Integrated 3D-stacked server designs for increasing physical density of key-value stores

Abstract

Key-value stores, such as Memcached, have been used to scale web services since the beginning of the Web 2.0 era. Data center real estate is expensive, and several industry experts we have spoken to have suggested that a significant portion of their data center space is devoted to key-value stores. Despite its wide-spread use, there is little in the way of hardware specialization for increasing the efficiency and density of Memcached; it is currently deployed on commodity servers that contain high-end CPUs designed to extract as much instruction-level parallelism as possible. Out-oforder CPUs, however have been shown to be inefficient when running Memcached.To address Memcached efficiency issues, we propose two architectures using 3D stacking to increase data storage efficiency. Our first 3D architecture, Mercury, consists of stacks of ARM Cortex-A7 cores with 4GB of DRAM, as well as NICs. Our second architecture, Iridium, replaces DRAM with NAND Flash to improve density. We explore, through simulation, the potential efficiency benefits of running Memcached on servers that use 3D-stacking to closely integrate low-power CPUs with NICs and memory. With Mercury we demonstrate that density may be improved by 2.9×, power efficiency by 4.9×, throughput by 10×, and throughput per GB by 3.5× over a state-of-the-art server running optimized Memcached. With Iridium we show that density may be increased by 14×, power efficiency by 2.4×, and throughput by 5.2×, while still meeting latency requirements for a majority of requests.

A scalable processing-in-memory accelerator for parallel graph processing

“…Specialized accelerators for database systems [7,30,59], key-value stores [34], and stream processing [49] have also been developed. Several studies have proposed 3D-stacked system designs targeting memory-intensive server workloads [13,28,35,50]. Tesseract, in contrast, targets large-scale graph processing.…”

“…In order to tackle the first challenge, recent studies have proposed specialized on-chip accelerators for a limited set of operations [13,30,34,59]. Such accelerators mainly focus on improving core efficiency, thereby achieving better performance and energy efficiency compared to general-purpose cores, at the cost of generality.…”

Architecting to achieve a billion requests per second throughput on a single key-value store server platform

“…Research efforts [14,15,28,32] in this direction achieve up to 13.2MRPS [14] with a 10GbE link and more than 10X improvements on energy efficiency compared to commodity servers running stock memcached. There are also non-FPGA-based architecture proposals [20,30,33,37] for accelerating memcached and/or improving performance and efficiency of various datacenter workloads.…”

Heterogeneous chip multiprocessor architectures for big data applications

“…However, while demand for data center computational resources continues to grow as the size of data grows, the semiconductor industry has reached its physical scaling limits and is no longer able to reduce power consumption in new chips. Physical design constraints, such as power and density, have therefore become the dominant limiting factor for scaling out data centers [3,4,5]. Current server designs, based on commodity homogeneous processors, will therefore not be the most efficient in terms of performance/watt to process big data applications [5,6].…”

“…Physical design constraints, such as power and density, have therefore become the dominant limiting factor for scaling out data centers [3,4,5]. Current server designs, based on commodity homogeneous processors, will therefore not be the most efficient in terms of performance/watt to process big data applications [5,6]. Big data applications require computing resources that can scale to manage massive amounts of diverse data.…”