sim 3.26\times $ , and $\sim 1.65\times $ improvement, respectively, compared with other in-memory computing platforms: Ambit (Seshadri et al., 2017), DRISA-1T1C (Li et al., 2017), DRISA-3T1C (Li et al., 2017), and ReDRAM (Angizi and Fan, 2019). As for DNA short read alignment, Aligner-D boosts the alignment throughput per Watt by $\sim 20104\times $ , $\sim 3522\times $ , $\sim 927\times $ , $\sim 88\times $ , $\sim 5.28\times $ , and $\sim 2.34\times $ , over ReCAM, CPU, GPU, FPGA, Ambit, and DRISA, respectively.