A Case for Micro-Cellstores: Energy-Efficient Data Management on Recycled Smartphones∗ Stavros Harizopoulos Spiros Papadimitriou HP Labs Google Research Palo Alto, CA, USA Mountain View, CA, USA [email protected] [email protected] ABSTRACT computational power of these phones is similar to that of Increased energy costs and concerns for sustainability make all 500 top supercomputers in the world combined|but at the following question more relevant than ever: can we turn a fraction of their energy needs. How can this power be old or unused computing equipment into cost- and energy- harnessed? efficient modules that can be readily repurposed? We be- A place where cost- and energy-efficient computing units lieve the answer is yes, and our proposal is to turn unused could be utilized at large numbers is a modern data center. smartphones into micro-data center composable modules. In Data center operating costs are characterized by a continu- this paper, we introduce the concept of a Micro-Cellstore ously growing energy cost component [5, 2]. Power and cool- (MCS), a stand-alone data-appliance housing dozens of re- ing costs are soon expected to surpass the (amortized) cost of cycled smartphones. Through detailed power and perfor- purchasing servers. Demand for new and bigger data centers mance measurements on a Linux-based current-generation is on the rise, fueled by both consumer and enterprise appli- smartphone, we assess the potential of MCSs as a data man- cations. However, could a significantly underpowered device agement platform. In this paper we focus on scan-based par- support applications that typically run on high-end servers? titionable workloads. We show that smartphones are overall In this paper, we argue that for certain classes of enterprise more energy efficient than recently proposed low-power al- data management problems, such as data warehousing and ternatives, based on an initial evaluation over a wide range analytics, there are several emerging trends that lend them- of single-node database scan workloads, and that the gains selves to a micro-data center design based on underpowered become more significant when operating on narrow tuples hardware (also known as\wimpy nodes"in the literature [3]). (i.e., column-stores, or compressed row-stores). Our initial These trends are (a) MPP-style processing (massively par- results are very encouraging, showing efficiency gains of up allel processing), (b) column-oriented and compressed data to 6×, and indicate several promising future directions. which ease pressure on the memory/network buses, and (c) offering reliability through replication instead of expensive hardware solutions. Categories and Subject Descriptors Our proposal is to repurpose old or unused smartphones H.4 [Information Systems Applications]: Miscellaneous and use them to assemble units, called Micro-Cellstores, that contain dozens of interconnected smartphones which collec- 1. INTRODUCTION tively act as a data-appliance mini-cluster. A concept dia- Modern smartphones have the computational power of a 5-year-old PC, but at a fraction of the size and energy con- sumption (110x smaller volume than a standard 1U server, and 200x less peak power). More than 1 billion cellphones SHIELD are shipped yearly; in 2010, according to IDC, over 300 mil- lion of those were smartphones (a 74.4% increase over 2009). Smartphones have a typical consumer refresh cycle of two to MICRO-USB DOCK three years. Over the next few years, we expect a total of one billion smartphones to become obsolete; the aggregate ∗The views contained herein are the authors' only and do not necessarily reflect the views of Hewlett-Packard or Google. ETHERNET Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies POWER ROUTER bear this notice and the full citation on the first page. To copy otherwise, to EXTERNAL STORAGE republish, to post on servers or to redistribute to lists, requires prior specific POWER SUPPLY permission and/or a fee. M C S U N I T DaMoN 2011 Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$10.00. Figure 1: Micro-Cellstore Architecture. gram of a Micro-Cellstore is shown in Figure 1. There are wimpy clusters were shown to be more energy efficient com- several interesting questions around architecting and manu- pared to traditional clusters built using more power-hungry facturing Micro-Cellstores that are beyond the scope of this server nodes. However, this result may not hold in scenarios paper, such as: What are viable methods for networking? such as database workloads which often exhibit sub-linear How can batteries and power management features be lever- scale-up characteristics, especially when full cluster cost is aged? What is the right ratio of phones, routers/hubs, and considered [14]. external storage? Is cooling a problem? Our work in this paper can be viewed as in the same cate- Our focus in this paper is exploring the types of data gory as the above-mentioned wimpy-node architectures. To management workloads that can be efficiently supported by our knowledge, we are the first to characterize the energy- MCS units, and compare the energy-efficiency of appliances efficieny of modern smartphones when running database- based on smartphones (ultra-wimpy nodes) against other style tasks. Throughout this paper, we define energy effi- low-power alternatives. Our contributions are the following: ciency of a workload as the ratio of the query completion rate (e.g., scans per hour) to the average power consumed • Detailed power-profile characterization of a modern by the system. smartphone (Nexus S, released in Q4 2010). • Power efficiency measurements for partitionable, scan- 3. A CASE FOR MICRO-CELLSTORES intensive database workloads on smartphones and two There are three main emerging trends in enterprise data types of wimpy platforms. management that lend themselves to a micro-data center • Introducing the case for Micro-Cellstores (MCS). design based on underpowered hardware: The rest of the paper is organized as follows. In Section 2 we • Massively Parallel Processing (MPP): Parallel DBMSs cover related work, including recent proposals for \wimpy" typically adopt the shared-nothing paradigm for scal- architectures. Section 3 motivates Micro-Cellstores and Sec- ing out (rather than scaling up) to deal with increas- tion 4 details the characteristics of modern smartphones. ingly larger data volumes. For queries that scale lin- Section 5 carries out our benchmarking and analysis of var- early with the number of nodes in a cluster, an under- ious single-node, scan-based database workloads. We con- powered cluster could reach acceptable performance clude in Section 6. levels by using more nodes. 2. RELATED WORK • Column-oriented and highly compressed data: Column- stores have emerged as the prevalent architecture for Energy concerns are important enough to often dictate high-performance data management. Operating on where data centers are built. A growing number of efforts columnar, highly compressed data eases pressure on to improve the energy efficiency of clusters and data centers the memory and network/IO buses (which are typi- include holistic redesigns that treat a data center as a sin- cally under-specced in a smartphone, due to concerns gle computer [4, 16], cluster workload consolidation to meet over manufacturing cost). power constraints and reduce energy requirements [15, 13, 12], and considerations of low-power architectures [3, 19]. In • Reliability through replication: Modern systems in- this section we briefly discuss recent efforts in improving the creasingly rely on replication for providing reliability, energy efficiency of database applications. rather than on expensive hardware-based solutions. Such techniques are particularly suitable for smart- Energy efficiency in databases. Traditionally, database phones which do not compare well to server-grade com- systems have been optimized for performance, ignoring power- ponents with respect to reliability. related costs. However, the proliferation of scale-out archi- Furthermore, recent work has demonstrated running MapRe- tectures has forced data management systems to consider duce jobs on a network of smartphones [6]. Micro-Cellstores energy as equally important to performance. Early research are inspired by the above observations, combined with the studies argued for the redesign of several key components expected abundance of used smartphones in the future (as such as the query optimizer, the workload manager, the explained in the introduction). scheduler and the physical database design [7, 11, 9, 20]. In the concept of Figure 1, the proposed housing struc- Many of these suggestions assumed that, like cars, computer ture contains standardized micro-USB connectors and, pos- systems had different optimal performance and energy effi- sibly, WiFi routers for connectivity. We also expect that ciency points. However, a subsequent detailed study on the there will be some form of storage directly connected to the energy efficiency of a single database server [18] found that, router/hub. Furthermore, batteries may be leveraged in in- because of the start-up power draw, the highest perform- teresting ways, e.g., to provide uninterrupted operation even ing configuration was also the most energy efficient. That under intermittent power availability, to charge during off- study did not consider multi-node configurations or low- peak hours at possibly cheaper rates, or to smooth out the power hardware. In this paper, we investigate the latter. cluster's power profile. Studying the tradeoffs between the different types of networking, deciding the best use for ex- Non-server architectures. In an effort to improve the en- ternal storage, and exploring ways to harness the batteries ergy efficiency of clusters, a number of studies have also are beyond the scope of this paper. considered the use of low-power \wimpy" nodes consisting of We expect MCSs based on cheaply acquired, used smart- low-power storage (SSDs) and processors (mobile CPUs) [3, phones to be environmentally sustainable, minimizing total 19, 14, 17].