Application-aware network sharing for enabling high progress of multi-tenants

Abstract

In multi-tenant shared clouds, applications of different tenants compete for the shared network and thus suffer significant unpredictability of performance. The progress is an essential metric to measure the overall data transfer rate of a tenant and to indicate how fast a tenant can complete data transfer. Most previous work of network allocation or schedule focus on achieving the tradeoff among the fairness, performance, and efficiency. They, however, often do not make each tenant achieve noticeable progress. In this paper, we motivate to maximize tenant's progress by using information about the underlying network and traffic patterns of tenants' applications. To illustrate the feasibility of maximizing progress, we develop a network sharing framework for multi-tenant datacenter, by exploiting the benefits of careful placement of tenant's applications. Furthermore, the tenants' progress and network utilization would be further improved without impacting performance isolation, if the bandwidth demand of each application can be rescheduled. Our evaluation results show that our method provides higher tenant progress and better tenant-level performance than HUG.