| dc.description.abstract | oday’s networks include decentralized intelligence,
scattered over multiple computing facilities, for the execution of
computation-intensive tasks that are necessary for both network
management and service provision to end users. This raises a
number of questions regarding task allocation, in terms of both
the task placement algorithm, which can be either centralized or
distributed, and the resulting task placement in the edge-cloud
continuum. Existing analytical models that tackle these issues
were conceived for network scenarios that include only cloud
computing facilities. In such settings, performance predictions
can be derived assuming that all users are located at practically
indistinguishable (hence irrelevant for the allocation decision)
distances in terms of delay from the cloud server. With the
diffusion of edge servers, the role of path latencies becomes
relevant, and just considering the response times (i.e., waiting
plus service times) at the servers is no longer sufficient. Thus,
we propose a model, called Minimum Delay Independent of
Traffic and service (MIND-IT), for the optimization of task
allocation. MIND-IT considers both the path latencies to reach
the edge/cloud servers and their response times, derives explicit
performance metrics and provides optimization algorithms. We
show, through numerical analysis and real experiments over the
Internet, that differences in path latencies to reach edge/cloud
servers cannot be neglected. By means of algorithmic game
theory, we also study the inefficiency of distributed and selfish
task allocation solutions, unveiling important properties such as
the fact that their performance loss with respect to centralized
optimization is generally limited, except when the system is
overloaded, and that worst-case guarantees can be computed
with low complexity. | es |
