鵝
歷
☌
←
→
⇡
07 Jan 2016

SoftCell: Scalable and Flexible Cellular Core Network Architecture

https://www.cs.princeton.edu/~jrex/papers/softcell13.pdf

ABSTRACT — SoftCell, a scalable architecture that supports fine-grained policies for mobile devices in cellular core networks, using commodity switches and servers… realize high-level service policies that direct traffic through sequences of middleboxes based on subscriber attributes and applications.

improves the scalability and flexibility of cellular core networks

Unlike traditional IP networks, cellular providers rely extensively on customized policies based on a wide variety of subscriber attributes and application types.

Typical subscriber attributes
the cell-phone model, the operating-system version, the billing plan, the options for parental controls
whether the total traffic exceeds a usage cap, or whether a user is roaming
Typical application types
video traffic (for transcoding), web traffic (for caching)
specific applications for which the developers pay the carrier on the user’s behalf [3] (for exempting that traffic from the user’s cap).

… fine-grained packet processing, cellular providers rely on specialized and proprietary devices — Serving Gateways (S-GWs) and Packet data network Gateways (P-GWs)

Centralizing nearly all data-plane functionalities in the P-GWs makes cellular core networks remarkably inefficient, complex, and inflexible [5, 3] for at least three reasons ... carriers cannot “mix and match” capabilities from different vendors (e.g., use a firewall from one vendor, and a transcoder from another), or “scale up” the resources devoted to a specific function [3, 6].

cellular networks are defined along many dimensions, leading to large number of packet classifiers.

Muti-dimensional aggregation
SoftCell significantly reduces the size of switch tables by aggregating entries along multiple dimensions, combining the benefits of traditional location-based routing and tag-based routing.

We built a SoftCell controller on top of Floodlight [13]. We then evaluated it using: i) real traces from a large LTE deployment; ii) micro-benchmarks on our prototype; and iii) large-scale simulation experiments. Our experiments and analysis show that SoftCell can easily handle the workload of a large LTE network and support thousands of service policy clauses with just a few thousand TCAM entries in the core switches.

2. SOFTCELL ARCHITECTURE

SoftCell’s goal is to support numerous fine-grained policies in a scalable manner for cellular core networks.

2.1 SoftCell Core Network Components

Controller
The controller implements high-level service policies by installing switch-level rules that direct traffic through middleboxes.
Middleboxes
SoftCell supports commodity middleboxes such as dedicated appliances, virtual machines, or packet-processing rules on switches.

2.3 SoftCell Design Challenges and Solutions

6. PERFORMANCE EVALUATION

6.1 LTE Workload Characteristics

As a first step towards SoftCell deployment, we measured the workload of a real cellular network to understand the performance requirements of the controller. In contrast to other LTE measurement works, we study the aggregate arrival rates of UEs and flows and show the implications on control-plane load.

Dataset description
We collected about 1TB traces from a large ISP’s LTE network

The dataset covers a large metropolitan area with roughly 1500 base stations and 1 million mobile devices (including mobile phones and tablets).

The trace includes various events such as radio bearer creation, UE arrival to the network, UE handoff between base stations, etc
QoS class, the flow will use the existing radio bearer.
to estimate flow activities…
Network-wide characteristics
As each of these events require the central controller to contact local agents (send packet classifiers) or update core switches (install short-cuts for long flows), it implies that the controller should be able to handle hundreds of such events per seconds …
Load on each base station

the CDF of active UEs per base station. We see that a typical base station handles hundreds of active UEs with a 99.999 percentile of 514.

flow arrival rate to be around several hundred per second

6.2 Controller Micro Benchmark

Cbench emulates a number of switches, generates packet-in events to the tested controller, and counts how many events the controller processes per second (throughput).

Central controller performance
throughput of the controller.
emulates 1000 switches and let these switches keep sending packet-in events to the controller

the controller can process 2.2 million of requests per second with 15 threads

Local agent performance
the local agent throughput

A cache hit ratio of 80% means that the local agent can handle 80% of the flows locally and needs to contact the central controller for the remaining 20%

7. DISCUSSION

Cellular network architecture
Several recent works have exposed the complexity and inflexibility of cellular networks [5, 3] and several research efforts [33, 31, 34, 35, 36] have aimed at fixing the problems. CellSDN [33] presents the first high-level design of software-defined cellular networks that SoftCell fully develops. [31] proposes to integrate the support of GTP tunnels within OpenFlow.
Software defined networks
DevoFlow [17] and DIFANE [37] improve upon Ethane [7] by moving some processing from the control plane to the data plane. However, their techniques cannot address specific cellular network requirements like fine-grained policies, or policy consistency under mobility.
“smart access edge, dumb gateway edge” to scale the system
a specific solution in the core to support fine-grained service routing, which is not addressed in Fabric and SDIA

reference