Data Stream Processing in HPC Systems: new frameworks and architectures for high-frequency streaming
• 大类 : 工程技术 - 4区
• 小类 : 计算机：理论方法 - 4区
In the Data Stream Processing (DSP) computing paradigm, applications continuously collect, organize and analyze unbounded streams of data generated by an increasing number of sensing devices. Typical sources of streams are health-care devices, financial trading systems, emergency management infrastructures, smart vehicles and many others. In these complex scenarios, unbounded flows of data must be gathered and analyzed in real-time to extract useful information and to make timely informed decisions.
The ever-increasing volume of data and the highly irregular nature of data rates pose new challenges to DSP systems mainly concerning low-latency processing capabilities. Existing DSP frameworks mainly target conventional distributed systems aiming at near real-time processing and scale-out scenarios.
High-volume, low-latency and full real-time data streaming processing may significantly benefit from the extensive exploitation of current and forthcoming highly-parallel heterogeneous rack-scale systems, where each node is equipped with multi-GPUs and multi-FPGA accelerators aggregated at rack level by low-latency/high-bandwidth networks. The capacity of these highly-dense/highly-parallel rack-scale solutions has grown remarkably over the years, offering tens of thousands of heterogeneous cores and multiple terabytes of aggregated RAM reaching computing, memory and storage capacity of a large warehouse-scale cluster of just few years ago.
The optimization of rack-scale DSP systems will also contribute to help to reduce the latency and to improve the efficiency of more massive distributed infrastructures.
While GPUs are well-suited for offline data-parallel computation and have emerged as the leading platform for the deep-learning domain, the online use of multiple GPUs for streaming computations is still an open research problem. Moreover, the emerging FPGAs-based appliances and CPU-FPGA hybrids are nowadays opening the doors to utilize these server-attached FPGAs as accelerators to drastically reduce latency in time-critical scenarios.
Such a new and challenging scenario demands new run-time mechanisms, data structures, strategies, and algorithms, whose implementation may nurture novel interdisciplinary approaches. This Special Issue aims at collecting innovative proposals on how to design and build DSP systems and frameworks explicitly targeting highly-parallel rack-scale heterogeneous systems.
We solicit papers covering various topics of interests that include, but are not limited to the following:
- Multi-GPUs accelerated data stream processing
- FPGA-based accelerators for real-time stream processing
- State-aware management of streaming data and operators for rack-scale DSP systems
- Efficient data-movement of streams in heterogeneous many-cores systems
- Non-intrusive autonomic/elastic supports for highly-parallel data stream processing
- QoS-driven performance modeling for topologies of streaming operators
- Power-aware, energy-efficient data streaming algorithms and techniques
- Concurrent data structures for data streaming
- Optimization of existing DSP frameworks for highly-parallel many-cores and hybrid systems
- Use cases and applications of real-time analytics based on rack-scale DSP systems in various domains, including cyber-physical systems, healthcare, Internet of Things, Smart Cities, and social networks