Temporal locality in today's content caching: why it matters and how to model it

Stefano Traverso, Mohamed Ahmed, Michele Garetto, Paolo Giaccone, Emilio Leonardi, Saverio Niccolini
Appears in: 
CCR October 2013
The dimensioning of caching systems represents a difficult task in the design of infrastructures for content distribution in the current Internet. This paper addresses the problem of defining a realistic arrival process for the content requests generated by users, due its critical importance for both analytical and simulative evaluations of the performance of caching systems. First, with the aid of YouTube traces collected inside operational residential networks, we identify the characteristics of real traffic that need to be considered or can be safely neglected in order to accurately predict the performance of a cache. Second, we propose a new parsimonious traffic model, named the Shot Noise Model (SNM), that enables users to natively capture the dynamics of content popularity, whilst still being sufficiently simple to be employed effectively for both analytical and scalable simulative studies of caching systems. Finally, our results show that the SNM presents a much better solution to account for the temporal locality observed in real traffic compared to existing approaches.
Public Review By: 
Augustin Chaintreau

“I’m not attracted and we don't get along, but the timing is right.” New Yorker Cartoon by Barbara Smaller If timing, as we are often reminded, is everything, it’s also a critical part of today’s caching system. The main reason is that the popularity of an item -- like a video on YouTube, as considered by the authors of this paper -- evolves quickly and it can easily fool a simplistic cache management strategy. In particular a heuristic model called independent reference is often used in evaluating caching networks even today, and this precisely states that popularity is fixed with time and that a new request is independent of the past. In reality requests typically occur in bursts, a phenomenon sometimes called temporal locality. How much of this locality matters in today’s caching? Which methods can be used in practice to extract and recognize it when it occurs? Are there models that retain some of the simplicity of the independent reference model while accounting for temporal locality? This paper answers the above questions by first conducting a longitudinal study of web-traffic inside a few operational Points of Presence (PoP) accounting for 60k users. It shows that temporal locality is present and that it generally implies that the performance derived for a simple caching scheme is unnecessarily pessimistic. It then makes two novel contributions: First, it proposes and evaluates a concrete method based on slicing and reshuffling to modulate temporal locality in a trace, showing how it directly impacts performance. Second, it builds a new temporal model of request generation that operates at a middle ground between mathematical simplicity and realism. The model reuses the above slicing and is inspired from Shot Noise Processes, a relatively mature tool of applied probability which allows one to sum independent processes that still exhibit themselves some temporal evolution. None of these tools and methods have ever been applied to caching for on-demand video or other content delivery. All reviewers appreciated the substance and the timing of this work to propose a different model of cache request. In spite of various classical results, we still rely too blindly on the independent reference model, owing to its simplicity. In addition, this work was quoted as a good example of applying sophisticated analysis to a real data set. Reviews also highlighted some important limitations of these current results: Relying on a single cache algorithm (LRU) for all claims makes the paper less relevant in practice and it reduces its surprise factor since LRU has well known deficiencies. This is exacerbated by the fact that only relatively small caches are considered, with associated low hit probability. It is a fact that the authors recognize and is dictated by the limitations of the traces themselves. Finally, one of the main advantages of this new model (over cascade-based requests etc.) is to permit analysis, but that point is not stricto sensu demonstrated here. For those interested, the authors cite a follow up paper, available but not currently published, that focuses on this point.We hope that our community will react to this work with a renewed interest for a more accurate model of caching requests, and that this would eventually narrow the gap between practical design and mathematical analysis for future networks.