1. Content caching and scheduling in wireless networks with elastic and inelastic traffic Group-VI 09CS1048 11CS30003 11CS30020 Performance Modelling in Computer Networks November, 13 2015

2. Contents  Abstract  System Model  Assumptions  Pure Unicast Elastic Scenario  Joint Elastic Inelastic Scenario  Observations  Conclusion  Future explorations

3. Abstract  Growth of wireless content access implies the need for content placement and scheduling at wireless base stations.  Users requests elastic and inelastic content which are stored in request queues at front ends which accumulates the requests by users.  Caches are of finite size and can be periodically refreshed.  Two cost models are considered, file distribution along with requests for streaming stored content and real-time streaming of events.

4. Assumptions  All pieces of content have same size and are called chunks, a chunk is unit of storage and transmission.  There is a number of requests for each elastic content in E at the beginning of each frame in each cluster, this number is assumed as bounded random variable with some mean and variance, and requests are i.i.d over frames.  To ensure requests are served in finite period, unicast transmission in assumed.  Each cache is assumed to have finite and equal capacity of chunks of content, v.  Request arrival and cache reloading is done in same timescale or frame.

5. Pure Unicast Elastic Scenario  In this module, we assume there are only requests for elastic content.  These requests are to be served using unicast communications.  We assume that transmissions are between base stations and front ends, rather than to the actual users making the requests.  We first determine the capacity region, which is the set of all feasible requests.  In this model front ends have independent and distinct channels to the caches, differs from the previously studied wired caching systems because the wireless channels are not always ON.  Therefore, the placement and scheduling must be properly coordinated according to the channel states.

6. Joint Elastic Inelastic Scenario  In this section, we study the general case where elastic and inelastic requests coexist in the system.  The base stations broadcast the inelastic contents to the inelastic users. We assumed servers can employ OFDMA method to simultaneously transmit over their single broadcast and multiple unicast channels.  These two types of traffic do not share the access medium, all the content must share the common space in the caches.  Consequently, we require an algorithm that jointly solves the elastic and inelastic scheduling problems.

7. Observations  If file distribution i.e. streaming of stored content is done then the cost will depend on the frequency with which cache is refreshed.  If streaming is done of content generated in real-time where content expires after a fixed time the cost on the placement of each packet in the cache is considered.  Decision on elastic and inelastic scheduling and allocation of cache spaces to elastic and inelastic traffic is based on the channel states and new request arrivals.  In file distribution case is complex optimization problem and is found using heuristic methods or by simulation.  Generation of content in real time and allocating static cache resources to different types of traffic prior will decompose the optimization problem and make it simple.

8. Conclusions  We incorporated the cost of loading caches in our problem with considering two different models.  In the first model, cost corresponds to refreshing the caches with unit periodicity.  In the second model relating to inelastic caching with expiry, we directly assumed a unit cost for replacing each content after expiration.  A policy was suggested for this model which can stabilize the deficit queues and achieves an average cost which is arbitrarily close to the minimum cost.

9. Future Explorations  The a priori distributions of elastic and inelastic traffics may be used to dynamically allocate spaces to both types of traffic instead of giving static spaces to them, this allocation will simplify the maximization problem as well as it will use the space more optimally.