Towards predictable datacenter networks

By Hitesh Ballani, Paolo Costa, Thomas Karagiannis, Antony Ian Taylor Rowstron. Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication, 2011

830 citations recorded by the source corpus.

The shared nature of the network in today's multi-tenant datacenters implies that network performance for tenants can vary significantly. This applies to both production datacenters and cloud environments. Network performance variability hurts application performance which makes tenant costs unpredictable and causes provider revenue loss. Motivated by these factors, this paper makes the case for extending the tenant-provider interface to explicitly account for the network. We argue this can be achieved by providing tenants with a virtual network connecting their compute instances. To this effect, the key contribution of this paper is the design of virtual network abstractions that capture the trade-off between the performance guarantees offered to tenants, their costs and the provider revenue. To illustrate the feasibility of virtual networks, we develop Oktopus, a system that implements the proposed abstractions. Using realistic, large-scale simulations and an Oktopus deployment on a 25-node two-tier testbed, we demonstrate that the use of virtual networks yields significantly better and more predictable tenant performance. Further, using a simple pricing model, we find that the our abstractions can reduce tenant costs by up to 74% while maintaining provider revenue neutrality.

Subfields: Computer Networks and Communications, Information Systems

View Paper: https://www.semanticscholar.org/paper/94ff2db9609acba01d47f89f41b7b2084d04784d

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Networks and Communications

In Information Systems

Citations in the corpus

Efficient Coflow Scheduling Without Prior Knowledge (2015, neutral): “Non-Blocking Fabric In our analysis, we abstract out the entire datacenter fabric as one non-blocking switch [10, 15, 20, 9, 31, 26] and consider machine uplinks and downlinks as the only sources of contention (Figure 1a).”
Achieving high utilization with software-driven WAN (2013, background): “Our current implementation draws on distributed rate limiting [5].”
Resource Management in Clouds: Survey and Research Challenges (2015, background): “Whenever a request for the provisioning of one or more VMs is made by a Cloud User, the Cloud Provider’s resource management system schedules the VMs by placing them onto PMs—the decision should help achieve the Cloud Provider’s current management objective(s).”
Efficient coflow scheduling with Varys (2014, neutral): “Network Sharing Among Tenants Fair sharing of network resources between multiple tenants has received considerable attention [11,33,35,39].”
Coflow: a networking abstraction for cluster applications (2012, neutral): “the network among multiple coflows [7, 23, 24].”
Data Center Network Virtualization: A Survey (2013, methodology): “Oktopus [16] is the implementation of two virtual network abstractions ( virtual cluster and virtual oversubscribed cluster ) for controlling the trade-off between the performance guarantees offered to tenants, their costs, and the provider revenue.”
FairCloud: sharing the network in cloud computing (2012, methodology): “A similar model is assumed by other efforts such as Oktopus [10] or Gatekeeper [24].”
End-to-end Performance Isolation Through Virtual Datacenters (2014, background): “These observations have led to proposals that provide performance isolation across the network [12] [14] [28] [43] [51] [52] [60] [73], storage servers [26] [62] [66] [72], and middleboxes [23].”
Making Sense of Performance in Data Analytics Frameworks (2015, background): “Researchers have used the argument that data-intensive application performance is closely tied to datacenter network performance to justify a wide variety of network optimizations [6,14,17–21,27,28,41,42,48,53].”
Apache Hadoop YARN: yet another resource negotiator (2013, methodology): “9While the RM uses locality as a weight for placing containers, network bandwidth is not explicitly modeled and reserved as in Oktopus [5] or Orchestra [9].”
Performance Isolation and Fairness for Multi-Tenant Cloud Storage (2012, neutral): “The static allocations by SecondNet [13] and Oktopus [6] guarantee bandwidth but can leave the network underutilized.”
this paper (2009, neutral): “For example, cloud gaming services deliver content with a resolution of ≥ 1080p [3] and frame-rate of 60fps [61], while requiring a tail end-to-end delay of less than 100ms [43].”