An FPGA approach for fast bitmap indexing

Abstract

In this paper, an efficient architecture of an FPGA-based bitmap index creation is proposed. The design utilizes a content addressable memory together with a bit-level transpose matrix to index multi-record documents by several given keywords. The experiments in a Cyclone V SX FPGA proved that our circuit could attain throughput of 330.6 million records per second while only using around 71% of embedded memory together with 45% of lookup tables and registers. In fact, achieved throughput is 2.8 times and 1.7 times as high as that of CPU-based and GPU-based design, respectively.

A 1.2-V 162.9 pJ/cycle bitmap index creation core with 0.31-pW/bit standby power on 65-nm SOTB

“…This brief extends our previous work [4], [16] by originally proposing a proof-of-concept BIC chip in a 65-nm SOTB CMOS process. This chip could operate at a wide-range supply voltage between 0.4 V and 1.2 V. At 1.2 V, it was fully operational at 41 MHz and required 6.68 mW.…”

“…This was exemplified by an FPGA-based large scale deep neural network system [3], whose processing throughput was around 3.5 and 2.5 times higher and power consumption was four and 22 times lower than the 80-W CPU and 225-W GPU, respectively. Following this research trend, our previous work [4] proposed a 150-MHz FPGA-based BI creation (BIC) system that achieved 2.8 times and 1.7 times higher indexing throughput than CPU- [2] and GPU-based designs [5], respectively. Despite the remarkable achievement of performance, eliminating as much as leakage power during the standby mode is still one of the most challenging tasks.…”

An FPGA approach for high-performance multi-match priority encoder

“…The formula of achieved throughput (THR) is described in (3), where N is the PE size and FREQ is the operating frequency.…”

“…Finally, the ARM system returns those results to the host server for information extraction. More detail of bitmap index creator and query processor can be found in [3] and [13], respectively. The results of timing analyzer point out that MPE4K can operate at 87 MHz and therefore deliver THR of 356.4 Gbps.…”

“…To accelerate IR, search data must be indexed in advance. Among many indexing techniques, bitmap index is well-known for handling multiple-condition queries effectively, because of its strong support of parallel processing [2,3]. Bitmap index returns the query result as a bit sequence, where bit i indicates whether data at address i satisfies all given conditions.…”