Towards bandwidth guaranteed energy efficient data center networking
 Ting Wang^{1}Email author,
 Bo Qin^{1},
 Zhiyang Su^{1},
 Yu Xia^{1},
 Mounir Hamdi^{1, 2},
 Sebti Foufou^{3} and
 Ridha Hamila^{3}
https://doi.org/10.1186/s1367701500357
© Wang et al.; licensee Springer. 2015
Received: 18 November 2014
Accepted: 23 April 2015
Published: 10 May 2015
Abstract
The data center network connecting the servers in a data center plays a crucial role in orchestrating the infrastructure to deliver peak performance to users. In order to meet high performance and reliability requirements, the data center network is usually constructed of a massive number of network devices and links to achieve 1:1 oversubscription for peak workload. However, traffic rarely ever hits the peak capacity in practice and the links are underutilized most of the time, which results in an enormous waste of energy. Therefore, aiming to achieve an energy proportional data center network without compromising throughput and fault tolerance too much, in this paper we propose two efficient schemes from the perspective of resource allocation, routing and flow scheduling. We mathematically formulate the energy optimization problem as a multicommodity minimum cost flow problem, and prove its NPhardness. Then we propose a heuristic solution with high computational efficiency by applying an AI resource abstraction technique. Additionally, we design a practical topologybased solution with the benefit of Random Packet Spraying consistent with multipath routing protocols. Both simulations and theoretical analysis have been conducted to demonstrate the feasibility and convincing performance of our frameworks.
Keywords
Introduction
The data center, as a centralized repository clustering a large number of servers, has become home to essential largescale computation, storage and Internetbased applications which provide various services like search, social networking, emails, gaming, cloud computing, and so on [1,2]. In order to provide high performance service with strong reliability to users, the data center network (DCN) architectures are usually overprovisioned and constructed aggressively with large number of switches and links to achieve highcapacity and high fault tolerance [3]. However, the research [4,5] shows that, in practice the average link utilization in different data centers ranges only between 5% and 25% and varies largely between daytime and night. This reveals that most network devices and links stay idle or underutilized most of the time, but an idle device consumes up to 90% of the power consumed at full loads [6], which leads to a great waste of energy. Apart from the energy wasted due to overrichly network interconnections, traditional nonenergyaware routing algorithms (like shortest path routing or its variations) can also lead to poor link utilization or even congestion, which worsens the situation.
According to current research findings [7,8], the power consumed by servers and infrastructure (i.e. power distribution and cooling) accounts for over 70% of overall power, while the network consumes around 15% of the total power budget. However, as the servers become more energy proportional, the fraction of the power consumed by the network in a data center grows correspondingly higher. As illustrated in [4], suppose the servers are totally energyproportional, when the data center is 15% utilized (servers and network), then the network will consume up to 50% of overall power. Even if the servers are not energyproportional, with 15% traffic load, making the network proportional still can save as much as 975 KW (for a data center with 10,000 servers) [4]. Unfortunately, today’s commodity network devices are not energy proportional, mainly because the components of the network devices (such as transceivers, line cards, fans, etc) are always kept on regardless of whether they have data packets to transfer or not, leading to a significant energy wastage.
Based on the above observations, this paper aims to achieve a bandwidth guaranteed energy proportional data center network, where the amount of power consumed by the network is proportional to the actual traffic workload. The key principle behind this approach is that most of time the traffic can be merged and satisfied by just a certain subset of network devices and links, and the remaining ones can be put onto sleep mode or powered off for the sake of power conservation. With this goal, we propose two efficient green frameworks from the perspective of bandwidth allocation and flow scheduling. However, the bandwidth in a data center is a scarce resource [9,10], and the energyaware routing problem is NPhard, which is proved in Section “Problem statement”. Besides, the time complexity of dynamically computing a feasible network subset to meet the traffic demands is horrible and unmanageable due to the exponential number of nodes and routes. In order to address these critical issues, derived from Artificial Intelligence our first framework employs a resource abstraction technique named Blocking Island (BI) and some well designed heuristic algorithms to efficiently reduce the searching space and significantly decreases the computation and time complexity. This framework can be applied in any arbitrary data center network topology. In the second framework, we put forward a topologybased energyaware algorithm by computing a network subset and adopting one recently proposed multipath routing mechanism RPS [11] for flow scheduling and packet transmission.
 1.)
We formulate the energy optimization problem in DCNs mathematically and prove its NPhardness.
 2.)
We propose two efficient general frameworks, which provide efficient solutions from the perspective of bandwidth allocation, routing and flow scheduling.
 3.)
To the best of our knowledge, we are the first to employ the AI model  Blocking Island Paradigm into data centers for resource allocation to achieve power savings.
 4.)
We conduct extensive simulations to evaluate and demonstrate the performance of our frameworks under various network conditions and reliability requirements.
The rest of the paper is organized as follows. First we review the related research literature in Section “Related work”. Then we formulate the energy optimization problem and prove its NPhardness in Section “Problem statement”. Afterwards, Blocking Island Paradigm is briefly reviewed in Section “Blocking island paradigm”. Then we propose two energyaware heuristic schemes in Section “Energyaware heuristic schemes”, followed by the evaluations and simulation results in Section “System evaluation”. Finally, Section “Conclusion” concludes the paper.
Related work
A considerable amount of investigation and research for achieving a green data center have been conducted in both academia and industry due to its great potential and benefits. Apart from the works on green/renewable resources [1214], lowpower hardware [1518], energyefficient network architecture [4,1923], and network virtualization techniques (VM migration and placement optimization) [24,25], there are also many networklevel proposals, which focus on traffic consolidation. The typical representatives include ElasticTree [5], and Energyaware Routing Model [26].
ElasticTree is a networkwide energy optimizer, which consists of three logical modules – Optimizer, Routing, and Power Control. Once the Optimizer outputs a set of active components, Power Control toggles the power states of ports, linecards, and entire switches, while Routing chooses paths for all flows, then pushes routes into the network. The authors proposed three types of optimizers with different quality of solution and scalabilities. The formal method achieves the best solution but is computationally very expensive and does not scale beyond a 1000node sized data center. The greedy binpacker ends up with suboptimal solutions as with any greedy approach but is much faster than the formal method. Lastly, the topology aware heuristic needs the smallest amount of computation time but the quality of its solution is inferior to both the greedy and formal method.
Energyaware Routing Model is also a networkwide approach, which aims to compute the routing for a given traffic matrix, so that as few switches are involved as possible to meet a predefined performance (throughput) threshold. The basic idea is that: Firstly, they take all switches into consideration and compute basic routing and basic throughput. Then, they gradually eliminate the switches from basic routing and recompute routing and throughput until the throughput reaches the predefined threshold. Finally, they power off the switches not involved in the routing. However, this approach suffers inefficient computation efficiency, where it takes several seconds to calculate a nonoptimal poweraware routing paths for thousands of flows and takes even hours to calculate a near optimal solution, which is intolerable for a latencysensitive data center network.
Problem statement
MCF problem description
The multicommodity flow (MCF) problem is a network flow problem, which aims to find a feasible assignment solution for a set of flow demands between different source and destination nodes. The MCF problem can be expressed as a linear programming problem by satisfying a series of constraints: capacity constraints, flow conservation, and demand satisfaction. This problem occurs in many contexts where multiple commodities (e.g. flow demands) share the same resources, such as transportation problems, bandwidth allocation problems, and flow scheduling problems. In the next subsection, we show that the energyaware routing problem can also be formulated as an MCF problem.
Problem formulation
 1.
Demand completion—each traffic demand specified as a tuple (i,j,d _{ ij }) should be satisfied with the required bandwidth simultaneously, with i,j,d _{ ij } (i,j∈V) as the source node, destination node and bandwidth request, respectively (i.e., Constraint (1));
 2.
Reliability requirement—each demand should be assigned FT number of backup routes (i.e., Constraint (2));
 3.
Capacity constraint—each link k∈E has a bandwidth capacity C _{ k } and none of the traffic demands ever exceed the link capacities (i.e., Constraint (3));
 4.
Flow conservation (i.e., Constraint (4)).
Note that if we assume the optimal rounting paths are linkdisjoint, we can simplify Constraint (2) as \(\forall i,j\in V, \sum _{k\in N_{i}}Y_{\textit {ji}}^{(k)}\geq FT, \sum _{k\in N_{i}}Y_{\textit {ij}}^{(k)}\geq FT\) with \(Y_{\textit {ji}}^{(k)}\geq x_{\textit {ij}}^{(k)}/C_{k}\) and \(Y_{\textit {ji}}^{(k)}\in \{0,1\}\).
NPhardness
For the MCF problem described above, we change to its corresponding decision problem (DMCF): Is there any set of rounting paths such that satisfy \(\Omega _{s} \sum _{i\in V}S_{i}+2\Omega _{p}\sum _{k \in E}L_{k}\leq N\), and all constrains in MCF. To prove the DMCF problem is NPhard, we show the classical 01 knapsack problem [27] can be reduced to a DMCF instance. Thus, both DMCF and MCF are NPhard due to the equivalence of hardness.
The formal definition of the 01 knapsack problem is given as below. There are n kinds of items I _{1},I _{2},...,I _{ n }, where each item I _{ i } has a nonnegative weight W _{ i } and a nonnegative value V _{ i }, and a bag with the maximum capacity as C. The 01 knapsack problem determines whether there exists a subset of items S (S⊆[n]) such that \(\sum _{i\in S}W_{i}\leq C\) and \(\sum _{i\in S}V_{i}\geq P\).
Proof.
Reduction: We first construct a specific instance G of the DMCF problem. Suppose there exists a source s and a sink t in G, and only one demand (s,t,d _{ st }=P). For each item I _{ i } in the knapsack problem, we build a path p _{ i } with W _{ i } links from s to t in G, and each link k in p _{ i } has capacity of C _{ k }=V _{ i }/α. The parameters are set as Ω _{ p }=1, Ω _{ s }=0, F T=1, and the predefined threshold of DMCF is set as N=2C.
(i) The solution for the 01 knapsack problem exists ⇒ The solution for the specific DMCF instance exists. Suppose there exists a subset of items S such that \(\sum _{i\in S}W_{i}\leq C\) and \(\sum _{i\in S}V_{i}\geq P\). Then, we can use S to construct a solution for the specific DMCF instance. For each item I _{ i } (i∈S), we choose the corresponding path p _{ i } in G, and assign a flow of size V _{ i } to this path, i.e., \(x_{\textit {st}}^{(k)}=V_{i}\) for all links in p _{ i }. Thus, the capacity constraint (3) holds since \(x_{\textit {st}}^{(k)}=V_{i}\geq \alpha C_{k}=V_{i}\), the flow conservation (4) holds naturally, and then the demand completion (1) is satisfied since \(\sum _{k\in N_{t}}x_{\textit {st}}^{(k)}=\sum _{k\in N_{s}}x_{\textit {st}}^{(k)}=\sum _{i\in S}V_{i}\geq P= d_{\textit {st}}\), and hence the reliability requirement (2) is met due to F T=1. Constraint (5) means we will choose all W _{ i } links in the path p _{ i }, and then the total number of chosen links is \(\sum _{i\in S}W_{i}\), leading to the value of the objective function \(2\Omega _{p}\sum _{k \in E}L_{k}=2\sum _{i\in S}W_{i}\leq 2C=N\). Therefore, the found solution is indeed a solution for the specific DMCF instance.
(ii) The solution for the specific DMCF instance exists ⇒ The solution for the 01 knapsack problem exists. Suppose there exists a set of S _{ i }’s and L _{ k }’s satisfying all constraint in the specific DMCF instance and \(2\Omega _{p}\sum _{k \in E}L_{k}\leq N\). If a link k (k∈N _{ t }) in the path p _{ i } has L _{ k }>0, then \(x_{\textit {st}}^{(k)}>0\) by Constraint (5) and \(x_{\textit {st}}^{(k)}\leq \alpha C_{i}=V_{i}\) by Constraint (3). For such a p _{ i }, we choose the corresponding item i in the 01 knapsack problem and form a subset of item S. Then, \(\sum _{i\in S}V_{i}\geq \sum _{k\in N_{t}}x_{\textit {st}}^{(k)}\geq d_{\textit {st}}=P\) due to Constraint (1). On the other hand, since \(x_{\textit {st}}^{(k)}>0\ (k\in N_{t})\) in p _{ i }, the flow values of all links in p _{ i } is equal to \(x_{\textit {st}}^{(k)}>0\) due to the flow conservation. This means all W _{ i } links in p _{ i } have L _{ k }=1 by Constraints (5). Then, the total number of chosen links is \(\sum _{i\in S}W_{i}=\sum _{k \in E}L_{k}\leq N/2\Omega _{p}=C\). Thus, we find the solution for the 01 knapsack problem. That ends the proof.
Blocking island paradigm

Unicity: Each node has one unique βBI. If S is the βBI for node x, then S is the βBI for all the nodes in S.

Route Existence: An unallocated demand d _{ u } = (x, y, β _{ u }) can be satisfied with at least one route if and only if both the endpoints x and y are in the same β _{ u }BI.

Route Location: The links of a route with β available bandwidth are all in the βBI of its endpoints.

Inclusion: If β _{ i } is larger than β _{ j }, then the β _{ i }BI for a node is a subset of β _{ j }BI for the same node.
Energyaware heuristic schemes
In this section, we propose two heuristic solutions to the energy optimization problem formulated in Section “Problem statement”, one of which is based on Blocking Island Paradigm while the other one is topology based. The BI based heuristic achieves bandwidth guaranteed green networks and enjoys low computation complexity with the help of Blocking Island Paradigm for resource allocation. The topology based heuristic holds the best scalability (O(N)) in computation time growth where N is the number of servers, but the resulting solution is not as good as the BI based solution. Comparatively, the BI based heuristic provides a more attractive and practical option.
Power conservation strategy
Most existing proposals apply devicelevel power conservation strategy, which intends to switch off the entire device (router/switch) including fixed overheads (like fans, linecards, transceivers, etc.) only when all ports on the device are idle. This means even if only one port has data to transfer, the device (including idle ports) should be kept alive all the time. Comparatively, in our energyaware heuristic schemes we apply the componentlevel strategy, which intends to power down the unused ports, and switch off the linecard if all the ports on this linecard are idle or disabled. If all linecards on a device are idle then to power off the entire device. Clearly, the componentlevel strategy achieves the most power savings.
where Ω _{ s } and Ω _{ p } are the same as described in Section “Problem statement”, and N _{ p } denotes the number of active ports on the switch.
BIbased heuristic scheme

Drawing support from BI model to guide the bandwidth allocation for the traffic demands in the most advantageous order.

Using the energyaware routing algorithm to compute the most beneficial routes for these allocated demands.

Switching off devices that are not involved in the final routings for power conservation.
Bandwidth allocation mechanism
In line with the data center policy which requires fast response to the request, the Route Existence property of BI enables much faster decisions in determining whether a request can be satisfied just by checking whether both the endpoints are in the same βBI, while traditional routing algorithms have to compute the routes before deciding the route’s existence. For example, if we want to assign a path for a traffic demand (H1, H3, 50) in the network as shown in Figure 1, then we can immediately know that the route does not exist since H1 and H3 are not in the same 50BI without any effort to search the whole network space and compute the routes. Moreover, if we need to find a path for (H1, H2, 50), then the search space can be reduced from the whole network to only {S1, S2, S3, S4}, which leads to a significant improvement in the efficiency of computation and bandwidth allocation. The unique βBI for a given node x can be obtained by a simple greedy algorithm (as depicted in Algorithm 1) whose complexity is linear in O(L), where L denotes the number of links. Additionally, querying two nodes in the same BI experiences a complexity of just O(1) since only two hashing operations are needed.
As a known NPhard MCF problem, it cannot be guaranteed to find an assignment to satisfy all the flows for all kinds of traffic matrix all the time. How to select the next demand to allocate bandwidth has a great impact on the future allocation success ratio, and also affects the search efficiency. There are some static methods for addressing this kind of MCF problem or constraint satisfaction problem (CSP) [29], such as firstfail principle based technique, or first selecting the largest demand. However, these static techniques are not suitable to be directly applied in the data center network which requires a more dynamic and efficient bandwidth allocation mechanism. In addition, considering the data center’s own particular characteristics, these traditional static techniques do not take the mean flow completion time and deadline into consideration as well. In our approach, these concerns are effectively resolved by exploiting the advantages of Blocking Island Paradigm.
 (i)
It firstly choose the demand of which the lowest common father (LCF) of the demand’s endpoints in the BIH tree is highest. The intuition behind this principle is to first allocate the more constrained demands, which follows the failfirst principle.
 (ii)
If there are multiple candidate demands after the first step, then the Shortest Demand First (SDF) principle is applied, which aims to meet as many deadlines as possible and reduce the mean flow completion time. The shortest demand indicates the demand whose expected completion time is minimum \(\left (\text {i.e.}~ min\left \{\frac {flow ~size}{required ~bandwidth}\right \}\right)\) where the f l o w s i z e and r e q u i r e d b a n d w i d t h are provided by the application layer [30,31].
 (iii)
In case there are still two or more satisfied demands, then the demand with the highest bandwidth requirement is preferentially selected. This criterion, apart from implying a near deadline flow, also follows the failfirst principle, where more bandwidth allocation more likely cause BI splittings and thus hinder any future allocation of other demands.
 (iv)
Finally, we randomly select one demand from the output of step (iii).
The demand selection rules for bandwidth allocation not only decreases the computation complexity and increases the search efficiency, but also takes the flow deadline into consideration. Moreover, they can also increase the success ratio of bandwidth allocation, which targets at simultaneously satisfying as many flows of the traffic matrix as possible. If some demands can not be allocated currently, they will be queued for a certain period until some allocated flows expire or departure so that some resources are released for further allocation.
Energyaware routing
 (i)
The traffic should be aggregated together to the greatest extent, which would allow us to conserve more energy in a tighter data center network.
 (ii)
The route should use as few critical links (interBI links) as possible, which aims to decrease the failure ratio of future allocations and also reduce the computation cost caused by splitting/merging BIs.
 (iii)
The route should use as few network devices as possible, which prefer to choose the shortest path.
 (iv)
The current allocation should impact on the future allocation as little as possible.
Guided by these rules, the ERA assigns a route for each requested demand in an efficient and dynamic way based on the current network status. Initially, ERA searches the lowestlevel BI where the two endpoints of the requested demand are located, and generates a set of feasible candidate routes. For example, as shown in Figure 1 the lowest level for demand (s1, s4, 45) is 50BI N _{1}. This procedure aims to meet the rule i and ii, which tries to aggregate the flows into the same subnet (lowest BI) and use as few critical links as possible. Afterwards, sort these candidate routes by the number of their induced BI splittings, and choose the route(s) that cause fewest BI splittings. This step complies with rule ii and iv, which takes the computation cost and future allocation into consideration. If there are more than one such route, then choose the shortest route which tries to meet the objective of rule iii. In case there are still multiple routes, then choose the route with the maximum number of flows which can contribute to the network traffic aggregation. Finally, we randomly choose one route or just choose the first route from the sorted candidate routes. The poweraware routing procedure terminates as long as the output of the above five procedures is unique, and allocates the best route with the required bandwidth to the current demand.
Reliability satisfaction
Admittedly, the energy conservation in the way of powering off devices sacrifices the network fault tolerance, which is an inevitable conflict between them. In order to improve the robustness of the network, we need to add additional number of available backup routes according to the reliability requirements as illustrated in Constraint (3). The selection of backup routes applies the shortestpath routing algorithm other than following the aforementioned multiple route selection rules. This strategy means to reserve as few devices as possible to meet the requirements of fault tolerance. From another perspective, as indicated in [32] the switches are fairly reliable (only 5% failure rates for ToR switches per year), hence it is not so wise to sacrifice a great deal (network resources, computation time, energy, etc.) for a small probability event. Therefore, the shortestpath routing algorithm is well suited and adequate for the backup routeselection.
Analysis of complexity
As aforementioned, the complexity of constructing a βBI is O(L), where L denotes the number of links. The route searching or routing decidability experiences a complexity of O(1). The Blocking Island Hierarchy (BIH) reflects the realtime state of the available network bandwidth, and it needs to be updated when the link state changes. Yet we only need to update the BIs, which are involved in allocating or deallocating bandwidths, by means of splitting or merging BIs. This means there is no need to compute the whole BIH again. The complexity of updating the BIH is O(rl), where r is the number of different resource requirements (β) and l indicates the number of involved links.
Topologybased heuristic scheme

How many switches and ports should be sufficient to support the network traffic.

How to distribute the traffic flows among the calculated network subset and achieve high network utilization.
Calculate the minimum network subset
In order to maximize power conservation, we need to only keep the required networking capacity available by merging the traffic flows and switch off idle devices. The minimum network subnet should be dynamically computed according to the statistics of traffic demands in runtime. The port statistics and switch status are collected by the centralized controller from OpenFlow enabled switches through the OpenFlow secure channel. It can accurately and directly obtain the statistics of the traffic matrix by using the builtin features (bytes, packet counters, etc.) for active flows kept in OpenFlow switches. In order to deal with the single point failure of the controller, THS provides multiple controllers with different roles (OFPCR_ROLE_EQUAL,OFPCR_ROLE_MASTER and OFPCR_ROLE_SLAVE) to guarantee the robustness of the system (the same as specified in [34]).
We assume that kport switches are used in the FatTree topology. Consequently, there are k Pods, \(\frac {k}{2}\) edge/aggregation switches in each Pod, and \(\frac {k^{2}}{2}\) edge/aggregation switches (Equation 9) in total. The number of core switches on the network is \(\frac {k^{2}}{4}\). Furthermore, there are \(\frac {k^{2}}{4}\) equalcost paths between any given pair of hosts in different Pods.
The number of active switches and ports on the network
Devices  Number of active  Total number of active  Uplink  Downlink  Number of active ports 

Switches in pod i  Switches on the network  ports  ports  per Switch in pod i  
Edge switch  \(\frac {k}{2}\)  \(\frac {k^{2}}{2}\)  \(NS^{Agg}_{i}\)  \(\frac {k}{2}\)  \(NP^{Edge}_{i}=\frac {k}{2}+NS^{Agg}_{i}\) 
Agg switch  \(NS^{Agg}_{i}\)  \({\sum \limits _{i}^{k}}NS^{Agg}_{i}\)  N S ^{ C o r e }  \(\frac {k}{2}\)  \(NP^{Agg}_{i}=\frac {k}{2}+NS^{Core}\) 
Core switch  /  N S ^{ C o r e }  /  /  N P ^{ C o r e }=k 
Traffic distribution using multipath routing
After obtaining the capable network subset, we need to distribute the traffic flows evenly among the subset. Since the size of each flow demand varies much, so the traffics that should be able to fill the network subset cannot fully utilize the network with unexpected low throughput when single path routing is applied. As aforementioned, the FatTree DCN holds many equalcost paths between any pair of servers. Therefore, we can divide the total flow demands by the switch capacity and distribute them evenly using multipath routing algorithms by splitting each flow into multiple subflows. However, the existing flowlevel multipath routing relying on perflow static hashing, like ECMPbased MPTCP [35], still cannot guarantee the traffic will be evenly distributed, which would lead to substantial bandwidth loss and significant load imbalance. Against this kind of binpacking problem, one may apply best fit decreasing (BFD) or first fit decreasing (FFD) heuristic algorithms to mitigate this issue, but still not achieve perfect network utilization and bisection bandwidth. Intuitively, the best method is to distribute all packets evenly among all equal cost paths using packetlevel multipath routing.
Based on some careful studies and extensive experiments, recently A. Dixit et al. proposed a packetlevel traffic splitting scheme named RPS (Random Packet Spraying) [11] for data center networks. RPS achieves near ideal load balance and network utilization, and causes little packet reordering by exploiting the symmetry of the multirooted tree topologies. Noticing that THS’s strategy of powering off switches barely affects the symmetry of the network since we always choose the leftmost switches. Therefore, after obtaining the required subset and adding FTredundancy, we directly borrow the packetlevel RPS multipath routing scheme to spread all flows equally among multiple different equal cost shortest paths. Then switch off the switches and ports which are not involved in final subset and update the network topology. The whole THS procedure is depicted in detail in Algorithm 2.
System evaluation
Simulation overview
In order to evaluate the performance and effectiveness of our proposed approaches (BHS and THS) to the power optimization problem, in this section we implement the blocking island paradigm and two heuristic schemes in the DCNSim simulator [36]. DCNSim can simulate several data center network topologies and compute many metrics, such as the power consumption, aggregate bottleneck throughput, network latency, average path length, faulttolerance, and statistics of device status. Without loss of generality, all simulations in this section are conducted based on FatTree topology, and all the links are capable of bidirectional communications, where the unidirectional link bandwidth is set to be 1 GBps. The default MTU of a link is 1500 bytes, the default packet size is one MTU, and the default buffer size of each switch is 10 MTU. The default processing time for a packet at a node is 10 μs while the default propagation delay of a link is 5 μs, and the TTL of a packet is set to be 128. The time interval between two packets from the same source to the same destination is set to be 5 ms as default.
Evaluation indicator
where PEC denotes the percentage of energy conservation, P _{ B H S/T H S } indicates the power consumed by BHS or THS, and P _{ a l w a y s−o n } represents the power consumed by the traditional alwayson strategy.
To calculate the power consumption, we use the real power consumption data of Cisco Nexus 3048 Data Center Switch. According to its switch data sheet [37], the typical operating power consumption of a Nexus 3048 switch is 120 watts at 100% loads, and powering off one port of the switch saves around 1.5 watts. Moreover, reducing the power consumed by the network can also result in cooling power savings proportionally, though this part of power savings is not taken into any calculation in this paper.
Network traffic matrix
Aside from the poweraware routing and resource allocation mechanisms which mainly determine how much power can be conserved, the traffic pattern also has a great impact on power savings and network performance. In data center networks, there are several typical types of traffic patterns, including OnetoOne, OnetoMany, and AlltoAll. In this section, all the simulations are conducted by applying the AlltoAll traffic pattern, which simulates the most intensive network activities and can evaluate the guaranteed performance under the most rigorous case. Furthermore, according to the findings in [38] about the characteristics of the packetlevel communications, the packet interarrival time reveals an ON/OFF pattern and its distribution follows the Lognormal mode for OFF phase, while the distribution varies between Lognormal mode, Weibull mode and Exponential mode in different data centers during the applicationsensitive ON phase. Here, the Exponential Flow Mode is applied to determine the distribution of packet interarrival times.
Simulation results
This subsection evaluates the overall performance of BHS and THS. The primary simulation results show that achieving 20% to 60% of power savings is feasible, and it varies under different network conditions (traffic loads, network scales, traffic patterns, etc.) and different reliability requirements.
System reliability
Network latency
Network scales
Computation efficiency
The implementation of BHS and THS in real world scenario
We have provided theoretical analysis and simulation studies for the proposed two green schemes BHS and THS. Although the simulation conditions are very close to the real world data center environments, there are still some issues needed to be considered in real world deployment of BHS and THS. Firstly, as aforementioned, the traffic patterns and packet interarrival time change time to time in the real world, though they may follow some disciplines (Lognormal, Exponential, Weibull, etc.) on the whole in the long run. We only simulated the Exponential flow mode according to the findings about the real world traffic characteristics in [38], and the performance of BHS and THS under other one or several mixed traffic patterns in a real world are left for further evaluation. Secondly, we also care about how the time cost in switching off/on a switch will affect the system performance in real data centers, which is actually a common concern of this research field. Another issue needed to be considered is the deployment of BHS. BHS requires a centralized controller which plays a very important role in the BI/BIG/BIH generation, bandwidth allocation and routing rules computation. How to guarantee the robustness of the centralized controller is a big concern. Besides, the choice of communication method (inband or outofband) between controller and switches is also needed to be weighted. All of the mentioned issues above are difficult to be simulated in simulators, which should be considered in real world scenarios.
Conclusion
In this paper, firstly we rigorously formulated the power optimization problem in data center networks into an MCF problem, and proved its NPhardness. In response to this NPhard problem, inspired by an Artificial Intelligence abstraction technique, we proposed a Blocking Island based heuristic scheme (BHS) by designing an energyaware bandwidth allocation mechanism and an energyaware routing algorithm, which can decrease the computation complexity and increase the success ratio of bandwidth allocation. To the best of our knowledge, we are the first to employ the BI paradigm into data center networks to achieve power conservation. Furthermore, we proposed a topologybased heuristic scheme (THS), which focuses on how to compute the minimum network subset and how to distribute the traffic flows properly among the network subset. THS performs faster than BHS and holds the best scalability (O(N)), but the quality of the resulting solution is not as good as BHS. BHS achieves a bandwidth guaranteed data center network irrespective of network topologies with a high quality solution and low computation cost. Comparatively, BHS provides a more attractive and practical solution. The conducted simulations further confirmed their feasibility and good performance.
Declarations
Acknowledgement
We thank Mrs. Shauna Dalton who carefully revised this paper for grammar and spelling.
Authors’ Affiliations
References
 Ting W, Zhiyang S, Yu X, Yang L, Jogesh M, Mounir H (2014) SprintNet: A high performance servercentric network architecture for data centers In: Communications (ICC), 2014 IEEE International Conference on, 4005–4010.. IEEE.Google Scholar
 Ting W, Zhiyang S, Yu X, Yang L, Jogesh M, Mounir H (2015) Designing efficient high performance servercentric data center network architecture. Comput Netw 79: 283–296.View ArticleGoogle Scholar
 Ting W, Yu X, Jogesh M, Mounir H, Sebti F (2014) A general framework for performance guaranteed green data center networking In: Global Communications Conference (GLOBECOM), 2014 IEEE, 2510–2515.. IEEE.Google Scholar
 Abts D, Marty MR, Wells PM, Klausler P, Liu H (2010) Energy proportional datacenter networks In: ACM SIGARCH Computer Architecture News, 338–347.. ACM.Google Scholar
 Heller B, Seetharaman S, Mahadevan P, Yiakoumis Y, Sharma P, Banerjee S, McKeown N (2010) Elastictree: Saving energy in data center networks In: NSDI, 249–264.Google Scholar
 Mahadevan P, Banerjee S, Sharma P, Shah A, Ranganathan P (2011) On energy efficiency for enterprise and data center networks. Commun Mag IEEE 49(8): 94–100.View ArticleGoogle Scholar
 Greenberg A, Hamilton J, Maltz DA, Patel P (2008) The cost of a cloud: research problems in data center networks. In: ACM SIGCOMM computer communication review 39(1): 68–73.Google Scholar
 Pelley S, Meisner D, Wenisch TF, VanGilder JW (2009) Understanding and abstracting total data center power In: Workshop on EnergyEfficient Design.Google Scholar
 Ting W, Zhiyang S, Yu X, Mounir H (2014) Rethinking the Data Center Networking: Architecture, Network Protocols, and Resource Sharing. In: Journal of IEEE Acess 2(1): 1481–1496.Google Scholar
 Dean J, Ghemawat S (2008) Mapreduce: simplified data processing on large clusters. Commun ACM 51(1): 107–113.View ArticleGoogle Scholar
 Dixit A, Prakash P, Hu YC, Kompella RR (2013) On the impact of packet spraying in data center networks. In: INFOCOM, 2013 Proceedings IEEE (pp. 21302138). IEEE.Google Scholar
 Arlitt M, Bash C, Blagodurov S, Chen Y, Christian T, Gmach D, Hyser C, Kumari N, Liu Z, Marwah M, McReynolds A, Patel C, Shah A, Wang Z, Zhou R (2012) Towards the design and operation of netzero energy data centers In: 13th IEEE Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems (ITherm).Google Scholar
 Goiri Í, Le K, Nguyen TD, Guitart J, Torres J, Bianchini R (2012) Greenhadoop: leveraging green energy in dataprocessing frameworks In: Proceedings of the 7th ACM european conference on Computer Systems, 57–70.. ACM.Google Scholar
 Nguyen KK, Cheriet M, Lemay M, Savoie M, Ho B (2013) Powering a data center network via renewable energy: A green testbed. Internet Comput IEEE 17(1): 40–49.View ArticleGoogle Scholar
 David H, Fallin C, Gorbatov E, Hanebutte UR, Mutlu O (2011) Memory power management via dynamic voltage/frequency scaling In: Proceedings of the 8th ACM international conference on Autonomic computing, 31–40.. ACM.Google Scholar
 Leverich J, Monchiero M, Talwar V, Ranganathan P, Kozyrakis C (2009) Power management of datacenter workloads using percore power gating. Comput Architecture Lett 8(2): 48–51.View ArticleGoogle Scholar
 Meisner D, Gold BT, Wenisch TF (2009) Powernap: eliminating server idle power. In: ACM SIGARCH Computer Architecture News 37(1): 205–216.Google Scholar
 Rangan KK, Wei GY, Brooks D (2009) Thread motion:finegrained power management for multicore systems. In: ACM SIGARCH Computer Architecture News (Vol. 37, No. 3, pp. 302313). ACM.Google Scholar
 Ting W, Zhiyang S, Yu X, Bo Q, Mounir H (2014) NovaCube: A low latency Torusbased network architecture for data centers. In Global Communications Conference (GLOBECOM), 2014 IEEE (pp. 22522257). IEEE.Google Scholar
 AbuLibdeh H, Costa P, Rowstron A, O’Shea G, Donnelly A (2010) Symbiotic routing in future data centers. ACM SIGCOMM Comput Commun Rev 40(4): 51–62.View ArticleGoogle Scholar
 Hong CY, Caesar M, Godfrey P (2011) Pcube: Improving power efficiency in data center networks In: Cloud Computing (CLOUD), IEEE International Conference on, 65–72.. IEEE.Google Scholar
 Singla A, Singh A, Ramachandran K, Xu L, Zhang Y (2010) Proteus: a topology malleable data center network In: Proceedings of the 9th ACM SIGCOMM Workshop on Hot Topics in Networks, 8.. ACM.Google Scholar
 Valancius V, Laoutaris N, Massoulié L, Diot C, Rodriguez P (2009) Greening the internet with nano data centers In: Proceedings of the 5th international conference on, Emerging networking experiments and technologies, 37–48.. ACM.Google Scholar
 Liu L, Wang Hm, Liu X, Jin X, He WB, Wang QB, Chen Y (2009) Greencloud: a new architecture for green data center In: Proceedings of the 6th international conference industry session on Autonomic computing and communications industry session, 29–38.. ACM.Google Scholar
 Mann V, Kumar A, Dutta P, Kalyanaraman S (2011) Vmflow: leveraging vm mobility to reduce network power costs in data centers In: NETWORKING, 198–211.. Springer Berlin, Heidelberg.Google Scholar
 Shang Y, Li D, Xu M (2010) Energyaware routing in data center network In: Proceedings of the first ACM SIGCOMM workshop on Green networking (pp. 18).. ACM.Google Scholar
 Martello S, Toth P (1990) Knapsack problems: algorithms and computer implementations. John Wiley and Sons, Inc.Google Scholar
 Frei C, Faltings B (1997) Simplifying network management using blocking island abstractions. Internal Note from the IMMuNe Project.Google Scholar
 Mackworth AK (1992) Constraint satisfaction problems. Encyclopedia of AI, 285, 293.Google Scholar
 Hong CY, Caesar M, Godfrey PB (2012) Finishing flows quickly with preemptive scheduling. ACM SIGCOMM Computer Communication Review 42(4): 127–138.View ArticleGoogle Scholar
 Wilson C, Ballani H, Karagiannis T, Rowtron A (2011) Better never than late: meeting deadlines in datacenter networks. In: ACM SIGCOMM Computer Communication Review (Vol. 41, No. 4, pp. 5061). ACM.Google Scholar
 Bodík P, Menache I, Chowdhury M, Mani P, Maltz DA, Stoica I (2012) Surviving failures in bandwidthconstrained datacenters In: Proceedings of the ACM SIGCOMM 2012 conference on, Applications, technologies, architectures, and protocols for computer communication, 431–442.. ACM.Google Scholar
 AlFares Mohammad, Loukissas Alexander, Vahdat Amin (2008) A scalable, commodity data center network architecture In: ACM SIGCOMM Computer Communication Review, 63–74.Google Scholar
 (2012) Openflow switch specification v1.3.0. www.opennetworking.org/images/stories/downloads/sdnresources/onfspecifications/openflow/openflowspecv1.3.0.pdf.
 Ford Alan, Raiciu Costin, Handley Mark, Bonaventure Olivier and others (2011) Tcp extensions for multipath operation with multiple addresses. Internet Engineering Task Force, RFC, 6824.Google Scholar
 Liu Y, Muppala J (2013) DCNSim: A data center network simulator In: Distributed Computing Systems Workshops (ICDCSW), 2013 IEEE 33rd International Conference on, 214–219.. IEEE.Google Scholar
 Nexus C (2012) 3048 switch data sheet.Google Scholar
 Benson T, Akella A, Maltz DA (2010) Network traffic characteristics of data centers in the wild In: Proceedings of the 10th ACM SIGCOMM conference on Internet measurement, 267–280.. ACM.Google Scholar
Copyright
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.