Current Projects
1. Data Center Network Architecture
The rising emergence of big data and cloud applications leads to the establishment of more and more data centers across the world. A typical data center is consisted of thousands to tens of thousands of servers, which pose a big challenge to the network architecture. To make the data movement efficient, the network topology, routing algorithm, flow control technique and protocol stacks all should be carefully revisited.
Technology Push
High-radix routers can help to reduce the network diameter and the average shortest path length (or the average number of hops). Since the switching latency dominates the zero-load round trip time, reducing the number of hops is the key way to reduce the packet latency. Although the high-radix routers have proved their potential, the way to efficiently exploit them to improve the network performance is still an open problem.
Optical switching, compared with electrical switching, has many advantages that network engineers have dreamed of, such as ultra-high bandwidth (with the help of DWDM technique), super-low latency and power consumption. There’s no such thing as a free lunch. Packet-level optical switching is not available yet, and buffing optical packet is almost impossible. How to leverage the optical switching technique to improve the performance/cost/energy is an open problem for us.
Software Defined Networking (SDN), whose key feature is the separation of control plane from the data plane, is an emerging technology and has proved itself for facilitating the traffic engineering. In this project, we will utilize it to solve the longstanding problems in managing the data center network.
Application Pull
Start Small and Expand Organically is the core business model of cloud computing. To this end, the network architecture should be reconfigurable. First, the topology should be reconfigurable. Once a group of nodes are added, a new topology should be constructed as soon as possible. Second, the routing algorithm also should be reconfigurable to fit the change of network topology. Now, the question is how to expand the data center network without reducing the network performance as well as keeping the cost as low as possible.
Server Consolidation is an important approach to improve system utilization and reduce the cost. Server consolidation often exploits the virtualization technology to provide a virtualized server pool. Thus, multiple tenants could share the physical resources, such as the network bandwidth. Unfortunately, without effective managements, the contention on physical resources may significantly reduce the user experience and even violate the SLA (service level agreement). This project will find the way to efficiently and fairly share the network resources between multiple tenants.
2. Network-on-Chip (NoC) Virtualization
According to Moore’s law, the number of transistors on integrated circuits doubles every two years. Therefore, the number of cores per chip increases continually in recent years. It is expected that thousands of cores will be available in a single chip in the near future. This kind of processors is often referred as man-core processor. Within a many-core processor, the Network-on-Chip (NoC) is responsible for forwarding packets between CPU cores. Unlike the machine-level networks, NoC is often designed as simple as possible to keep it fast and meet the constraints of resources.
Researchers have presented that, within a many-core processor, it is also possible to do the cloud computing. This is the so-called single chip cloud computing (SCC). Unfortunately, according to the Amdahl’s law, it is very difficult for a single application to provide enough parallelism to fully utilize the resources. Therefore, concurrently running multiple applications on a single chip is a promising way to improve the utilization rate of chip resources. However, the NoC virtualization techniques are far from mature.
First, the state-of-the-art NoC techniques can’t fulfill the various requirements of different applications. Different applications may generate quite different traffic patterns. However, an identical virtualized NoC is often provided to all kinds of applications. For example, each application is allocated a 2-D mesh sub-network that utilizes the dimension-order routing (DOR) to route packets. Obviously, this kind of solutions is not efficient. To address this problem, smart NoC virtualization technique is expected to provide application-specific virtualized NoC to each application. Therefore, both of the topology and routing algorithm should be reconfigurable.
Furthermore, it has been proved that the contentions between multiple applications significantly affect the application performance. Particularly, to eliminate the contentions on NoC, state-of-the-art solutions try to isolate the traffics belonging to different applications to different regions. According to our work, however, strict restrictions either reduce network performance or reduce the system utilization rate.
Solving above open problems is the main objective of this project.