In this chapter we provided an overview on cloud datacenter, especially topology design and traffic management. We described multipath routing technique in cloud datacenter and also gave a briefly summary of congestion voidance control in network. We note that there are many variants of TCP. The main variants of TCP in use today are favours of TCP Reno and TCP Sack. For high speed networks, many other TCP variants proposed, such as AIMD, STCP, HSTCP. The main difference between these TCP variants lies in the manner in which they deal with lost packet recovery. However, there’s no TCP variant pays attention to the congestion problem in cloud datacenters. Hence, there is a need to study how to enhance the TCP performance in high speed and low latency cloud datacenter networks. In this work, we propose a new TCP protocol, namely Adaptive Bandwidth Congestion Control (ABCC) which modifies the TCP behavior with more aggressive bandwidth probing and better utilization. A brief introduction to ABCC is given in the chapter 4.
Table 2.1 will explain what we do in the rest of thesis.
Table 2.1: Literature Survey
Topology Multipath TCP variants Simulation
VL2 Fat tree o x o
PortLand Clos o x o
Experimental Multi-root x x x
Our work Fat tree, Inventec o o o
Table 2.2: Comparisions of TCP Variants and our proposed
Types Slow Start Phase CA without Loss CA with Loss
TCP cwnd = cwnd + 1 cwnd = cwnd + cwnd1 cwnd = cwnd 1 −12
AIMD cwnd = cwnd + 32 cwnd = cwnd + cwnd1 cwnd = cwnd 1 −12
STCP cwnd = cwnd + cwnd × 0.01 cwnd = cwnd + cwnd1 cwnd = cwnd 1 −18
HSTCP cwnd = cwnd + inc(cwnd) cwnd = cwnd + cwnd1 cwnd = cwnd + dec(cwnd)
Our work Bandwidth Adaptation Update Sending Rate cwnd = cwnd 1 − 1δ
Chapter 3
System Model and Problem Formulation
3.1 System Model
3.1.1 Model
The network model under consideration in this research is multi-root tree network topology in cloud datacenters. We focus on the simulation of proposed mechanism in two models: a m-port n-tree fat-tree topology and an experimental datacenter. The two models belong to typical cloud datacenter architectures that are using widely in many datacenters nowadays.
They commonly consist of three-level trees of switches or routers. A three-tiered design has a core tier in the root of the tree, an aggregation tier in the middle and an edge tier at the third level of the tree. In the edge level or aggregation level, every device has two outlets for the connection of high level devices.
3.1.1.1 Model 1: Fattree topology
There are many ways to construct the fat-tree network topology. The root switches in traditional k-ary n-trees only use half of their communication ports. In order to utilize every port of root switches, we use an m-port n-tree approach [22] to construct the cloud datacenter. In this construction, the number of ports in each switch is denoted by m, and n is signed for the number of tiers in network. An m-port n-tree consists of 2 × (m/2)n hosts
Figure 3.1: 4-port 3-tree Fat tree topology with 16 hosts and 20 switches.
and (m/2)n−1+ mn−1 switches. Each switch has m communication ports 1, 2, 3,..., m that are attached to other switches or hosts. There are (m/2)n−1 root switches. Every switch other than root switches uses port 1, 2,..., m/2 connecting with its descendants or hosts, and uses port m/2+1, m/2+2,..., m connecting with its ancestors. Hosts are connected at leaf switches. In this study, we evaluate the proposed mechanism in two m-port n-tree fat-tree topologies, a 4-port 3-fat-tree constructed for small scale and 8-port 3-fat-tree one for medium scale cloud datacenter. The large-scale cloud datacenter is considered in the experimental topology. Small-scale datacenter consists of 20 1Gb switches with 4 ports, where 4, 8 and 8 switches are in core, aggregate and edge tiers, respectively. This network serves for 16 hosts.
Medium-scale datacenter consists of 80 8-ports 1Gb switches with 16 in core, 32 in aggregate and 32 in edge tier, respectively. This network serves for 128 hosts.
Figure 3.2: 8-port 3-tree Fat tree topology with 128 hosts and 80 switches.
3.1.1.2 Model 2: Experimental topology
A real datacenter designed by Inventec Inc. has 65 switches. There are 5 switches (10G, 48 ports) in core tier, each switch connects to 10 switches in Aggregate (Region) tier and remains 10 links to 10 switches in region tier. Under the core tier, the network is divided into two regions. Each region has 5 switches (10G, 48 ports) which aggregates 5 racks consist of 5 switches and 1 data node. Four switches in aggregate tier remain fully down-links to each rack; there are always 2 links from 2 switches in aggregate tier to data node in each rack.
Each switch also connects to 10 data nodes. Each switch in region tier always has 5 up-links to connect to core tier.
3.1.2 Assumptions
The multi-root cloud datacenter network is a packet-switching based network. Routing in a subnet is based on the forwarding table stored in each switch. When a packet arrives in a switch, the packet will be forwarded to the corresponding output port via the forwarding table lookup. The table look up will find the destination in the shortest path condition or send same content in the shortest path and one or several backup paths. When multiple hosts want to send packets to the multiple other hosts at the same time, the link congestion may occur while the bandwidth of multiple paths offered by the multi-root topology is wasted.
We assume applications that generate a series of relatively short data transfers for that reason, agile rate adaptation is very important to improve communication efficiency.
Figure 3.3: An experimental cloud datacenter with 65 switches and 1000 hosts.
Figure 3.4: Assumption model.
Traffic will be created to fully load the network under all-to-all traffic with large and small distributions so that we can observe the congestion phenomenon.