Approximate Compression – Enhancing Compressibility rough Data Approximation Harini Suresh, Shashank Hegde, and John Sartori University of Minnesota Abstract CCS Concepts •Information systems ! Compression strategies Internet-connected mobile processors used in cellphones, tablets, 1 Introduction and internet-of-things (IoT) devices are generating and transmit- Traditional computing systems are increasingly being replaced and ting data at an ever-increasing rate. ese devices are already the outnumbered by mobile and cloud computing systems that rely most abundant types of processor parts produced and used today heavily on communication between internet-connected devices. and are growing in ubiquity with the rapid proliferation of mo- Along with this shi toward communication-centric computing, bile and IoT technologies. Size and usage characteristics of these technological advances and proliferation of internet-connected de- data-generating systems dictate that they will continue to be both vices are constantly increasing the amount of data produced and bandwidth- and energy-constrained. e most popular mobile ap- transmied in these computing systems [1–4]. is explosion in plications, dominating communication bandwidth utilization for data communication is particularly prominent for image, video, and the entire internet, are centered around transmission of image, audio data. For example, mobile phones with high-resolution cam- video, and audio content. For such applications, where perfect eras like the iSight camera in iPhone6s regularly exchange 12MP data quality is not required, approximate computation has been images through popular applications like Facebook and Instagram. explored to alleviate system bolenecks by exploiting implicit noise is represents a more than 100× increase in transmied data for a tolerance to trade o output quality for performance and energy single image compared to the images transmied by earlier phone benets. However, it is oen communication, not computation, models. Similar, or even greater increases in communication band- that dominates performance and energy requirements in mobile width usage are observable across the gamut of popular mobile systems. is is coupled with the increasing tendency to ooad applications, including streaming digital audio and video content, computation to the cloud, making communication eciency, not social media, video conferencing, and cloud storage [5]. In fact, computation eciency, the most critical parameter in mobile sys- nearly all of the applications that consume the vast majority of all tems. Given this increasing need for communication eciency, data internet bandwidth (e.g., Netix, Youtube, Facebook, iTunes, Insta- compression provides one eective means of reducing communica- gram, Snapchat, etc.) are centered around transmission of video, tion costs. In this paper, we explore approximate compression and image, and audio content [6]. Network technologies continue to communication to increase energy eciency and alleviate band- evolve to support increasing loads from an ever-increasing num- width limitations in communication-centric systems. We focus ber of communication-centric computing devices and applications, on application-specic approximate data compression, whereby a but device technology and application advances, along with usage transmied data stream is approximated to improve compression trends, ensure that the demand for more bandwidth always exceeds rate and reduce data transmission cost. Whereas conventional lossy the advances in communication and network technology. compression follows a one-size-ts-all mentality in selecting a com- With applications and computing platforms straining available pression technique, we show that higher compression rates can communication bandwidth, data compression presents one means be achieved by understanding the characteristics of the input data of reducing bandwidth utilization. Given trends toward more stream and the application in which it is used. We introduce a suite mobile, parallel, and cloud-based computing, it may oen be ad- of data stream approximations that enhance the compression rates vantageous to compress data locally before transmiing over a of lossless compression algorithms by gracefully and eciently trad- bandwidth-constrained link. At the same time, given that many of ing o output quality for increased compression rate. For dierent the killer applications for communication bandwidth operate on classes of images, we explain the interaction between compression noise-tolerant data (e.g., image, video, audio), lossy compression [7] rate, output quality, and complexity of approximation and establish presents a potentially-aractive approach to reduce bandwidth us- comparisons with existing lossy compression algorithms. Our ap- age further by sacricing some data delity for increased compres- proximate compression techniques increase compression rate and sion rates. In this paper, we recognize that exploiting noise toler- reduce bandwidth utilization by up to 10× with respect to state-of- ance in these applications to perform less costly approximate data the-art lossy compression while achieving the same output quality communication may be more eective for emerging applications and beer end-to-end communication performance. than exploiting noise tolerance to perform approximate computing. ACM Reference format: Along this vein, we expand on the eld of lossy compression by Harini Suresh, Shashank Hegde, and John Sartori designing a suite of approximate compression techniques for noise- University of Minnesota. 2017. Approximate Compression – Enhancing tolerant applications.1 We use classication of dataset features to Compressibility rough Data Approximation. In Proceedings of ESTIMe- create data-aware approximations that enhance the compressibility dia’17, Seoul, Republic of Korea, October 15–20, 2017, 10 pages. DOI: 10.1145/3139315.3139319 of data while maintaining acceptable data quality. e goal of our approximate compression techniques is to maximize the increase Permission to make digital or hard copies of all or part of this work for personal or in compression rate provided by approximation per unit reduction classroom use is granted without fee provided that copies are not made or distributed in quality. Our paper makes the following contributions. for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components of this work owned by others than ACM • We propose a suite of approximation techniques that enhance the must be honored. Abstracting with credit is permied. To copy otherwise, or republish, performance of existing lossless compression algorithms for image to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. Request permissions from [email protected]. ESTIMedia’17, Seoul, Republic of Korea 1To distinguish approximate compression from existing lossy compression techniques, © 2017 ACM. 978-1-4503-5117-1/17/10...$15.00 we describe approximate compression as data approximation followed by lossless DOI: 10.1145/3139315.3139319 compression. Harini Suresh, Shashank Hegde, and John Sartori ESTIMedia’17, October 15–20, 2017, Seoul, Republic of Korea University of Minnesota data. Our techniques leverage image characteristics to increase gains in compression rate per unit quality reduction. • We show that the approximation technique that maximizes com- pression rate depends on image characteristics and desired output quality and demonstrate that learning-based techniques (e.g., neu- ral networks) can be used to select an appropriate approximate compression technique for a given data stream. Figure 1. We propose a suite of approximation techniques to en- • Our approximation techniques allow graceful trade-os between hance the compressibility of the input stream fed to a lossless data quality and compression rate, allowing a sweet spot to be compression algorithm. selected based on user preferences or application constraints. of bits required for encoding is equivalent to the entropy of the • We perform a thorough evaluation of our suite of approximate symbols generated by the source. Reducing the number of bits compression techniques over the trade-o space between quality required to encode the input stream to below its entropy becomes and compression rate and show that our data-aware approximate possible with the introduction of distortions to the data stream. compression techniques result in up to 10× improvement in com- Lossy compression introduces distortions through transforma- pression rate with respect to state-of-the-art lossy compression tion and quantization of the input data, which are then losslessly techniques, for the same output quality. compressed. Transformation algorithms like Discrete Cosine Trans- • We evaluate the performance of our approximate compression form, and Discrete Wavelet Transform are used to transform the techniques against state-of-the-art lossy compression and show that input to a form with lower entropy that requires fewer bits for they improve end-to-end performance by over 2×, on average, while encoding. Transformation may also aid in identifying components achieving beer compression rates for the same output quality. We of a data stream that are irrelevant to human perception. also show that our approximation techniques achieve performance is paper demonstrates how dierences in data characteristics scalability on massively-parallel processors, indicating