HAEC – Highly Adaptive Energy-Efficient Computing

About

The collaborative research center HAEC is a first attempt to achieve high adaptivity and energy efficiency in such an integrated approach. At the circuit level, we focus on innovative ideas for optical and wireless chip-to-chip communication. At the network level, we research secure, high performance network coding schemes for wired and wireless board-to-board communication. Innovative results at the hardware/software interface level will include energy control loops, which allow hardware to adapt to varying software requirements and vice versa. Software development in general is supported by energy-aware runtimes, energy-aware resource, stream and configuration management schemes and by an analysis framework for high performance/low energy applications. New internet applications are supported by innovations in energy-aware service execution. And, last but not least, formal methods are developed to offer a new quality of assurance in our systems of tomorrow. Demonstrating our results in a joint prototype – the HAEC Box – our goal is to become a pace setter for industry and academia on the design of future energy efficient-computing systems.

The Database Systems Group is involved into two projects of HAEC which are briefly described in the following.

Project B05

The targeted architecture of the HAEC Box allows us to place millions of cores into a single box that are hierarchically organized in sockets and boards communicating via an adaptive communication network, which uses optical and short-range wireless interconnects. To apply energy-related optimizations on such an architecture
at runtime, high-level components of the energy-control loop require facilities to get a consistent view of the system as well as to recongure its hardware and software components. Thus, two main application areas for this project in HAEC are (1) to monitor the current system state by processing the vast amount of hardware
and software data captured by sensors spread all over the HAEC Box and (2) to act as the central “Knowledge Plane” where depending projects are able to access the current and historic hardware topology and state of the HAEC Box, to reconfigure the topology of the system, and to share data with other components of the energy-control loop. The foundation of the project is the in-memory storage engine ERIS, which was developed during Phase I, that implements the core data management functionality. Our storage engine employs a novel index-centric query processing model to be update-friendly on the one hand and to offer a superior analytical performance on the other hand. From the HAEC applications perspective, we employ multiple concepts to satisfy the needs of our HAEC application areas. In particular, we focused on synopses and the “need-to-know” principle in Phase I to collect only sensor data that is necessary to fulfil the requirements of the consuming applications. In Phase II, we will shift our focus to the concept of external programmability of our storage engine to give depending projects the ability to get a more sophisticated access to the data by injecting custom operators and procedural code to ease the efficient implementation of energy-efficient analytical algorithms that leverage our processing infrastructure. From the hardware and software perspective, the aforementioned architecture of the HAEC Box challenges us in multiple facets to accomplish our primary goal of energy awareness. For instance, the memory model of the HAEC Box will exhibit strong non-uniform memory access (NUMA) effects which lead to higher latencies and a lower bandwidth when accessing remote memory based on the current configuration of the adaptive communication network of the box. Additionally, runtime adaptivity, i.e. the adaption of hardware components like powering off cores or sockets at runtime, as well as the instruction set architecture (ISA) of the employed cores have a huge impact on the overall energy efficiency of the system. In Phase I, we already did fundamental research on the concepts of NUMA awareness, runtime adaptivity, and ISA awareness and built a first prototype of our infrastructure. For Phase II, we intent to continue research on those concepts and add data placement adaptivty, storage adaptivity as well as code adaptivity as additional vehicles to achieve energy awareness. Besides investigating the four concepts in isolation, we will orchestrate all vehicles for energy awareness in a common energy/utility-based model and employ our data management infrastructure as a concrete example for the HAEC energy-control loop.

Project B08

HAEC takes a comprehensive approach to energy eciency, where individual components of hardware and software are not studied in isolation but in a greater context that allows the system to make optimal decisions at any point in time. To achieve this – and indeed even to evaluate whether it has been achieved yet – it is essential to confront the system with practically relevant computational workloads that can challenge the novel, adaptive architecture of HAEC on all levels. Important steps have already been taken towards this goal.
Project B05 has produced an efficient in-memory storage engine that will be adapted to the HAEC memory model in Phase II. Project B02 (Baader/Turhan) has developed new approaches to context awareness, which require such large scale data management facilities These advances have created a new opportunity for establishing a powerful knowledge processing pipeline within HAEC, which encourages the creation of a dedicated project. To bring the HAEC platform to its full potential in this field, it is of critical importance to expand the interface between application level and data management level. In Phase I, a simple form of SQL queries was used to exchange information between these levels. Based on the insights gathered so far, it is now possible to replace this initial solution with one that takes the specifific capabilities of HAEC into account.
To address this task, we plan to create a new bridge between high-level knowledge-driven applications and low-level data management infrastructure in HAEC. This should enable applications to express complex information needs in a concise, efficient, and declarative way while exploiting the specific hardware and software characteristics of the HAEC Box. We believe that navigational query languages are the appropriate tool for this: they can capture advanced computations, such as graph search operations, yet they have a fully declarative semantics that is a sound basis for data exchange across application boundaries. Given this “middle ground,” we obtain two clearly defined main targets: (1) high-level applications must exploit the power of navigational queries to reduce communication, and (2) low-level components must answer such queries efficiently on the given computational architecture. Neither of these would make sense in isolation, since the gap can only be closed if both sides agree on the details of navigational query language to be used. The evaluation of the work will consider two highly relevant use cases. The first is ontology-based query answering over large knowledge graphs, which is an important challenge for modern data management systems. Our work will be based on Wikidata, the new knowledge graph of Wikipedia, as an ideal example of a complex, dynamic dataset of high practical relevance. As a second use case, we will consider the context awareness approach developed in B02.