Digging deeper into content to improve caching efficiency

Speaker:        Dr. Lazaros Gkatzikis
                Huawei France Research Center
                Paris, France

Title:          "Digging deeper into content to improve caching efficiency"

Date:           Thursday, 30 June 2016

Time:           4:00pm - 5:00pm

Venue:          Room 3501 (near lifts 25/26), HKUST


Abstract:

Content caching at the edge is promising for the sustainability and
performance of future wireless networks. Access latency and core network
traffic can be reduced by bringing content closer to the user. However,
caching in the access network suffers from two unavoidable technical
problems: 1) Due to the large number of caches to be deployed, each
individual cache has to be small (in comparison to the file catalogue),
and hence only a negligible fraction of the catalog can be stored in a
single cache. 2) The user population accessing a given cache at the access
level is typically small compared to caching at the CDN level, and hence
estimating the popularity of dynamic content in time is challenging.

In this talk, we present two methods based on content chunking to improve
caching efficiency. First, we focus on video content which represents a
significant portion of the network traffic. Rarely do users watch online
contents entirely. Several video delivery platforms, such as YouTube,
collect statistics on the user engagement performance of their videos,
usually called "Audience retention rate". We demonstrate how this
information can be used to achieve a significant traffic reduction on the
core network. We characterize the performance upper bound of a cache able
to store parts of videos and the best performance achievable by chunk-LRU.
Next, we address the problem small sample sizes focusing on L local caches
and one global cache. On one hand we show that the global cache learns L
times faster by aggregating all requests from local caches, which improves
hit rates. On the other hand, assuming locality of interest, aggregation
washes out local characteristics which leads to a hit rate penalty. This
motivates coordination mechanisms that combine global learning of
popularity in clusters and LRU with prefetching.


*******************
Biography:

Lazaros Gkatzikis (S'09-M'13) received the Ph.D. degree in computer
engineering and communications at the University of Thessaly, Volos,
Greece. He is a Research Staff Member with the Huawei France Research
Center, Paris, France. In the fall of 2011, he was a Research Intern with
the Technicolor Paris Research Laboratory. He was a Postdoctoral
Researcher with the University of Thessaly, Volos, Greece (2013) and the
KTH RoyalInstitute of Technology, Stockholm, Sweden (2014). His research
interests include network optimization, game theory, and performance
analysis.