Professional Documents
Culture Documents
Wireless Facility Scheduling For Data Center Network
Wireless Facility Scheduling For Data Center Network
Wireless Facility Scheduling For Data Center Network
CHAPTER 1
INTRODUCTION
The control packets in the Data Center Networks (DCNs) have to contest with the data
packets although they are usually much shorter in size and much more important in network
management. Moreover, the uneven distribution of the packets may create potential hotspots
in the DCN which could degrade network performance drastically. To bridge these gaps, a
few proposals have been put forward to construct extra wireless facilities in the DCNs to help
promote the performance of the control traffic and relieve the burden of the hotspots.
However, little attention has been paid on how to efficiently schedule the wireless facilities.
1.1 Introduction
Data Center is a pool of resources inter connected using a communication network.
Data Center network holds a important role in a Data Center, as it interconnects all the data
Center resources together. In this paper, a scheduling method is put forward which contains
two steps,
1. Route calculation
2. Flow control
In the former step, a route set between each node pair is calculated in advance for later
usage. Then, arrived data and control flows are scheduled according to multiple policies based
on the given route sets in the flow scheduling step.To meet the increasing bandwidth demand
of cloud-based computation paradigms, many novel network structures, such as FatTree, DCell,
BCube , Jellyfish , PortLand, VL2, FastPass , have been presented to replace or enhance the
traditional tree-based data center networks (DCNs), whose aggregation links are potential
bottlenecks of the network and may hinder the applications’ Quality of Service.
These novel architectures usually provide multiple paths between any two nodes in the
DCN. However, many of them introduce revolutionary changes to the structure like random
cabling in Jellyfish or multi-ports nodes in BCube . These solutions cannot be applied to
existing DCNs since it is very hard, if not impossible, to redesign or reassign the DCNs without
interrupting existing applications For instance, Zhou et al. have designed a flexible wireless
topology using Line-of-Sight (LOS) and indirect line-of-sight (ILOS) wireless links based on
3D beamforming technology [8]. However, how to efficiently schedule the wireless facilities in
the DCN has drawn little attention.
These novel DCNs which consist of both wired links as well as wireless/optics ones are
named as hybrid DCNs These augmented links are usually utilized to offer extra bandwidth to
promote the QoS of the DCN. Recent researches have revealed that HDCN can remarkably
reduce the number of cables and switches in the DCN and thus can reduce not only equipment
costs but also server installation and reconfiguration costs.
CHAPTER 2
RELATED WORK
To enhance the performance of the traditional DCNs, a few proposals have been
proposed based on introducing wireless links or facilities to the data centers.In reference [9],
a wireless facilities network named ANGORA is presented, which can provide robust paths
decoupled from the wired network, and flexibility to adapt to workloads and network
dynamics. In ANGORA, a wired data plane is in charge of switching data packets while the
wireless facilitates deliver control traffic only.
Halperin et al. have explored using 60GHz wireless technology to relieve hotspots in
oversubscribed data center (DC) networks [10]. By experimenting with prototype equipment,
they have shown that the DC environment is well suited to the deployment of 60 GHz links
contrary to concerns about interference and link reliability. The 60GHz radio is named
wireless which is equipped on every rack. The improvement has been testified by analyzing
production traces of DC traffic for four real applications.
Cui et al. have presented a novel DCN architecture, Diamond, which nests the wired
DCN with radios equipped on all servers [13]. To harvest the gain allowed by the rich
reconfigurable wireless resources, they proposed the low-cost deployment of scalable 3D
Ring Reflection Spaces (RRSs) which are interconnected with streamlined wired herringbone
to enable large number of concurrent wireless transmissions through high-performance multi-
reflection of radio signals over metal.
Farrington et al. have presented a generalized hotspot scheduling, called traffic matrix
scheduling, where most or even all bulk traffic is routed over circuits [14]. In other words,
they don’t just hunt elephants, they also hunt mice. Traffic matrix scheduling rapidly time-
shares circuits across many destinations at microsecond time scales. The traffic matrix
scheduling algorithm can route arbitrary traffic patterns and runs in polynomial time.
Besides, there are a few proposals which introduce optical links and switches in to the
DCNs to construct an optical/electricity switch structure. Aktas et al. have put forward
WiCOD, which relies on a wireless control plane serving an all-optical data plane [11]. In
reference [12], Hamedazimi have presented FireFly, an inter-rack network solution using
free-space optics (FSO) that pushes DC network design to the extreme on three key fronts,
(1) All links are reconfigurable.
While beamforming can bound the transmission energy in a “narrow” direction, it still
creates interference at some receivers. Furthermore, ra-dio design artifacts lead to signal leaked
outside of the intended direction [20, 27]. These factors, together with the dense rack deployment,
create harmful interference for concurrent links. For example, consider a link using a typical horn
antenna with 10O beam width and 10dBm transmit power. A single link can interfere1 with up to
27 racks in a typical deployment with 160 racks in a 11m×33m area [17].
Radio interference constrains the number of concur-rent links, and thus also network
throughput. Separat-ing links in the frequency domain limits link capacity. Alternatively,
increasing spacing between racks leads to inefficient space usage.
To address the above limitations, we propose 3D beam-forming, a new beamforming
approach that leverages ceiling reflections to connect racks wirelessly. As shown in Figure 1(d),
each sender points its beam towards a point on the ceiling, which reflects the signal to the desired
receiver. This creates an indirect line-of-sight (LOS) path between the sender and receiver,
bypassing obstacles2 and reducing interference footprint. Next we discuss the feasibility of 3D
beamforming, and show how it addresses the limitations of 2D beamforming
CHAPTER 3
CONCEPTS
3.1 60 GHz Wireless Technology
License-free 60GHz radios can achieve multi-gigabit Radio Frequency (RF) links
using the allocated sufficient spectrum. Moreover, the very narrow beam associated with
60GHz radios enables multiple 60GHz radios to be easily and accurately installed by a
non-expert installer on the same roof top or mast, even if they are all operated at the same
transmitting and receiving frequencies.
Existing designs adopt 60 GHz wire-less technologies because of two reasons. First,
the 7GHz available spectrum offers ample opportunity to achieve multi-Gbps data rates
required by data centers. Second, operating at a high carrier frequency, 60 GHz links
generate limited interference, which is highly beneficial to data centers with dense rack
deployments.
To enhance link rate and further suppress interfer-ence, 60 GHz links use
beamforming, a physi-cal layer technique to concentrate transmission energy in desired
directions. Recent advances in radio design make 60 GHz beamforming radios readily
available and affordable, either as directional (horn) antennas or antenna arrays [4]. They
can adjust beam direction in fine-grain either mechanically or electronically.
However, 60GHz signals can be absorbed by the Oxygen. This attribute limits the
coverage of 60GHz links.In the same time, compared with other wireless technologies, it
also includes some interference and security advantages. In a word, these unique
characteristics make 60GHz wireless technology suitable for short-to-medium distance,
high-bandwidth applications. TP-LINK has unveiled the first 802.11ad wireless router
named Talon AD7200 which works at the 60GHz spectrum at Consumer Electronics
Show (CES) 2016. This router can cumulatively transfer data at the rate of 7133Mbps [16].
In the DCNs, 60GHz radios are usually installed on the top of the racks and
connected to the ToR switches. The quantity needed of radio depends on the size of the
rack. For example, today’s standard rack is 4ftx2ft and a 60GHz radio is 1ftx1ft [8], so at
most 8 radios can sit atop each rack. Because 60GHz links are highly directional, each
rack can only communicate with a small, constant number of peers in parallel.
1. Direct Wireless Link. Two 60GHz radios deployed at neighbor racks can directly
establish a LOS 60GHz wireless link.
2. Raised Wireless Link. A raised radio can establish a wireless link with another raised
radio with the same height.
3. 3D Beamforming Link. Two radios can build a wireless link through the reflector
(mirror) deployed on the ceiling.
Here, all these three link types are adopted. As long as there is no barrier between two
racks and their distance is less than some particular distance, there could be a potential wireless
link between them. Then, a few wireless links can be built between the racks. All the wireless
links make up a Wireless Flyway Network (WFN) in the DCN.
Based on the topology of the DCN, we can calculate a set of feasible routes for each
pair of nodes. Note that, these sets can be obtained in advance since the topology is relatively
static in DCN. After obtaining these sets, we can schedule the given set of flows
{f1,f2,…,fn}.Therefore, with a given DCN topology and a number of flows, the following
step is to calculate the route sets and then decide the routes for each flow.
CHAPTER 4
METHODOLOGY
In the DCN, the topology is more stable than the Internet. Moreover, multiple paths
exist among each node pair. Therefore, we can calculate the routes among each node pair in
advance for the later use. Then, the scheduler can estimate the load of each link based on
ongoing flows on the routes. Finally, the newly arrived flows will be assigned to the routes
based on the traffic estimation results.
They treated the topology as a graph G(V, E), where V and E are the sets of links and
edges respectively. The problem is to decide the routes between any two nodes i and j. We
start from node i, and apply the width first search until node j is found in the tree or the depth
of the searching tree reaches k, which is the maximum length of a route. A tabu list is kept
during the searching process to cut those tree branches encountered before or interacting with
existing routes. The algorithm of this searching process for a pair of nodes i and j is shown in
Algorithm 1.
1. Wired-only. Only the wired networks are utilized for carrying the data and control flows.
2. Wireless-for-control. The wireless links are utilized for carrying control flows while the data
flows are carried on the wired network.
3. Partial-for-data-with-priority. The wireless links are utilized for carrying control flows
and it can also carry a few offloaded data flows if idle bandwidth is available. Here, if a
data flow is partially offloaded to the WFN, we move 10% of its total packets to the WFN
and remaining 90% packets will still be transmitted through the wired network. Moreover,
the control flow’s priority is higher than the offloaded data flows. The offloaded data
flows will be transferred back to the wired network if the control flows need to use its
occupied wireless capacity.
4. Failure-tolerant. The wireless links are utilized for carrying control flows and it carries
the flows for those ToRs whose links to the aggregation switches are failure.
CHAPTER 5
SIMULATION
To evaluate the performance of the WFN construction and its scheduling methods
proposed in this paper, a series of simulation experiments are conducted based on extending
the flyway simulator .
tolerant scenario, two wire links are randomly removed from the DCN.
Radio interference constrains the number of concur-rent links, and thus also network
throughput. Separat-ing links in the frequency domain limits link capacity. Alternatively,
increasing spacing between racks leads to inefficient space usage.
Average one-way delay: It is defined as the average value of a flow’s all packets’ one-
way delays. A packet’s one-way delay is defined as its receiving time at the destination
minus its sending time at the source.
Rack-based: This is the topology used in [15]. Racks are grouped into 4×4 clusters, and
each cluster is a row of 10 racks with no inter-spacing. Racks are of 0.6m×1.2m in size
and aisles separating the clusters are 3m (between columns) and 2.4m (between rows)
wide. In total, the data center is 11m×33m and contains 160 racks.
Container-based: Offered by Sun Microsystems [5], this data center consists of 2×2
container clusters. Each cluster has 8 containers in a row with an inter-spacing of 0.61m.
Each container is of 6.1m×2.4m in size, containing 8 racks laid out in two rows.
Container clusters are separated by 3.7m (between columns) and 6.7m (between rows).
The entire data center is 15m×50m and contains 256 racks.
Assign one radio transceiver per rack and allow each transceiver to associate with one link.
We build bi-directional links by randomly selecting transceiver pairs, forming arbitrary rack
communication. Each receiver antenna also points to its sender. The maximum number of links for
the two topologies are N = 80 and 128, respectively. For each group of N transceiver pairs, we
determine the number of concurrent links as follows. We admit links one by one in a random order,
compute their cumulative interference to each other, and only admit a link if all links after
admission achieve a minimal rate of 5.53Gbps in the presence of interference. We consider both
cases where all the links operate on a single 2.16GHz channel, and cases where three 2.16GHz
channels (for the US 60 GHz band) are available.
CHAPTER 6
RESULTS
There is a LOS path between i and j and the length of this path is less than 10 meters.
Based on this setting, Figure 2 demonstrates the wireless relationship among the racks. Each
rectangle represents a rack. The dots on the racks are the wireless radios and the dotted line
between two radios represents a wireless link.
The CTF of the data and the control flows are shown in Figure 3 and Figure 4 respectively. In
Figure 3, each data flow is indicated by the number of the data flow plus a
character ‘D’. For example, 1D is the 1th data flow. Analogously, 1C is the 1th control flow.
From Figure 3 and Figure 4, we can draw the following observation conclusions:
1) Generally speaking, Wired-only scenario has the worst performance among the four
considered scenarios. This is due to the fact that the latter three scenarios can potentially
provide more bandwidths for the data flows as well as the control flows. Note that sometimes,
Partial-for-data-with-priority cannot ensure a better performance for the offloaded data flows
due to the handover time introduced when the packets are offloaded from the wired network
to the wireless links. Moreover, if the length of two wireless antennas is large, the
transmitting delay of the wireless path is non-neglected for the CTF of a flow.
2) Through efficiently scheduling the wireless links, we can promote the performance of
the control flows. This is very critical to ensure the quality of the time-sensitive control flows.
Moreover, timeless delivery of the control flows is also very helpful for the management of
the network.
3) In the Failure-tolerant scenario, we randomly make two wired links fail. Here, the
selected fail links influence the 18th and the 25th data flows since their ToRs’ aggregator links
to the aggregator switches fail. Therefore, the 18th and the 25th data flows’ CTFs will be
infinite in the former four scenarios. Under this condition, Failure-tolerant scenario can
deliver their packets through the WFN. This is true for the 13th control flow since it is
influenced by the fail links too. This is critical for ensuring the continuity of the DC service.
CHAPTER 7
CONCLUSION
Recent evidences have shown that it is feasible to introduce wireless links into the
data center networks (DCNs) to relieve the hotspot wired links as well as to promote the
performance of the applications. However, wireless link scheduling has drawn little attention.
In this paper, we have presented a wireless links or facilities scheduling method to efficiently
schedule the wireless resources in the DC. It mainly contains three steps, i.e. routes
calculation, traffic estimation and flow scheduling. Routes calculation is in charge of
computing the routes between each node pair in advance since the topology of the DCN is
stable. Then, the traffic on each link is estimated by a scheduler based on the historical
assignment results. In the last step, the newly arrived flows are assigned to the routes
obtained in the first step based on four different kinds of scheduling policies.