QoS provisioning for various types of deadline-constrained bulk data transfers between data centers

Future Generation Computer Systems 105 (2020) 162–174

Contents lists available at ScienceDirect

Future Generation Computer Systems journal homepage: www.elsevier.com/locate/fgcs

QoS provisioning for various types of deadline-constrained bulk data transfers between data centers✩ ∗

Aiqin Hou a , Chase Q. Wu b , , Ruimin Qiao a , Liudong Zuo c , Michelle M. Zhu d , Dingyi Fang a , Weike Nie a , Feng Chen a a

School of Information Science and Technology, Northwest University, Xi’an, Shaanxi 710127, China Department of Computer Science, New Jersey Institute of Technology, Newark, NJ 07102, USA c Computer Science Department, California State University, Dominguez Hills, Carson, CA 90747, USA d Department of Computer Science, Montclair State University, Montclair, NJ 07043, USA b

article

info

Article history: Received 21 February 2019 Received in revised form 27 September 2019 Accepted 28 November 2019 Available online 3 December 2019 Keywords: Big data Data center High-performance networks Software-defined networking Bandwidth scheduling

a b s t r a c t An increasing number of applications in scientific and other domains have moved or are in active transition to clouds, and the demand for big data transfers between geographically distributed cloudbased data centers is rapidly growing. Many modern backbone networks leverage logically centralized controllers based on software-defined networking (SDN) to provide advance bandwidth reservation for data transfer requests. How to fully utilize the bandwidth resources of the links connecting data centers with guaranteed quality of service for each user request is an important problem for cloud service providers. Most existing work focuses on bandwidth scheduling for a single request for data transfer or multiple requests using the same service model. In this work, we construct rigorous cost models to quantify user satisfaction degree, and formulate a generic problem of bandwidth scheduling for multiple deadline-constrained data transfer requests of different types to maximize the request scheduling success ratio while minimizing the data transfer completion time of each request. We prove this problem to be not only NP-complete but also non-approximable, and hence design a heuristic algorithm. For performance evaluation, we establish a proof-of-concept emulated SDN testbed and also generate large-scale simulation networks. Both experimental and simulation results show that the proposed scheduling scheme significantly outperforms existing methods in terms of user satisfaction degree and scheduling success ratio. © 2019 Elsevier B.V. All rights reserved.

1. Introduction Similar to Internet Service Providers (ISPs), Cloud Service Providers (CSPs) aim to satisfy as many customers as possible with high throughput. Most of the leading CSPs deploy geographically-distributed data centers (DCs) to provide various types of services to their customers. These DCs are typically connected by links of high bandwidths across wide-area networks (WANs). The network infrastructures that connect these geographically distributed DCs cost millions of dollars annually [2], but unfortunately have not been fully utilized. The average utilization of network resources on even those busy inter-DC links is 40%– 60% [2], and is 30%–40% [3] on many others, partially due to the traditional best-effort transfer method on the Internet. As ✩ Some preliminary results in this manuscript were published in INDIS’18 in conjunction with SC’18 Hou et al. (2018) [1]. ∗ Corresponding author. E-mail address: [email protected] (C.Q. Wu). https://doi.org/10.1016/j.future.2019.11.039 0167-739X/© 2019 Elsevier B.V. All rights reserved.

the number of cloud-based applications continues to increase, it has become a significant challenge to fully utilize the bandwidth resources of inter-DC network links to accommodate as many data transfer requests as possible and meanwhile maximize the throughput of the entire network system. Nowadays, the backbones of many WANs employ new technologies to create high-performance networks (HPNs) (e.g., ESnet [4], Internet2 [5], etc.), which provide the capability of advance bandwidth reservation over dedicated channels provisioned by circuit-switching infrastructures or IP-based tunneling techniques for big data transfer. Particularly, the emerging softwaredefined networking (SDN) technologies greatly facilitate HPN deployments, and in fact, many HPNs have incorporated SDN capabilities into their network infrastructures to provide better Quality of Service (QoS). Such networks feature a virtual singleswitch abstraction on top of data planes that employ both a bandwidth reservation system and SDN concepts [6,7]. Generally, bandwidth reservation allows a batch of user transfer requests accumulated over a period of time to be scheduled collectively in advance, and has proven to be an effective solution to providing

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

QoS guaranteed transfer services and meanwhile achieving a high utilization of network resources. There exist different types of inter-DC data transfers, among which, bulk data transfer requests (BDTRs) for large data volumes on the order of terabytes to petabytes with deadline constraints account for a major portion of traffic (e.g., 85%–95% in some WANs) [2,8–10]. However, most existing solutions for BDTRs are tailored for private cloud services, hence limiting their generalization and scope of application. For example, both Software-Driven WAN (SWAN) [2] and B4 [3] take traffic engineering approaches to improve the inter-DC WAN utilization by considering traffic characteristics and priorities (e.g., interactive > elastic > background). However, neither of them addresses the deadline constraint of BDTRs, one of the most common performance requirements from users [9]. As an increasing number of applications in scientific and many other domains have migrated from local computing and storage platforms to clouds, the demand for inter-DC data transfer with different types of BDTRs is rapidly growing, but the bandwidth scheduling problem in the emerging cloud environment still remains largely unexplored. In this paper, we investigate a bandwidth scheduling problem for two types of BDTRs with fixed or variable bandwidth. Given multiple such BDTRs, we aim to fully utilize inter-DC link bandwidth resources and schedule as many BDTRs as possible while minimizing the earliest complete time (ECT) of each request. Specifically, we construct a rigorous cost model, define a new performance metric named user satisfaction degree, and then formulate a generic problem, Bandwidth Scheduling for Multiple Requests of Various Types, referred to as BS-MRVT. We prove BSMRVT to be not only NP-complete but also non-approximable, and then propose an efficient heuristic scheduling algorithm. We conduct proof-of-concept experiments on a Mininet-based emulated testbed and also extensive simulations for bandwidth scheduling in both simulated and real-life networks. Both experimental and simulation results show that our proposed algorithm significantly outperforms existing methods in terms of user satisfaction degree and scheduling success ratio. The rest of this paper is organized as follows. We conduct a survey of related work in Section 2. We construct network models and formulate BS-MRVT with complexity analysis in Section 3. We design the algorithm with a detailed illustration in Section 4. We conduct performance evaluation in Section 5 and conclude our work in Section 6. 2. Related work There have been a number of successful research efforts in making full use of network resources for bulk data transfer between data centers using centralized traffic engineering techniques. For example, Jain et al. developed B4 [3], which is able to globally schedule massive bandwidth requirements at a modest number of sites. Hong et al. designed SWAN [2], a centrally controlling system that enables inter-DC WANs to carry more traffic. Kandula et al. developed TEMPUs [8], an online temporal planning scheme that packs long-running transfers across network paths and future time steps, while leaving enough capacity slack for future high-priority requests. However, the aforementioned work does not consider data transfer deadline. Moreover, most of these existing schemes are tailored for their private clouds and may not work as well for public clouds. Zhang et al. designed Amoeba [9] for inter-DC WANs to ensure that data transfer be completed before a hard or soft deadline through centralized traffic engineering and bandwidth reservation scheme. Jin et al. presented Owan [10] to schedule bulk data transfer requests to meet their deadlines in a modern WAN infrastructure with an intelligent optical network-layer.

163

Nandagopal et al. proposed GRESE, which attempts to minimize the overall bandwidth costs by leveraging the flexible nature of the deadlines of certain bulk data transfers [11]. Lin and Wu formulated and solved a class of transport-support workflow optimization problems for end-to-end data transfer path compositions and transport method selections to meet data transfer requests [12]. There also exists some work on routing policy for bulk data transfer. Yassine et al. designed a multi-rate bandwidth-ondemand broker [13], which employs a scheduling algorithm that considers both delay tolerant and intolerant multimedia data by using the concept of dense wavelength division multiplexing in backbone optical networks. Bandwidth-on-demand provides dynamic multi-rate for communication between geographically distributed cloud-based data centers to meet the requirements specified in the service level agreement. In [14], Shu and Wu investigated the problem of bandwidth scheduling for energy efficiency in high-performance networks. Some efforts have been made to address the problem of scheduling multiple bandwidth reservation requests for bulk data transfer over a single path. Sharma et al. studied the problem of accommodating as many bandwidth reservation requests (BRRs) as possible while minimizing the total time to complete all data transfers on the same path. The problem was proved to be NP-hard and a heuristic algorithm was proposed. Zuo et al. investigated the problem of scheduling as many concurrent bandwidth reservation requests as possible on one dedicated channel in an HPN [15]. Wang et al. studied a periodic bandwidth scheduling problem to maximize the number of satisfied user requests for bandwidth reservation with deadline constraint on a fixed network path [16]. Note that these problems only consider a single network path, which is not common in real applications. Some researchers investigated the problem of bandwidth scheduling for multiple user requests in HPNs. Zuo and Zhu studied the problem of scheduling as many concurrent bandwidth reservation requests as possible in an HPN while achieving the average earliest complete time (ECT) and the average shortest duration (SD) of scheduled BRRs [17]. Given multiple bandwidth reservation requests in a batch awaiting to be scheduled in a dedicated network, Zuo et al. studied two scheduling maximization problems: maximize the amount of data to be transferred and maximize the number of requests to be scheduled [18]. Srinivasan et al. proposed a bandwidth allocation scheme that flexibly and adaptively allocates bandwidth for big data transfer requests with the objective to maximize the acceptance ratio of the requests while satisfying the deadline constraints [19]. Wang et al. leveraged the temporal and spacial characteristics of inter-DC bulk data traffic and investigated the problem of scheduling multiple bulk data transfers to reduce network congestion [20]. Given one data transfer request aiming to achieve the ECT, Lin and Wu in [21] investigated bandwidth scheduling with an exhaustive combination of different path and bandwidth constraints: (i) fixed path with fixed bandwidth (FPFB), (ii) fixed path with variable bandwidth (FPVB), (iii) variable path with fixed bandwidth (VPFB), and (iv) variable path with variable bandwidth (VPVB). VPFB and VPVB are further divided into VPFB-0/1 and VPVB-0/1, respectively, depending on whether or not the path switch delay is negligible. They also analyzed the complexities of these problems and proposed optimal and heuristic algorithms. The above survey of related work shows that although bandwidth scheduling has been studied extensively in various contexts in the literature, there are very limited efforts to address the problem of scheduling multiple concurrent bandwidth reservation requests of different types for big data transfer among data centers, which is the focus of our research.

164

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

Fig. 2. An example network topology.

Fig. 1. An example HPN with bandwidth reservation service.

3. Problem formulation In this section, we construct rigorous network models to define a new performance metric named user satisfaction degree, and then formally formulate the problem along with detailed complexity analysis. 3.1. Network model To support data movement across a backbone network between two geographically distributed data centers, a central management system (such as a control plane) maintains global information about the link bandwidths and traffic demands. As the core unit of the entire system, the bandwidth scheduler is responsible for reserving bandwidths for BDTRs and releasing the reserved bandwidths after the corresponding data transfers have been successfully completed. For illustration purposes, we provide an example backbone network with bandwidth reservation service in Fig. 1, where the control plane with Global Network View (GNV) provides realtime network status information to the bandwidth scheduler. Fig. 2. shows the topology of a simple backbone network. For simplicity, we represent the network topology as a graph G(V , E) with |V | nodes and |E | links, where each link l ∈ E maintains a list of residual or available bandwidths specified as a segmented constant function of time [21]. We use a 3-tuple of time-bandwidth (TB) (tl [i], tl [i + 1], bl [i]) to denote the residual or available bandwidth bl [i] of link l during the ith time-slot (i.e., the time interval [tl [i], tl [i + 1]]), i = 0, 1, 2, . . . , Tl − 1, where Tl is the total number of time-slots on link l. We combine the TB lists of all links to build an Aggregated TB (ATB) list, where we store the residual or available bandwidths of all links in each intersected time-slot. As shown in Fig. 3, we create a set of new time slots by combining the time slots of all links, and then map the residual bandwidth of each link to the ATB list in each new time slot. We denote the ATB list as (t [0], t [1], b0 [0], b1 [0], . . . , b|E |−1 [0]), . . . , (t [T − 1], t [T ], b0 [T − 1], b1 [T − 1], . . . , b|E |−1 [T − 1]), where T is the total number of new time-slots after the aggregation of TB lists of all |E | links. 3.2. Problem formulation The data transfer in a BDTR is guaranteed to complete before the specified deadline once the request has been successfully scheduled. In this work, we consider two types of BDTRs that are common in big data applications: Fixed Bandwidth Bulk data Request (FBBR) and Variable Bandwidth Bulk data Request (VBBR). A BDTR ri is in the form of (vis , vid , tid , δi , bmax , bti ) denoting that i the user needs to transfer δi amount of data by the deadline tid

Fig. 3. Aggregated time bandwidth list of two links.

from source node vis to destination node vid , and bti is a variable indicating whether it requires a fixed bandwidth during the entire transfer (FBBR when bti = F ) or not (VBBR when bti = V ). Different from our previous work [22,23], where we consider priority and preemption for different types of network applications (e.g., on-demand vs. long-term), we do not prioritize the user requests in this work as they belong to the same class of network applications that require bulk transfer of big data. The reserved bandwidth, either fixed or variable, is upper limited by the maximum local area network (LAN) bandwidth bmax [18], and i the default earliest data transfer start time is 0. For example, a BDTR ri that requires a fixed bandwidth to transfer 50 Gb/ of data from vs to vd within time interval [0, 5 s] under the constraint of the maximum LAN bandwidth of 6 Gb/s is denoted as (vs , vd , 5 s, 50 Gb/, 6 Gb/s, F ). In general, given multiple concurrent BDTRs, the scheduling success ratio (SSR), defined as the percentage of BDTRs that can be successfully scheduled, serves as a good indicator for bandwidth scheduling performance. To obtain a reasonable measurement of finegrained QoS, we propose to combine user experiences in terms of both collectivity and individuality to measure the QoS of data transfer. Towards this goal, we define a new performance metric named User Satisfaction Degree (USD) to quantify the transfer performance of each BDTR. We use a boolean variable ai to denote whether BDTR request ri can be successfully accommodated in the network (ai = 1) or not (ai = 0). The USD of ri , denoted by usdi , is defined as

tid

tie +tid

, where tie denotes the data

transfer completion time of ri . Obviously, usdi is maximized if tie is minimized, namely, the ECT of ri is achieved. Since tie ∈ (0, tid ] when ai = 1, we know that usdi ∈ [0.5, 1). Given a set of BDTRs to be scheduled, we define the normalized USD of all BDTRs as: 1

|BDTR|

∑

·

ri ∈BDTR

(ai ·

tid tie

+ tid

),

(1)

subject to

∑

ai · bli (t) ≤ C l , ∀l ∈ E , ∀t ∈ [0, T ),

(2)

ri ∈BDTR

where C l represents the bandwidth capacity of link l and bli (t) denotes the reserved bandwidth for ri on link l within time

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

165

Definition 1. Given a backbone network G(V , E) with an ATB list for all links and multiple BDTRs (vis , vid , tid , δi , bmax , bti ) of two i types, either FBBR or VBBR, we wish to maximize the SSR and the normalized USD of all BDTRs defined in Eq. (1) under the constraint of Eq. (2).

further consider a specific network topology where the bandwidth could only be reserved on one single path from vis to the same destination v d . Therefore, ri always takes one unit of time to complete, and the objective to maximize the normalized 1 total ∑ USD of all BDTRs is essentially equivalent to maximize N · ri ∈BDTR ai , namely, to maximize the number of BDTRs that can be successfully scheduled. The above instance of BS-MRVT is actually the same as maxR considered in [18], which has been proved to be NP-complete. We now show that any instance of the known NP-hard SSUSF problem [24] with the objective to maximize cardinality can be transformed into the above instance of BS-MRVT in polynomial time. In SSUSF [24], each demand requires to transfer di amount of flow from source node si to a unique sink node t over a single path, and the flow route profit can be regarded as the required transfer time, which is 1 for each demand under the bandwidth capacity constraints. For each request ri to be scheduled in BSMRVT with a particular structure shown in Fig. 4, we let vis = si , v d = t , wi = 1, δi = di , and |BDTR| = k. The objective of SSUSF to maximize cardinality is equivalent to the objective of BS-MRVT to maximize |BDTR| that can be successfully scheduled. Hence, if we have a solution to the instance of SSUSF, we have a solution to the instance of BS-MRVT with a particular structure, and vice versa. Since a special case of BS-MRVT with a particular structure is NP-hard, so is the original BS-MRVT problem. Along with BS-MRVT ∈ NP, we conclude that BS-MRVT is NP-complete. Proof ends. □

3.3. Complexity analysis

Theorem 2. BS-MRVT is non-approximable.

We prove BS-MRVT to be both NP-complete and nonapproximable. We first introduce the SSUSF (single-sink unsplittable flow) problem with the objective to maximize cardinality in [24] as follows: given a network graph G′ (V ′ , E ′ ), a bandwidth capacity of each link and a batch of demands {D1 , D2 , . . . , Dk }, where demand Di requires to route di amount of flow from source node si to a unique destination sink node t over a single path, and the corresponding flow route profit wi = 1 (i.e., unit-profit case) for each demand, the goal is to find a schedule that maximizes the number of demands routed successfully under the link bandwidth capacity constraints.

Proof. We prove this theorem using proof by contradiction. From Theorem 1.3 in [24], we know that we cannot approximate SSUSF with a factor of O(|E |1/2−ε ) for any ϵ > 0. Assuming that there exist an approximate algorithm with an approximation ratio of O(|E |1/2−ε ) for a certain ϵ > 0 for BS-MRVT, we show that this assumption implies a polynomial-time optimal solution to SSUSF [24]. We denote an arbitrary demand Di in SSUSF (unit-profit case) as (si , t , 1, di , Ci ), where 1 is the required value of the flow route profit wi by every demand and can be considered as the transfer time for Di . It means that the demand Di needs to transfer di amount of data for a time duration of 1 from source node si to a unique destination node t over a single path. Suppose that the transfer start time is 0, and then the transfer time interval is [0, 1]. Note that the maximum amount of flow is limited by the bandwidth capacity Ci of the link. The objective of SSUSF is to maximize the number of demands routed successfully under the link bandwidth capacity constraints. We then construct a corresponding instance of BS-MRVT in polynomial time. We firstly consider a special BDTR ri : (vis , vid , tid , δi , bmax , bti ) by setting vis = si , v d = t , tid = 1, δi = i di , bmax = δ = C , and bti = F (i.e., it requires a fixed bandwidth i i i for data transfer). It means that request ri needs to transfer δi amount of data with a fixed bandwidth from a source node vis to a unique destination node v d within time interval [0, 1] under the link bandwidth capacity constraints bmax = δi . We further i consider a specific network topology for BS-MRVT as shown in Fig. 4, where the bandwidth could only be reserved on one single path from vis to the same destination v d . Obviously, the reserved bandwidth is the maximum fixed bandwidth δi within time interval [0, 1], and the data transfer completion time of ri is always 1. Hence, maximizing the normalized total USD of all BDTRs is ∑ equivalent to maximizing N1 · ri ∈BDTR ai , namely, to maximize the number of BDTRs that can be successfully scheduled We then apply the assumed approximate algorithm to the instance of BS-MRVT as described above. Obviously, the approximate algorithm finds a path from vis to v d such that the data of

Fig. 4. An instance of BS-MRVT with a particular network structure.

slot t. This constraint denotes that within any time slot t, the total reserved bandwidth on a link cannot exceed the bandwidth capacity of that link. We formally define Bandwidth Scheduling for Multiple Reservations of Various Types (BS-MRVT) as follows [1]:

Theorem 1. BS-MRVT is NP-complete. Proof. The decision version of BS-MRVT is as follows: given a backbone network G(V , E) with an ATB list for all links and multiple BDTRs (vis , vid , tid , δi , bmax , bti ), does there exist a scheduling i strategy such that the SSR ≥ m and the normalized USD of all BDTRs ≥ Nn , where N = |BDTR|? Given the scheduling options of all BDTRs, we can calculate the SSR and USD, and then compare them with m and Nn , respectively, to find the correct answer in polynomial time. Hence, BS-MRVT ∈ NP. We now prove that BS-MRVT is NP-hard. According to the principle of proof by special case in [25], a problem is NP-hard if a special case of this problem with a particular input is equivalent to a known NP-hard problem. Similar to the NP-hardness proofs provided in [17,26], we first construct a special case of BS-MRVT with a particular network structure, as illustrated in Fig. 4. In a backbone network G(V , E) with the ATB list of all links and multiple BDTRs, we consider a special request ri in the form of (vis , v d , 1, δi , δi , F ), which means that ri needs to transfer δi amount of data from a source node vis to a unique destination node v d within time interval [0, 1], the maximum bandwidth constraint is also δi , and the reserved bandwidth must be fixed and no greater than δi within time interval [0, 1]. We

166

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

size δi can be completely transferred along the path during time interval [0, 1], and then we obtain the number of BDTRs successfully scheduled. Obviously, solving such an instance of BS-MRVT is equivalent to solving an arbitrary instance of SSUSF (unitprofit case) with the objective to maximize cardinality. Therefore, this approximate algorithm finds an optimal solution to SSUSF (unit-profit case) whenever one exists. This is in conflict with the NP-completeness of SSUSF. Hence, the given theorem stands. Proof ends. □ 4. Algorithms design and analysis The NP-completeness of BS-MRVT indicates that there does not exist any polynomial-time optimal algorithm for BS-MRVT unless P = NP. In this section, we focus on the design of a heuristic algorithm. 4.1. Design of a heuristic flexible scheduling algorithm for BS-MRVT We design a heuristic algorithm, Flexible Multiple Scheduling for MRVT, referred to as FMS-MRVT, to solve the problem of BSMRVT, whose pseudocode is provided in Algorithm 1. The time complexity of this algorithm is O(|BDTR| · T 2 · (|E | + |V | · log(|V |)) in the worst case, where |BDTR| denotes the total number of requests, T denotes the total number of time-slots on the aggregated TB list, and E and V denote the set of edges and nodes in the network, respectively. Given an HPN with an ATB of all links, for a batch of BDTRs of various types, our goal is to maximize the SSR and the newly defined normalized USD. Therefore, we consider scheduling priority according to: (i) deadlines in the ascending order; (ii) data sizes in the ascending order if they have the same deadline; (iii) FBBR over VBBR if they have the same deadline and data size, as an FBBR request requires fixed bandwidth while a VBBR request is more flexible in bandwidth use. Furthermore, we design different scheduling algorithms for FBBR and VBBR requests, taking into consideration of their bandwidth requirements, as shown in Algorithms 2 and 3, respectively. Algorithm 1 FMS-MRVT Input: an HPN graph G(V , E) with an ATB list of all links, multiple BDTRs (vis , vid , tid , δi , bmax , bti ) i Output: the normalized total user satisfaction degree usd 1: Initialize variable usd = 0; 2: Sort all BDTRs by their deadlines in an ascending order. For BDTRs with the same deadline, further sort them by their data sizes in an ascending order, and for BDTRs with the same data size, FBBRs are placed ahead of VBBRs; s d d max t 3: for each ri : (vi , vi , ti , δi , bi , bi ) in the set of BDTRs do 4: if bti == F then 5: Call Algorithm 2; 6: else 7: Call Algorithm 3; 8: 9: 10: 11:

usdi = ai ·

tid tie +tid

to compute the user satisfaction degree usdi of ri , and add usdi to the total user satisfaction degree and compute the normalized usd. For a FBBR request ri , we design an algorithm named FBBR Scheduling shown in Algorithm 2, which computes a path with the maximum fixed residual bandwidth across different time slots before the deadline of ri . Obviously, this path could be either a fixed path with fixed bandwidth (FPFB) or a variable path with fixed bandwidth (VPFB). We select the one with a larger bandwidth. If they have the same bandwidth, a FPFB path is generally preferred to a VPFB path for the simplicity of implementation as it does not require path switching between adjacent time slots. Given a transfer request ri with start time slot p and end time slot q, which falls in a time range from time slot 0 to the deadline k, the key idea is to compute an FPFB/VPFB path with the maximum residual bandwidth within time slot [p, q]. If the selected path completes the data transfer, we obtain the transfer end time, which is guaranteed to be the earliest completion time; otherwise, we increase q by 1 and continue to the next round. This process is repeated until the transfer end time is achieved or q > k. Algorithm 2 FBBR Scheduling Input: an HPN graph G(V , E) with an ATB list of all links, a FBBR ri (vis , vid , tid , δi , bmax , F) i Output: ai to denote whether or not ri could be successfully scheduled, and tie , ECT of ri , if ai = 1 e 1: Initialize variables ai = 0 and ti = ∞; d 2: Identify time slot k such that t [k] < ti ≤ t [k + 1]; 3: for 0 ≤ q ≤ k do 4: Initialize variable bq = +∞; 5: for q ≥ p ≥ 0 do 6: Use modified Dijkstra’s algorithm on graph G(V , E) to compute the path with the maximum residual bandwidth b from vis to vid within time slot p; 7: bq = min(bq , b); 8: for each l ∈ E do 9: bl = min(bl , bl [p], bmax ), bl is initialized to +∞ at the i beginning of each outer loop (Line 3); 10: Use modified Dijkstra’s algorithm on graph G(V , E) to compute the path with the maximum residual bandwidth b′ from vis to vid within time slots [p, q]; 11: b[p,q] = max(bq , b′ ); ( ) δ 12: if b[p,q] · min(t [q + 1], tid ) − t [p] ≥ δi and t [p]+ b i < tie [p,q]

13: 14: 15: 16: 17: 18:

then ai = 1; tie = t [p] +

δi b[p,q]

;

if ai == 1 then Update the residual bandwidths of the links on the selected path during time interval [0, tie ]; Break; return ai and tie .

;

usd = usd + usdi ; usd = usd/|BDTR|; return usd.

Algorithm 1 is briefly explained as follows. Line 2: To schedule as many BDTRs as possible, we sort the given set of BDTRs according to different criteria (deadline, data size, and type). Lines 3–10: We call Algorithms 2 and 3 to compute the ECTs of FBBRs and VBBRs, respectively. We then use the returned ECT

Algorithm 2 is briefly explained as follows: Line 2: We compute the latest time slot k within which the data transfer deadline tid of the given FBBR ri falls in. Lines 3–17: We consider each time slot q within the range [0, k]. A FBBR requires fixed bandwidth within the corresponding data transfer interval, and we consider two different paths within time slot interval [p, q]: fixed path with fixed bandwidth (FPFB) and variable path with fixed bandwidth (VPFB). For simplicity, we do not consider the path switching delay in VPFB. We initialize bandwidth of the VPFB path, bq , to +∞ (Line 4). We consider each time slot [p, q] in the reverse order (Line 5). After we obtain the

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

path with the maximum residual bandwidth from vis to vid within time slot p, we calculate the maximum fixed residual bandwidth of the variable paths within the time slot range [p, q] (Line 7). We update the available bandwidths of all links within the time slot range [p, q] (Lines 8–9), and then use modified Dijkstra’s algorithm to compute a single path with the maximum residual bandwidth (i.e., the FPFB path with the maximum residual bandwidth b′ ) from vis to vid within time slot range [p, q] (Line 10). We choose the larger bandwidth of the FPFB and the VPFB (Line 11), and determine if the data transfer could be completed using the larger bandwidth within time slot range [p, q] (Line 12). If so, ri can be successfully scheduled within time slot [p, q], and hence we set ai to 1 and calculate the corresponding data transfer completion time tie . If ri can be successfully scheduled within time slot range [p, q], ∀p ∈ [0, q], we do not need to consider any other q, we update the residual bandwidths on the corresponding path that gives us the ECT for ri and break the for loop (Line 15–17). Algorithm 3 VBBR Scheduling Input: an HPN graph G(V , E) with an ATB list of all links, a VBBR ri (vis , vid , tid , δi , bmax , V) i Output: ai to denote whether or not ri could be successfully scheduled, and tie , ECT of ri , if ai = 1 1: Initialize variable ai = 0 and q = 0; d 2: Identify the time slot k such that t [k] < ti ≤ t [k + 1]; 3: for each l ∈ E do 4: bl = +∞; 5: while δi > 0 && q ≤ k do 6: for each l ∈ E do 7: bl = min(bl , bl [q], bmax ); i 8: Use modified Dijkstra’s algorithm on graph G(V , E) to compute the path pi [q] with the maximum residual bandwidth bi [q] from vis to vid within time slot q; ( ) 9: if δi ≤ bi [q] · min(ti [q + 1], tid ) − ti [q] then 10: ai = 1 ; δ 11: tie = ti [q] + b [iq] ; i 12: Update the residual bandwidths of the links on paths pi [0], pi [1], . . . , pi [q]; 13: Break; 14: else 15: δi − = bi [q] · (ti [q + 1] − ti [q]); 16: q + +; e 17: return ai and ti . For VBBR request ri , which requires maximum bandwidth for transfer, we employ a more flexible scheduling scheme, referred to as VBBR Scheduling. The key idea is to use the path set with the maximum residual bandwidth in each time slot for the data transfer of ri . Given any VBBR request, we should start its data transfer as early as possible since we can allocate variable bandwidths on variable paths across different time slots. Similarly, we do not consider the path switching delay. Algorithm 3 is briefly explained as follows: Lines 5–16: We consider each time slot within the range [0, k]. Within time slot q, we use modified Dijkstra’s algorithm to compute the path with the maximum residual bandwidth from vis to vid , and then calculate the maximum amount of data that can be transferred. If the boolean expression in Line 9 is true, the remaining data of ri can be successfully transferred within time slot q, and hence we set ai to 1 (Line 10) and calculate the corresponding data transfer completion time (Line 11). We then update the bandwidths of links on the data transfer paths as noted in Line 12. In Line 14, if the remaining data of ri cannot be transferred within time slot q, we then update the remaining data size of ri (Line 15) and consider the next time slot (Line 16).

167

Table 1 Aggregated time-bandwidth (ATB) list of the links in Fig. 2.

4.2. Illustration of FMS-MRVT We use the example network topology in Fig. 2 along with the following set of BDTRs to illustrate the execution of FMS-MRVT:

• r1 (FBBR) : (vs , vd , 4 s, 10 Gb/, 15 Gb/s, F ) • r2 (FBBR) : (vs , vd , 4 s, 15 Gb/, 17 Gb/s, F ) • r3 (VBBR) : (vs , vd , 5 s, 20 Gb/, 12 Gb/s, V ) We consider that the smallest time-slot is 1 unit, bandwidth capacity of each link is 30 Gb/s and available bandwidths of the network links across six time slots are shown in Table 1. According to FMS-MVTR in Algorithm 1, we first sort the requests in the order of r1 , r2 , and r3 , and then schedule them one at a time. Step 1: We call Algorithm 2 to schedule r1 (FBBR) : (vs , vd , 4 s, 10 Gb/, 15 Gb/s, F ). The computed FPFB path is vs − v2 − vd with the maximum fixed residual bandwidth of 14 Gb/s within time interval [1 s, 2 s] (i.e., time slot (1). Note that the VPFB path and the FPFB path are identical within time interval [1 s, 2 s], so they have the same bandwidth of 14 Gb/s. Hence, the data transfer start time is t1S = 1 s, and the data transfer completion time is t1E = 1.714 s. The bandwidths of the links on vs − v2 − vd within time interval [1 s, 2 s] are updated as follows: the available bandwidths of links vs − v2 and v2 − vd become 2 Gb/s and 0 Gb/s, respectively. We then calculate usd1 = 4 = 0.700. 4+1.714 Step 2: Similarly, we call Algorithm 2 to schedule r2 (FBBR) : (vs , vd , 4 s, 15 Gb/, 17 Gb/s, F ). The computed FPFB path is vs − v2 − vd with the maximum fixed residual bandwidth of 17 Gb/s within time interval [2 s, 3 s] (i.e., time slot (2). Hence, the data transfer start time is t1S = 2 s, and the data transfer completion time is t1E = 2.882 s. We have the same result for VPFB. The bandwidths of the links on vs − v2 − vd within time interval [2 s, 3 s] are updated as follows: the available bandwidths of links vs − v2 and v2 − vd become 1 Gb/s and 0 Gb/s, respectively. We then calculate usd2 = 4+24.882 = 0.581. Step 3: we call Algorithm 3 to schedule VBBR r3 (VBBR) : (vs , vd , 5 s, 20 Gb/, 12 Gb/s, V ). The computed VPVB paths are vs − v1 − vd , vs − v1 − vd , vs − v1 − vd , vs − v2 − vd and vs − v2 − vd with the maximum residual bandwidths of 4 Gb/s, 2 Gb/s, 2 Gb/s, 9 Gb/s and 10 Gb/s within time slots 0, 1, 2, 3 and 4, respectively. The data transfer start time is 0 s, and the data transfer completion time is 4.3 s. The bandwidths of the links on vs − v1 − vd within time slots [0, 2] are updated as follows: the bandwidths of link vs − v1 become 0 Gb/s, 0 Gb/s, and 0 Gb/s while those of link v1 − vd become 2 Gb/s, 4 Gb/s, and 1 Gb/s, respectively. Similarly, the bandwidths of the links on vs −v2 −vd within time slot [3, 4] are updated as follows: the bandwidths of link vs − v2 become 4 Gb/s and 1 Gb/s while those of link v2 −vd become 0 Gb/s and 0 Gb/s, respectively. We then calculate usd3 = 5+54.3 = 0.538.

168

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

Algorithm 4 Fixed-MRVT

Algorithm 5 OptFPFB-MRVT

Input: an HPN graph G(V , E) with an ATB list of all links, multiple BDTRs (vis , vid , tid , δi , bmax , bti ) i Output: the normalized user satisfaction degree usd 1: Initialize variable usdi = 0; s d d max t 2: for each ri : (vi , vi , ti , δi , bi , bi ) in BDTRs do 3: Identify the time slot k such that t [k] < tid ≤ t [k + 1]; 4: for each l ∈ E do 5: bl = +∞; 6: for 0 ≤ q ≤ k do 7: for each l ∈ E do 8: bl = min(bl , bl [q], bmax ); i 9: Use Dijkstra’s algorithm on graph G(V , E) to compute an FPFB path with the maximum residual bandwidth b from vis to vid within time slot interval [0, k]; 10: if b · tid ≥ δi then

Input: an HPN graph G(V , E) with an ATB list of all links, multiple BDTRs (vis , vid , tid , δi , bmax , bti ) i Output: the normalized user satisfaction degree usd 1: Initialize variable usd = 0; s d d max t 2: for each ri : (vi , vi , ti , δi , bi , bi ) in BDTRs do e 3: Initialize variable ti = ∞; 4: Identify the time slot k such that t [k] < tid ≤ t [k + 1]; 5: for 0 ≤ q ≤ k do 6: for each l ∈ E do 7: bl = +∞; 8: for q ≥ p ≥ 0 do 9: The same as Lines 7–9 of Algorithm 4; ( ) δ 10: if b · min(t [q + 1], tid ) − t [p] ≥ δi and t [p] + bi < tie then δ 11: tie = t [p] + bi ; e 12: if ti < +∞ then

δi

11:

tie =

12:

usd+ =

13: 14: 15:

b

; tid e ti +tid

;

Update the residual bandwidths of the links on the corresponding path within time interval [0, tie ] ; usd = usd/|BDTR|; return usd.

13: 14: 15: 16: 17:

usd+ =

tid

tie +tid

;

The same as Line 12 of Algorithm 4; Break; usd = usd/|BDTR|; return usd.

Step 4: All of these three BDTRs can be successfully scheduled, and their normalized USD is calculated as usd = (0.700 + 0.581 + 0.538)/3 = 1.819/3 = 0.606 and SSR = 100%. 4.3. Other heuristic algorithms for BS-MRVT For performance comparison, we design two heuristic algorithms for BS-MRVT, referred to as Fixed-MRVT shown in Algorithm 4 and OptFPFB-MRVT shown in Algorithm 5. Fixed-MRVT is based on the well-known Dijkstra’s algorithm, and OptFPFBMRVT is based on the existing optFPFB algorithm proposed in [21]. Given a BDTR, Fixed-MRVT directly employs modified Dijkstra’s algorithm to compute an FPFB path with the maximum residual bandwidth within time interval [0, t [k + 1]] and checks if the data transfer of the given BDTR can be successfully completed on the path, while OptFPFB-MRVT identifies an FPFB path with the earliest completion time of the data transfer within time interval [0, t [k + 1]]. 5. Performance evaluation For performance evaluation, we implement the proposed algorithms and conduct (i) proof-of-concept experiments on an emulated SDN testbed based on Mininet [27] system, and (ii) extensive simulations in randomly generated networks as well as a real-life HPN topology. 5.1. Experiment-based performance evaluation 5.1.1. Mininet testbed setup The Mininet emulation tool has been widely used for constructing a virtual network topology [28]. It is also suitable for the performance evaluation of bandwidth scheduling algorithms as the overhead introduced by the scheduling process in Mininet emulation is marginal compared to the scheduling algorithm running time, and is almost negligible for large cases [7]. Based on the Mininet emulation tool, we emulate each switch in the virtual network topology using a virtual instance of Open

Fig. 5. A Mininet emulated testbed.

vSwitch [29], and choose OpenDaylight [30] as the OpenFlow controller, which is a Java-based modular open platform for customizing and automating networks of any size and scale. We deploy a small virtual network testbed emulating a 4-site inter-DC WAN, as shown in Fig. 5. On this testbed, we set the bandwidth capacity of each link between two OpenFlow switches to 10Gbps, and set the bandwidth capacity of each link connecting each host to the switch to 1Gbps, as upper-limited by the NIC speed of the host. 5.1.2. Performance comparison 5.1.2.1. Illustration of a scheduling instance. For illustration, we first conduct a scheduling experiment on the above emulated testbed over a period of total 10 time slots, among which the smallest one is of 1 time unit. The available bandwidths of the network links across [0, 9] time slots are provided in Table 2. In addition, a controller supervisory controls all of these four switch nodes and receives various requests from different users on hosts. In this experiment, we consider five BDTRs as follows:

• r1 (VBBR) : (h2 , h4 , 7 s, 7 Gb, 1 Gb/s, V ) • r2 (VBBR) : (h1 , h4 , 9 s, 8 Gb, 1 Gb/s, V )

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

169

Table 2 Available link bandwidths in Gb/s in Fig. 5 across [0, 9] time slots.

• r3 (VBBR) : (h3 , h2 , 6 s, 4 Gb, 1 Gb/s, V ) • r4 (FBBR) : (h1 , h2 , 6 s, 7 Gb, 1 Gb/s, F ) • r5 (FBBR) : (h4 , h3 , 8 s, 3 Gb, 1 Gb/s, F ) In the virtual network topology shown in Fig. 5, the controller receives these BDTR requests from different applications running on different hosts and schedules them according to the available link bandwidths provided in Table 2 using FMS-MRVT and the other two heuristic algorithms, i.e., Fixed-MRVT and OptFPFBMRVT, for performance comparison. Their scheduling outcomes and corresponding performance measurements are tabulated in Tables 3, 4, and 5, respectively. These results show that our proposed FMS-MRVT algorithm achieves 100% SSR and the highest USD among all. Fixed-MRVT achieves the worst performance because it directly employs modified Dijkstra’s algorithm to compute a FPFB path with the maximum residual bandwidth, while OptFPFB-MRVT achieves better performance because it computes the optimal FPFB path with the earliest completion time. Note that both Fixed-MRVT and OptFPFB-MRVT employ the least flexible scheduling model FPFB to serve different types of requests, without considering the maximum throughput requirements of VBBR requests. FMS-MRVT outperforms both algorithms in comparison because it prioritizes BDTRs strategically and applies suitable scheduling models to different types of requests, as explained in detail below: Firstly, we sort BDTRs by their deadlines in the ascending order to ensure that urgent requests are scheduled with priority, and if they have the same deadline, we sort them by their data sizes in the ascending order to accommodate as many requests as possible. For requests with the same deadline and data size, the FBBR type is given a higher priority than the VBBR type because the former has a more stringent bandwidth requirements (constant and continuous) than the latter. Secondly, for each FBBR request, we use the FPFB and VPFB scheduling models to compute candidate transfer paths and select a fixed or a variable path set with a larger bandwidth; while for each VBBR request, we employ the VPVB scheduling model with the most flexible bandwidth utilization to compute a set of paths with the maximum residual bandwidth in the corresponding time slots. This way, it utilizes the bandwidths of any size within any time slots during the transfer period, and hence is able to schedule more requests and outperform the other algorithms. 5.1.2.2. Performance evaluation under various loads. For a thorough performance evaluation, we conduct more emulation-based scheduling experiments with different numbers of requests on the virtual 4-site network testbed in Fig. 5. Each of these experiments spans across a total of 20 time slots, among which, the smallest one is of 1 time unit, and the link bandwidth between any two switches is initialized to be the link capacity of 10Gb/s, as determined by each switch’s linecard speed. We run different network applications on these hosts, which make multiple BDTRs, each of which is either an FBBR or a VBBR with a requested bandwidth up to 1 Gb/s, as limited by each host’s NIC speed. We

Fig. 6. The number of successfully scheduled requests achieved by three algorithms under various loads on the emulation testbed.

Fig. 7. Comparison of the normalized USD achieved by three algorithms under various loads on the emulation testbed.

repeat the scheduling experiments under different loads from 5 to 75 user requests with an interval of 5 requests. The number of successfully scheduled requests and the normalized USD of all requests are plotted in Figs. 6 and 7, respectively. In all of the above experiments, we observe that the proposed FMS-MRVT algorithm significantly outperforms both Fixed-MRVT and OptFPFB-MRVT on this Mininet testbed. As the number of requests increases, FMS-MRVT achieves a better performance improvement because it makes use of limited network resources

170

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174 Table 3 Performance measurements of FMS-MRVT on the emulated testbed. Requests

Time slots

Paths

Bandwidths (Gb/s)

ECT (s)

r1 (VBBR)

0–2 3 4–6

S2 − S4 S2 − S3 − S4 S2 − S4

1, 1, 1 1 1, 1, 1

7

r2 (VBBR)

0–1 2 3 4–5 6 7

S1 S1 S1 S1 S1 S1

1, 1 1 1 1, 1 1 1

8

r3 (VBBR)

0-4

S3 − S2

1, 1, 0.34, 1, 0.66

4.66

r4 (FBBR)

0-6

S1 − S3 − S4 − S2

1

7

r5 (FBBR)

0-2

S4 − S3

1

3

− S2 − S4 − S3 − S4 − S2 − S3 − S4 − S2 − S4 − S3 − S4 − S2 − S4

SSR

Normalized USD

100%

0.588

Table 4 Performance measurements of OptFPFB-MRVT on the emulated testbed. Requests

Time slots

Paths

Bandwidths (Gb/s)

ECT (s)

r1 (VBBR) r2 (VBBR) r3 (VBBR)

0–6 0–7 3–6

S2 − S4 S1 − S3 − S4 S3 − S2

1 1 1

7 8 7

r4 (FBBR)

No path for r4

r5 (FBBR)

0–2

S4 − S3

1

3

SSR

Normalized USD

80%

0.468

SSR

Normalized USD

60%

0.368

Table 5 Performance measurements of Fixed-MRVT on the emulated testbed. Requests

Time slots

Paths

Bandwidths (Gb/s)

ECT (s)

r1 (VBBR) r2 (VBBR)

0–6 0–7

S2 − S4 S1 − S3 − S4

1 1

7 8

r3 (VBBR)

No path for r3

r4 (FBBR)

No path for r4

r5 (FBBR)

0–2

S4 − S3

in a more flexible way. Note that OptFPFB-MRVT computes the optimal FPFB path with the ECT for any request, so each request may occupy the bandwidth during some of the intervals, while Fixed-MRVT directly computes the FPFB path with the maximum residual bandwidth for any request, so each request may take up a fixed amount of bandwidth resources for a long period. 5.2. Simulation-based performance evaluation 5.2.1. Simulation setup To evaluate the performance of the proposed scheduling algorithms on large-scale problems, we conduct scheduling experiments on a real-life HPN, ESnet5 [4] (in Section 5.2.2), and a set of simulated networks (in Section 5.2.3). We also compare the scheduling performance of these algorithms when both reservation loads and simulation networks vary (in Section 5.2.4). We set the time slots to span across 20 time units, and the start time t [0] = 0. The link bandwidths in the network follow a 1 2 normal distribution: b = bmax · e− 2 (x) , where bmax is set to be 100 Gb/s, and x is a random variable within the range of (0, 1]. There may exist multiple types of data flows in public shared networks. However, our work is focused on the scheduling of different user requests for the same class of data flows, i.e., bulk data transfer. Generating such user requests in a random manner reasonably reflects the distribution of network traffics in highperformance network environments, as commonly modeled and adopted in many similar efforts including [17,19,20]. In each run of the simulation, we randomly generate 100–1500 BDTRs. For each BDTR (vis , vid , tid , δi , bmax , bti ), vis and vid are two randomly i

1

3

selected nodes, tid is a random integer between 1 to 20 s, bmax is a i random integer between 1Gbps to 20Gbps, δi is a random integer no larger than bmax · tid , and bti is a random variable (F or V). i 5.2.2. Performance comparison in ESnet5 To mimic the real ESnet scenario, we perform our simulations on the ESnet topology in Fig. 8, which has 57 nodes and 65 links. For each reservation load, we run these algorithms 10 times, each using a different set of generated BDTRs. We evaluate the scheduling performance of these algorithms in terms of two metrics: USD and SSR. The performance measurements are plotted in Figs. 9 and 10. We observe that FMS-MRVT outperforms OptFPFB-MRVT and Fixed-MRVT by 18%–22% and 3–4 times in terms of normalized USD and 50% and 3–5 times in terms of SSR, respectively. 5.2.3. Performance evaluation simulation in randomly generated networks For a scalability test, we randomly generate 15 different largescale networks as shown in Table 6. Without loss of generality, we set the number of BDTRs to be 500, and run each algorithm in these networks. The corresponding performance measurements of these algorithms are plotted in Figs. 11 and 12. We observe that FMS-MRVT achieves consistently better performance over OptFPFB-MRVT and Fixed-MRVT. Specifically, FMS-MRVT outperforms OptFPFBMRVT and Fixed-MRVT by 23%–24% and 2–3 times in terms of normalized USD, and 50% and 3 times in terms of SSR, respectively.

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

171

Table 6 Index of 15 large-scale networks. Index of network

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Number of nodes Number of links

40 80

50 100

60 120

70 140

80 160

90 180

100 200

120 240

150 300

200 400

230 450

260 500

290 520

320 540

350 560

Fig. 8. The topology of ESnet [4].

Fig. 9. Comparison of normalized USD achieved by three algorithms in ESnet5.

Fig. 11. Comparison of normalized USD with a given load in random networks.

Fig. 10. Comparison of SSR achieved by three algorithms in ESnet5.

Fig. 12. SSR comparison with 500 BDTRs in different random networks.

172

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

maximize BDTRs scheduling success ratio while minimizing the data transfer completion time of each request. We considered two different types of BDTRs: FBBRs and VBBRs. We proved BS-MRVT to be both NP-complete and non-approximable, and proposed an efficient heuristic algorithm, FMS-MRVT. For performance comparison, we also designed two heuristic algorithms, namely, OptFPFB-MRVT and Fixed-MRVT, by leveraging from two existing bandwidth scheduling algorithms. The performance superiority of FMS-MRVT was verified by proof-of-concept experiments on a Mininet-based SDN testbed and extensive simulations on both a real-life network topology and a large set of randomly generated networks. We plan to implement and test the proposed bandwidth scheduling solution in real-life high-performance networks such as OSCARS of ESnet [31]. It is also of our future interest to combine the proposed scheduling solution with existing highperformance transport methods to support big data transfer between data centers over wide-area networks. Declaration of competing interest Fig. 13. Normalized USD comparison with variable loads in different networks.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This research is sponsored by National Natural Science Foundation of China under Grant No. U1609202, Key Research and Development Plan of Shaanxi Province, China under Grant No. 2018GY-011, and Xi’an Science and Technology Plan Project under Grant No. GXYD18.2 with Northwest University, China. The authors would also like to acknowledge the anonymous reviewers’ constructive comments. References

Fig. 14. SSR comparison with variable loads in different networks.

5.2.4. Performance evaluation with different reservation loads in different networks Using the same set of randomly generated networks shown in Table 6, we run these algorithms to schedule different numbers of BDTRs ranging from 100 to 1500. The performance measurements, which are qualitatively similar to the previous results, are plotted in Figs. 13 and 14. We observe that FMS-MRVT achieves consistently better performance than OptFPFB-MRVT and FixedMRVT in terms of both USD and SSR. As the network size scales up, both USD and SSR increase; on the other hand, as the number of BDTRs increases, USD increases while SSR decreases. 6. Conclusion and future work In this paper, we investigated an advance bandwidth scheduling problem, i.e., Bandwidth Scheduling for Multiple Reservations of Various Types, referred to as BS-MRVT, with the objective to

[1] A. Hou, C. Wu, D. Fang, L. Zuo, M. Zhu, X. Zhang, R. Qiao, X. Yin, Bandwidth scheduling for big data transfer with deadline constraint between data centers, in: Workshop on Innovating the Network for Data-Intensive Science , INDIS, in Conjunction with the International Conference for High Performance Computing, Networking, Storage, and Analysis, SC’18, Dallas, Texas, USA, 2018. [2] C.Y. Hong, S. Kandula, R. Mahajan, M. Zhang, V. Gill, M. Nanduri, R. Wattenhofer, Achieving high utilization using software-driven WAN, in: Proceedings of SIGCOMM, 2013, pp. 15–26, http://dx.doi.org/10.1145/ 2534169.2486012. [3] S. Jain, A. Kumar, S. Mandal, J. Ong, L. Poutievski, A. Singh, S. Venkata, J. Wanderer, J. Zhou, M. Zhu, J. Zolla, U. Hlzle, S. Stuart, A. Vahdat, B4: Experience with a globally-deployed software defined WAN, in: Proceedings of SIGCOMM, Vol. 43, (4) 2013, pp. 3–14, http://dx.doi.org/10.1145/2486001. 2486019. [4] ESnet, 2018, https://www.es.net. (Online; Accessed May. 1, 2018). [5] Internet2, 2018, http://www.internet2.edu/ion. (Online; Accessed May. 1, 2018). [6] I. Monga, E. Pouyoul, C. Guok, Software-defined networking for big-data science - architectural models from campus to the WAN, in: Proceedings of SuperComputing, SC’12, 2012, http://dx.doi.org/10.1109/SC.Companion. 2012.341. [7] M.F. Aktas, G. Haldeman, M. Parashar, Scheduling and flexible control of bandwidth and in-transit services for end-to-end application workflows, Future Gener. Comput. Syst. 56 (2015) 284–294, http://dx.doi.org/10.1016/ j.future.2015.09.011. [8] S. Kandula, I. Menache, R. Schwartz, S. Babbula, Calendaring for wide area networks, in: Proceedings of SIGCOMM, 2014, pp. 515–526, http: //dx.doi.org/10.1145/2740070.2626336. [9] H. Zhang, K. Chen, W. Bai, D. Han, C. Tian, H. Wang, H. Guan, M. Zhang, Guaranteeing deadlines for inter-data center transfers, IEEE/ACM Trans. Netw. 25 (1) (2017) 579–595, http://dx.doi.org/10.1109/TNET.2016. 2594235.

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174 [10] X. Jin, Y. Li, D. Wei, S. Li, J. Gao, L. Xu, W.X. G. Li, J. Rexford, Optimizing bulk transfers with software-defined optical WAN, in: Proceedings of SIGCOMM, 2016, pp. 87–100, http://dx.doi.org/10.1145/2934872.2934904. [11] T. Nandagopal, K. Puttaswamy, Lowering inter-datacenter bandwidth costs via bulk data scheduling, in: Proceedings of the 12th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing, 2012, pp. 244–251, http://dx.doi.org/10.1109/CCGrid.2012.70.

173

Aiqin Hou received the Ph.D. degree in computer science from Northwest University, Xi’an, China, in 2018. She is currently a faculty member with the School of Information Science and Technology, Northwest University, Xi’an, China. Her research interests include big data, high-performance network, and bandwidth scheduling.

[12] D. Yun, C. Wu, An integrated transport solution to big data movement in high-performance networks, in: Proceedings of the 23rd International Conference on Network Protocol, 2015, pp. 460–462, http://dx.doi.org/10. 1109/ICNP.2015.55. [13] A. Yassine, A. Asghar, S. Shirmohammadi, Bandwidth on-demand for multimedia big data transfer across geo-distributed cloud data centers, IEEE Trans. Cloud Comput. 99 (2016) 1, http://dx.doi.org/10.1109/TCC.2016. 2617369. [14] T. Shu, C.Q. Wu, Bandwidth scheduling for energy-efficiency in highperformance networks, IEEE Trans. Commun. 65 (8) (2017) 3359–3373, http://dx.doi.org/10.1109/TCOMM.2017.2693262. [15] L. Zuo, M.M. Zhu, C.Q. Wu, Concurrent bandwidth scheduling for big data transfer over a dedicated channel, Int. J. Commun. Netw. Distrib. Syst. 15 (2/3) (2015) 169–190, http://dx.doi.org/10.1504/IJCNDS.2015.070970. [16] Y. Wang, C. Wu, A. Hou, On Periodic Scheduling of Bandwidth Reservations with Deadline Constraint for Big Data Transfer, in: Proceedings of the 41th IEEE Conference on Local Computer Networks, Dubai, UAE, November 7–10, pp. 224–227, http://dx.doi.org/10.1109/LCN.2016.48. [17] L. Zuo, M.M. Zhu, Concurrent bandwidth reservation strategies for big data transfers in high-performance networks, IEEE Trans. Netw. Serv. Manag. 12 (2) (2015) 232–247, http://dx.doi.org/10.1109/TNSM.2015.2430358. [18] L. Zuo, M.M. Zhu, C.Q. Wu, Bandwidth reservation strategies for scheduling maximization in dedicated networks, IEEE Trans. Netw. Serv. Manag. 15 (2) (2018) 544–554, http://dx.doi.org/10.1109/TNSM.2018.2794300. [19] S. Srinivasan, T. Huu, M. Gurusamy, Flexible bandwidth allocation for big data transfer with deadline constraints, in: IEEE Symposium on Computers and Communications, 2017, pp. 347–352, http://dx.doi.org/10.1109/ISCC. 2017.8024554.

Chase Q. Wu completed his Ph.D. dissertation with Oak Ridge National Laboratory, Oak Ridge, TN, USA, and received the Ph.D. degree in computer science from Louisiana State University, Baton Rouge, LA, USA, in 2003. He was a Research Fellow with Oak Ridge National Laboratory during 2003–2006 and an Associate Professor with the University of Memphis, Memphis, TN, USA, during 2006–2015. He is currently a Professor of computer science and the Director of the Center for Big Data, New Jersey Institute of Technology, Newark, NJ, USA. His research interests include big data, parallel and distributed computing, high-performance networking, sensor networks, and cybersecurity.

Ruimin Qiao received the B.S. degree in mathematics from Northwest University, Xi’an, China, in 2017. She is currently an M.S. student in the School of Information Science and Technology at Northwest University, Xi’an, China. Her research interests include big data computing and high-performance networking.

[20] Y. Wang, S. Su, A.X. Liu, Z. Zhang, Multiple bulk data transfers scheduling among datacenters, in: Computer Networks, 2014, pp. 123–137, http: //dx.doi.org/10.1016/j.comnet.2014.02.017. [21] Y. Lin, Q. Wu, Complexity analysis and algorithm design for advance bandwidth scheduling in dedicated networks, IEEE/ACM Trans. Netw. 21 (1) (2013) 14–27, http://dx.doi.org/10.1109/TNET.2012.2189127. [22] L. Zuo, C. Wu, N. Rao, A. Hou, C.-H. Chang, Bandwidth preemption for high-priority data transfer on dedicated channels, in: Proc. of the 27th International Conference on Computer Communications and Networks, Hangzhou, China, 2018. [23] P. Dharam, C. Wu, Y. Wang, Advance bandwidth reservation with deadline constraint in high-performance networks, in: Proc. of the International Conference on Computing, Networking and Communications, ICNC, CNC Workshop, Honolulu, Hawaii, USA, 2014. [24] F.B. Shepherd, A. Vetta, The inapproximability of maximum single-sink unsplittable, priority and confluent flow problems, Theory Comput. 13 (20) (2017) 1–25, http://dx.doi.org/10.4086/toc.2017.v013a020. [25] R.L.R.T.H. Cormen, C.E. Leiserson, C. Stein, Introduction to Algorithms, third ed., MIT Press, Cambridge, MA, USA, 2009, http://dx.doi.org/10.1002/ 9783527649846.ch17. [26] S. Sharma, D. Katramatos, D. Yu, End-to-end network qos via scheduling of flexible resource reservation requests, in: Proceedings of 2011 International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’11, 2011, pp. 68:1–68:10, http://dx.doi.org/10. 1145/2063384.2063475.

Liudong Zuo received the Ph.D. degree in computer science from Southern Illinois University Carbondale in 2015. He received the B.E. degree in computer science from University of Electronic Science and Technology of China in 2009. He is currently an assistant professor in Computer Science Department at California State University, Dominguez Hills. His research interests include computer networks, algorithm design, and big data.

Michelle M. Zhu received her Ph.D. degree in computer science from Louisiana State University in 2005. She finished her dissertation research in the Computer Science and Mathematics Division at Oak Ridge National Laboratory. She was an associate professor in the Computer Science Department at Southern Illinois University, Carbondale, until 2016. She is currently an associate professor in the Department of Computer Science at Montclair State University. Her research interests include high-performance computing, grid and cloud computing, and big data.

[27] Mininet: an instant virtual network on your laptop (or other pc), 2019, http://mininet.org/. (Online; Accessed Jan. 1, 2019). [28] C. Paasch, S. Ferlin, O. Bonaventure, O. Bonaventure, Experimental evaluation of multipath TCP schedulers, in: ACM SIGCOMM Workshop on Capacity Sharing Workshop, 2014, http://dx.doi.org/10.1145/2630088. 2631977. [29] Open vSwitch: an open virtual switch, 2019, http://openvswitch.org/. (Online; Accessed Jan. 1, 2019). [30] SDN controller, 2019, http://www.opendaylight.org/. (Online; Accessed Jan. 1, 2019). [31] OSCARS, 2018, http://www.es.net/oscars. (Online; Accessed May. 1, 2018).

Dingyi Fang is currently a Professor with the School of Information Science and Technology, Northwest University, Xi’an, China. His current research interests include mobile computing and distributed computing systems, network and information security, localization, social networks, and wireless sensor networks.

174

A. Hou, C.Q. Wu, R. Qiao et al. / Future Generation Computer Systems 105 (2020) 162–174

Weike Nie received the B.S. degree in electronic engineering, the M.S. degree in electronic and information engineering, and the Ph.D. degree in information and telecommunication engineering from XiDian University, Xi’an, China, in 1997, 2004, and 2009, respectively. Since September 2009, he has been with the Department of Information Science and Technology School, Northwest University, Xi’an, China, where he is currently an Associate Professor. He was a visiting scholar with New Jersey Institute of Technology from February 2017 to February 2018. His current research interests include array signal processing, blind signal processing, and wireless sensor network localization.

Feng Chen received the M.S. degree in computer science from Northwest University, Xi’an, China, in 2007, and the Ph.D. degree in computer science from Northwestern Polytechnical University, Xi’an, in 2012. He is currently a faculty member with Northwest University, Xi’an. His research interests are in the area of wireless networks, social networks, and Internet of Things.