Energy Efficient Heterogeneous Datacenter-on-Chip for Big Data Computing
Abstract
The importance of big data has been demonstrated in numerous applications that belong to diverse areas ranging from engineering and biology, to social media, economics and military. These applications are extremely relevant for big data centers employing machine learning and data analysis to solve problems once deemed too complex to produce any practical results. However, the design of data centers is dominated by power, thermal, and physical constraints. On the contrary, emerging heterogeneous manycore processing platforms consisting of CPU and GPU cores along with memory controllers (MCs) have small footprints and offer power and area-efficient trade offs for running big data applications. Consequently, heterogeneous manycore computing platforms represent a powerful alternative to the data center oriented type of computing. Under this new computing paradigm, ensuring an efficient communication among cores represents a major challenge. Indeed, the state-of-the-art on-chip communication infrastructures employed for conventional manycore platforms are sub-optimal at handling both CPU and GPU communication requirements efficiently. In addition to the communication bottleneck, finding optimal power and thermal management strategies represents another major concern. To address these challenges, we propose a holistic approach to design a high performance and energy-efficient heterogeneous datacenter-on-chip (HetDoC) for big data computing, which is relevant for several army applications. To this end, the main tasks of this proposal are to: 1) Design an optimal network-on-chip (NoC) architecture as the interconnection backbone for the HetDoC that can handle both CPU and GPU communication requirements efficiently 2) Develop a voltage-frequency island (VFl)-based power management strategy to make the HetDoC architecture energy-efficient 3) Evaluate power-thermal-performance (PTP) trade-offs for the proposed HetDoC by considering big data applications relevant to Anny s core mission (e.g., battle field planning, emergency response and logistics management, anomaly detection, etc.) We envision a HetDoC architecture targeting big data applications where the entire system (or a large part thereof) can be designed using a heterogeneous manycore-based single-chip architecture. At this massive level of integration, traditional NoC architectures, e.g., mesh, tree, ring, cannot provide a scalable, low latency and energy-efficient communication backbone, which is essential for solving the big data problems targeted in this work. On the other hand, wireless NoCs (WiNoCs) outperform conventional wireline NoC architectures both in terms of achievable bandwidth and energy dissipation. Consequently, we plan to explore design of heterogeneous (i.e., combination of CPUs and GPUs) systems combined with hybrid (i.e., combination of wired and wireless links) NoC architectures for big data computing. Moreover, to make the proposed HetDoC energy-efficient, it is imperative to incorporate suitable power management strategies. Big data application like deep learning has been successfully leveraged to improve the performance of planning and decision-making under uncertainty algorithms. The key idea is to use deep learning techniques to automatically learn generalized representations for policies and value functions that form the core components of these algorithms. Anny applications including real-time battle field planning, emergency response and logistics management will greatly benefit from the proposed HetDoC instantiation for deep learning techniques. Similarly, graph pattern matching over large-scale cyber networks forms the core computational process for cyber security and anomaly detection detection applications that are extremely important for army. Our proposed HetDoC design can provide a high-performance and energy-efficient computing platform for these security applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1710485
Entities
People
- Partha Pande
Organizations
- Army Contracting Command
- United States Army
- Washington State University