Abstract
In this article, we study the suitability of different computational accelerators for the task of real-time data processing. The algorithm used for comparison is the polyphase filter (PPF), a standard tool in signal processing and a well-established algorithm. We measure performance in FLOPs and execution time, which is a critical factor for real-time systems. For our real-time studies, we have chosen a data rate of 6.5 GB/s, which is the estimated data rate for a single channel on the Square Kilometre Array's (SKA) Low Frequency Aperture Array. Our findings show that GPUs are the most likely candidate for real-time data processing, excelling in both performance and power consumption.
Introduction
Modern radio astronomy projects, such as the Square Kilometre Array (SKA), represent a significant challenge in big data management. These telescopes are essentially massive data factories, generating information at rates that make long-term storage of raw data financially and physically impossible. For instance, a single channel from the SKA’s Low Frequency Aperture Array (LFAA) is predicted to output data at approximately 6.5 GB/s. To extract scientific value, this data must be processed in real-time to identify transient events like pulsars or fast radio bursts, allowing the system to discard the vast majority of "uninteresting" background noise.
A central component of this real-time pipeline is the Polyphase Filter (PPF). The PPF serves as a high-quality bandpass filter, preparing data for subsequent Fourier Transforms while minimizing artifacts such as spectral leakage and scalloping loss. However, the PPF is computationally intensive, necessitating a rigorous evaluation of modern hardware accelerators to determine which can sustain the required throughput within a reasonable power budget.
Methodology and Filter Design
The researchers implemented the PPF algorithm across three distinct hardware architectures to compare their efficacy:
- CPUs: Dual Intel Xeon E5-2650 processors utilizing OpenMP for parallelization and AVX intrinsics for vector processing.
- Xeon Phi: An Intel 5110P many-core accelerator, designed to offer a familiar x86 programming environment with high core counts.
- GPUs: Three generations of NVIDIA architectures, including Fermi (GTX 580), Kepler (Tesla K40), and Maxwell (GTX 750 Ti), programmed using CUDA.
A critical design choice was the use of a Finite Impulse Response (FIR) filter rather than an Infinite Impulse Response (IIR) filter. While IIR filters can be more computationally efficient for certain responses, FIR filters offer superior stability and are significantly easier to parallelize across many-core architectures. In the context of scientific observation, the finite nature of the FIR response ensures that a single data error or spike does not "ring" or propagate indefinitely through the output stream.
Performance Analysis and Results
Performance was evaluated using the "multiple of real-time" metric ($N$). A value of $N > 1$ indicates that the hardware can process the data faster than it arrives, a prerequisite for real-time deployment. The benchmarks accounted for the PCIe bottleneck, distinguishing between "Single Mode" (including host-to-device transfer) and "Streaming Mode" (processing data already resident in device memory).
Key Findings
- GPU Dominance: The NVIDIA Kepler Tesla K40 emerged as the performance leader, achieving a real-time multiple of 6.89. This indicates substantial headroom for additional processing tasks beyond the PPF.
- Power Efficiency: The Maxwell-based GTX 750 Ti demonstrated remarkable efficiency, handling the workload while drawing only 60W, a fraction of the power required by high-end enterprise cards or dual-CPU setups.
- The Xeon Phi Trade-off: The Xeon Phi 5110P achieved a multiple of 2.25. While it outperformed the dual-Xeon CPU setup (1.92), it lagged behind the GPUs. Its primary advantage remains the lower barrier to entry for developers, as it allows for the reuse of standard C++ code with minimal modification.
Conclusion
The study concludes that GPUs are currently the most viable solution for the SKA's real-time processing needs, offering the best balance of raw throughput and power efficiency. While CPUs and the Xeon Phi are capable of meeting the real-time threshold, they provide less headroom for the complex pipelines required in modern radio astronomy. Future work will focus on integrating these optimized implementations into the Astro-Accelerate library and exploring the role of FPGAs in the broader processing chain.
Summary updated