Data scientists spend most of their time dealing with data preparation, rather than doing what they know best: build machine learning models and algorithms to solve previously unsolvable problems. In this paper, we describe the Visual Data Management System (VDMS), and demonstrate how it can be used to simplify the data preparation process and consequently gain in efficiency simply because we are using a system designed for the job. To demonstrate this, we use one of the largest available public datasets (YFCC100M), with 100 million images and videos, plus additional data including machine-generated tags, for a total of about ~12TB of data. VDMS differs from existing data management systems due to its focus on supporting machine learning and data analytics pipelines that rely on images, videos, and feature vectors, treating these as first class citizens. We demonstrate how VDMS outperforms well-known and widely used systems for data management by up to ~364x, with an average improvement of about 85x for our use-cases, and particularly at scale, for a image search engine implementation. At the same time, VDMS simplifies the process of data preparation and data access, and provides functionalities non-existent in alternative options.

Breadth-First Search (BFS) is the core of many graph analysis algorithms and it is used in many problems, such as social network, computer network analysis, and data organization. BFS is an iterative algorithm that due to its irregular behavior is quite challenging to parallelize. Several approaches implement efficient algorithms for BFS for multicore architectures and for Graphics Processors, but it is still an open problem how to distribute the work among the main cores and the accelerators. In this paper, we assess several approaches to perform BFS on different heterogenous architectures (highend and embedded mobile processors composed of a multi-core CPU and an integrated GPU) with a focus on social network graphs. In particular, we propose two heterogenous approaches to exploit both devices. The first one, called Selective, selects on which device to execute each iteration. It is based on a previous approach, but we have adapted it to take advantage of the features of social network graphs (fewer iterations but more unbalanced). The second approach, referred as Concurrent, allows the execution of specific iterations concurrently in both devices. Our heterogenous implementations can be up to 1.56x faster and 1.32x more energy efficient with respect to the best of only-CPU or only-GPU baselines. We have also found that for a highly memory bound problem like BFS, the CPU-GPU collaborative execution is limited by the shared-memory bus bandwidth.