If you did not already know

shining
Spark is an in-memory analytics platform targeted at today’s commodity server environments. It relies on Hadoop Distributed File System (HDFS) to hold intermediate checkpoint states and final processing results. Spark uses immutable data to store data updates for each iteration, making it inefficient for long-running iterative workloads. A non-deterministic garbage collector aggravates this problem. Sparkle is a library that optimizes memory usage in Spark. Leverages large shared memory for better data shuffling and intermediate storage. Sparkle replaces the current TCP/IP-based shuffle with a shared memory approach and proposes an off-heap memory store for efficient updates. We ran a series of experiments on a scale-out cluster and a scale-up machine. An optimized shuffle engine that leverages shared memory provides 1.3x to 6x faster performance than Vanilla Spark. Off-heap memory store and shared-memory shuffle engine deliver over 20x performance improvement for probabilistic graph processing workloads with large real-world hyperlinked graphs. Sparkle benefits most from running on large memory machines, but it also delivers 1.6x to 5x performance improvement over scale-out clusters with comparable hardware settings. …

multi-layer relationship network
A relational network (RN) introduced by Santoro et al. (2017) demonstrated powerful relational reasoning capabilities in a fairly shallow architecture. However, its single-layer design only considers pairs of information objects, so it is not suitable for problems that require reasoning over a large number of facts. To overcome this limitation, we propose a multi-layer relation network architecture that allows continuous refinement of relational information through multiple layers. With increasing depth, we apply it to the bAbI 20 QA dataset and solve all 20 tasks with joint training, outperforming state-of-the-art results, showing that more complex relational inference is possible. . …

Short-time Fourier transform (STFT)
Recently, we proposed a short-time Fourier transform (STFT)-based loss function for training neural speech waveform models. In this paper, we generalize the above framework and propose training schemes for such models based on spectral magnitude and phase loss obtained by STFT or Continuous Wavelet Transform (CWT), or both. Since CWT can have different time and frequency resolution than STFT and is cable considering closer to the human auditory scale, the proposed loss function can provide complementary information about speech signals. Experimental results showed that it is possible to train a high-quality model using the proposed CWT spectral loss and is as good as using STFT-based loss. …

Supermodel
The supermarket model usually refers to a system with a large number of queues. Arriving customers randomly choose a $d$ queue and join the queue with the fewest customers. The supermarket model outperforms even a small selection compared to simply joining a uniformly randomly selected queue due to the load balancing system. …