Profiling the Performance Anomalies of Multi-Source Media Downloading at Scale in the Wild

As one of the most fundamental and pervasive Internet services, me- dia file downloading has undergone several generations of enabling technologies. Unfortunately, the performance today is still far from satisfactory. As the state-of- the-art approach to accelerating media file downloads, multi-source downloading enables user clients to utilize multiple data sources and various content deliv- ery techniques. However, without careful designs, multi-source downloading can result in worse performance with higher overhead, referred to as an anomaly. This paper conducts the first empirical study to quantitatively understand the per- formance anomalies of multi-source media downloading, based on the production logs of a large-scale system serving 179M media file downloads for 37M users (including both PC and mobile users) per day. We reveal the characteristics and root causes of manifold anomalies with regard to seven types of data sources. In particular, 23% of the downloads accelerated by using multiple data sources become slower than the original single-source downloading, and there are sweet spots between the number of data sources used and the download speed. Also, we exploit some unconventional metrics (e.g., diversity of participation time) to explain some counter-intuitive anomalies. Accordingly, we provide a series of practical and applicable implications to effectively address the anomalies.