RUGAL: Provisioning of og Services over 5G Slices To Meet QoS with Miniml ost Ashkn Yousefpour, Json P. Jue The University of Texs t Dlls Emil: shkn, jjue}@utdlls.edu Abstrct Recent dvnces in the res of Internet of Things (IoT), cloud computing nd big dt hve been ttributed to the rise of growing number of complex nd useful pplictions. On the other hnd, The fifth genertion (5G) wireless technology is envisioned to provide fster Internet ccess, with lower ltency, nd ubiquitous mobile coverge compred to its predecessors. As IoT becomes more prevlent in our dily life, more dtintensive, dely-sensitive, nd rel-time pplictions re expected to emerge. Ensuring Qulity of Service (QoS) in terms of bndwidth nd low-ltency for these pplictions is essentil in 5G, nd fog computing is seen s one of the primry enblers for stisfying these QoS requirements. og puts compute, storge, nd networking resources closer to the user. In this report, we show the system model of RUGAL, frmework for QoS-wre og-supported 5G Slice Provisioning (QSP). QSP concerns the dynmic deployment or relese of ppliction services over 5G slice on fog nodes, or dimensioning (enlrge or shrink) of fog-supported 5G slice, in order to meet the ltency nd QoS constrints of pplictions while minimizing cost. I. SYSTEM MODEL In this pper we study the QSP problem, which is to dimension (enlrge or shrink) fog-supported 5G slice to comply with the ltency (i.e. QoS) constrints of the ppliction while lso minimizing the cost of the slice resources. The Slice Owner (SO) wishes to run (possibly ltency-sensitive) ppliction(s) over her slice. The SO solves n instnce of the QSP problem to find good solution for utilizing her 5G slice; the SO wishes to comply with the ltency constrints of her ppliction nd wishes to do so with miniml resource cost. A slice cn be shrunk to free up extr resources, which my sve costs, or my be enlrged if needed, to serve the rising demnds, which my reduce the violtion costs. Moreover, within slice, ppliction components (i.e. services) my be dimensioned or plced in different loctions to comply with the ltency constrints nd/or to minimize the cost. A 5G slice is define s networking, compute, storge, nd memory resources reserved to serve one or more pplictions in the 5G networks. Networking resources re modeled in terms of reserved bndwidth between two physicl nodes, wheres compute, storge, nd memory re modeled in terms of reserved resources within physicl node. A slice cn be comprised of severl pplictions tht spn cross fog, cloud, nd edge networks. An IoT ppliction my be composed of severl services tht re essentilly the components of the ppliction nd cn run in different loctions. Such services re normlly implemented s continers, virtul mchines (VMs), or unikernels. or instnce, n ppliction my hve services such s uthentiction, firewll, cching, nd encryption. Some ppliction services re dely-sensitive nd hve tight dely thresholds, nd my need to run closer to the users/dt sources (e.g. cching), for instnce t the edge of the network or on fog computing devices. On the other hnd, some ppliction services re dely-tolernt nd my hve high vilbility requirements, nd my be deployed frther from the users/dt sources long the fog-to-cloud continuum or in the cloud (e.g. uthentiction). We lbel these fog-cpble IoT services tht could run on the fog computing devices or the cloud servers s fog services. To formulte the QSP problem, we introduce some nottion. Let set of (telecom) edge nodes be denoted by, set of cloud servers by, nd set of fog services running over the SO s slice by A. Some services comprise the components of complex IoT ppliction tht will eventully be run over 5G slice. Some edge nodes hve fog computing cpbilities (e.g. compute, storge, memory), while some edge nodes solely ct s the ccess points in the edge network nd re the entry points of the incoming IoT trffic to the telecom network. Let f(j) = 1, if edge node j hve fog computing cpbility (hence clled fog node), nd f(j) = 0, otherwise. Let the desired QoS level for service be denoted by q (0, 1), nd dely threshold for service by th. Let p denote the penlty tht the SO experiences if ltency requirements of service re violted by 1% per unit time, nd let V % be the percentge of IoT service dely smples of service tht do not meet the dely requirement. or instnce if q = 97%, ny violtion percentge V % greter thn 3% must be compensted for by the SO. The underlying communiction network, set of edge nodes nd cloud servers re modeled s grph G = (V, E), such tht the node set V includes the edge nodes nd cloud servers (V = ), nd the grph edge set E includes the logicl links between the nodes in V. Ech grph edge e(src, dst) E is ssocited with two numbers: r e, the trnsmission rte of logicl link e (megbits per second); nd d e, the propgtion dely of logicl link e (milliseconds). or the ske of correct indexing, if e(src, dst) is the logicl link between n edge node nd cloud server, we renme r e nd d e to r e nd d e, respectively. These prmeters re mintined
by the Physicl Infrstructure Provider (PIP) nd re shred with the SO. The min decision vribles of the QSP problem re the resource lloction vribles, defined in the tble I. A. Optimiztion Problem The QSP problem cn be formulted s the following minimiztion of ost(t) over ll time intervls (strting from t = 0 to t = T end ): P1 : min ost(t), 0 t T end ost(t) = [ proc (t) + proc (t)]+ (1) (t) + stor (t)] + [ mem (t) + mem (t)]+ [ stor [ net (t) + net (t)] + viol (t), Subject to QoS constrints. The cost components re defined below. proc (t) = X k P k P T t, (2) k proc (t) = Xj P j P T t, (3) j stor (t) = (X k S k S + λ in k k R )T t, (4) k stor (t) = (Xj S j S + λ in jj R )T t, (5) j mem (t) = X k M k M T t, (6) k mem (t) = j X M j M j T t, (7) net (t) = j net (t) = j viol (t) = j Y jh (j) u (j,h (j))t t, (8) Y jj u (j,j )T t, (9) j [V % (1 q )] + λ in jp T t. (10) The nottion [.] + in the definition of viol is defined s [x] + = mx(x, 0). k R nd j R re the unit cost of request in cloud server k nd fog node j, respectively (e.g. per 1000 requests). proc nd proc re cost of processing in cloud nd fog, re cost of storge in cloud nd fog, nd mem stor nd stor nd mem re cost of memory in cloud nd fog, respectively. The cost of processing is defined similr to the pricing of compute instnces in Amzon E2 [1] nd the cost of storge is defined similr to the pricing of storge instnces in Amzon S3, which is the storge cost plus the request cost (GET, POST, etc.) [2]. comm is the cost of communiction between fog nd cloud, comm is the cost of communiction between fog nodes, nd viol is the cost (penlty) of dely violtions. A service deployed on fog node my be relesed when the demnd for the service is smll. Therefore, we ssume the services re stteless, tht is they do not store ny stte informtion on fog nodes [3], [4], nd we do not consider costs for stte migrtions. We consider discrete-time system model where time is divided into time periods clled reconfigurtion periods. T is the time intervl between two instnces of solving the optimiztion problem. B. onstrints The constrints of the optimiztion problem re introduced here. 1) Mximum Allowed Slice pcity: The PIP exerts mximum llowed slice cpcity (MAS) for given slice. In other words, the SO cnnot llocte to her slice more resources thn the mximum llowed slice cpcity imposed by the PIP. Let the MAS for compute, storge, nd memory t fog node j be Kj P, KS j, KM j, nd t cloud server k be K P k, K S p, K k M, respectively. Also, let the MAS for networking (bndwidth) between fog node j nd fog node j be Kjj N nd between fog node j nd cloud server k be K jk N. irst, we express the MAS constrints for the nodes in the SO s slice: Xj P < Kj P, j ; X k P < K k P, k ; (11) Xj S < Kj S, j ; X k S < K k S, k ; (12) Xj M < Kj M, j ; X k M < K k M, k. (13) Note tht in certin scenrios the MAS for cloud servers my be very lrge vlue, nd thus we cn relx those constrints in such scenrios. Similr to the MAS constrints for the nodes, MAS constrints for the links in the SO s slice re Y jj K N jj, (j, j ) E, (14) Y jk K N jk, (j, k) E. (15) 2) Trffic Offloding: We im to define the QoS-wre og-supported 5G Slice Provisioning (QSP) problem generl enough nd gnostic to the underlying trffic offloding scheme such tht different offloding frmeworks cn be used. The simplest offloding scheme of fog nodes is no-offlod. In the no-offlod frmework, when request rrives to the fog node, the fog node will process it if its queue is not full, or drops the request if the queue is full. It cn be seen tht no-offlod is not good trffic offloding policy, since requests my be blocked or overloded fog nodes my hve to process more requests which will result in high ltency. Another simple offloding policy is cloud-only-offlod, where offloding hppens only from fog node to the cloud. Another offloding policy is single-neighbor-multi-offlod, when fog node cn offlod trffic to one of its neighbors (e.g. the one with the smllest witing time, s in our prior work [5]) nd offloding cn hppen multiple times mong the fog nodes until it reches the mx-offlod-limit nd is hence offloded to the cloud. One cn envisge mny such offloding policies. In this subsection we select one such offloding policy to show how n rbitrry offloding scheme cn fit in our
x P j 0 x S j 0 x M j 0 k 0 x S k 0 TABLE I VARIABLES O QSP PROBLEM Service Vribles omputing resources llocted to service over the SO s slice on fog node j (in MIPS) Storge llocted to service over the SO s slice on fog node j (in megbits) Memory llocted to service over the SO s slice on fog node j (in megbits) omputing resources llocted to service over the SO s slice on cloud server k (in MIPS) Storge llocted to service over the SO s slice on cloud server k (in megbits) x k M 0 Memory llocted to service over the SO s slice on cloud server k (in megbits) Trffic Vribles rction of trffic for service offloded from fog 0 t jj 1 node j to fog node j 0 t jk 1 rction of trffic for service offloded from fog node j to cloud server k Slice Node Vribles Xj P 0 omputing resources llocted to the SO s slice on fog node j (in MIPS) Xj S 0 Storge llocted to the SO s slice on fog node j (in megbits) Xj M Memory llocted to the SO s slice on fog node j (in 0 megbits) X k P 0 omputing resources llocted the SO s slice on cloud server k (in MIPS) X k S 0 Storge llocted to the SO s slice on cloud server k (in megbits) X M k 0 Y jj 0 Y jk 0 Memory llocted to the SO s slice on cloud server k (in megbits) Slice Link Vribles Networking resources (bndwidth) llocted to the (logicl) link between edge node j nd edge node j in the SO s slice (in megbits per second) Networking resources (bndwidth) llocted to the (logicl) link between edge node j nd cloud server k in the SO s slice (in megbits per second) frmework. We choose nd describe the multi-neighbor-singleofflod, where the trffic from fog node cn be offloded to ll of the fog node s fog neighbors nd lso to the cloud. In multi-neighbor-single-offlod n offloded trffic to fog node cnnot be offloded gin nd must be processed by the receiving node. Let I j denote the incoming IoT trffic for service over the SO s slice t fog node j, nd let λ j j denote the offloded trffic for service from fog node j to fog node j over the SO s slice. Then the totl incoming trffic of fog node j for service over the SO s slice, v j, will be v j = I j + λ j j. (16) j Note tht in multi-neighbor-single-offlod n offloded trffic to fog node cnnot be offloded gin; nonetheless, not ll of I j is ccepted by fog node j for processing. Let us define t jj s the frction of trffic for service tht is offloded from fog node j to fog node j (when j = j, t jj indictes the frction of ccepted trffic by fog node j). Similrly, t jk is the frction of trffic for service tht is offloded from fog node j to cloud server k. Obviously, these frctions should dd up to one for given service t given fog node j, or t jh + (j) t jj = 1, A, j. (17) j The ccepted incoming trffic to fog node j for service over the SO s slice is λ in j = t jj I j + λ j j. (18) j The frction of I j tht is not ccepted by the fog node will be offloded either to nother fog node, or to cloud server. The disptched trffic from fog node j to cloud server k = h (j) for service over the SO s slice is denoted by λ out j nd is derived by λ out j = t jh (j) I j. (19) The incoming trffic to cloud server k for service over the SO s slice is λ in k = I k + λ out j, (20) j H 1 (k) where H 1 (k) set of indices of ll fog nodes tht route the trffic for service to cloud server k, nd I k is the incoming IoT trffic for service over the SO s slice t cloud server k. Similr to Eq. (18) nd Eq. (19), the offloded trffic from fog node j to fog node j for service over the SO s slice is λ jj = t jj I j. (21) 3) og omputing pbility: If edge node j does not hve fog computing cpbilities, it should not ccept ny trffic for processing nd it should offlod ll of the incoming trffic either to fog node or cloud server; tht is, if f(j) = 0, then t jj = 0. Similrly, if edge node j does not hve fog computing cpbilities, other edge nodes j should not offlod ny trffic to this node; tht is, if f(j) = 0, then t j j = 0. These two constrints cn be linerly described by t j j f(j), A, j, j. (22) 4) Service Dely: IoT service dely is defined s the time intervl between the moment when n IoT node sends service request nd when it receives the response for tht request. (The closed-form eqution of IoT service dely is explined in our prior study [6]). To obtin the service dely, we need to hve the verge propgtion dely nd verge trnsmission rte between IoT nodes nd their corresponding fog nodes. These vlues must be known by the SO, nd in some cses cn be pproximted by round-trip dely mesurement techniques. However, since obtining these vlues re not lwys fesible, we chnge the scope of the definition of the IoT service dely to consider the dely only within fog to cloud domins. IoT service dely cptures the dely from the moment n IoT node sends request until it receives the response for tht request. This cn be chnged to cpture the dely
from the moment the request reches fog node. This new service dely, d j, cn be relized s the verge dely budget for service t fog node j within fog-cloud. The verge dely budget for the multi-neighbor-single-offlod scheme for service t fog node j is equl to d j = w j t jj + [2d (j,j ) + lrq + l rp + w j ] t jj + [2d (j,k) + lrq j Y jk Y jj + w k] t jk; k = h (j).(23) Similrly, the verge dely budget for the single-neighborsingle-offlod scheme for service t fog node j is equl to d j = w j t jj + [2d (j,j ) + lrq + [2d (j,k) + lrq Y jk Y jj + w j ] t jj + w k] t jk; k = h (j).(24) where j is the best neighbor of fog node j. The vribles t jj, t jj, t jk, Y jj, nd Y jk re clculted by the QSP problem. In order to evlute the dely budget, we need to hve the verge size of requests nd reply of service (l rq nd l rp ) nd the propgtion dely between fog nodes nd from fog nodes to cloud servers (d (j,j ) nd d (j,k)), which re known or mesured by the SO. We lso need the verge witing times (queueing time plus processing time), w j, w j, nd w k, which cn be obtined either from the corresponding M/M/c queueing models of the fog nodes nd the cloud servers (discussed in Section I-B8), or bsed on predictive performnce modeling nd blck-box monitoring techniques [7]. In either cse, the incoming trffic to fog nodes is required to obtin the verge witing times. The incoming trffic to the fog nodes (I j ) cn either be directly monitored by the SO (e.g. using the monitoring gent of n SDN controller [6]) or cn be predicted by the SO using lerning pproch (to be discussed in Section II). In the first pproch where I j is monitored (t the beginning of ech configurtion intervl), the QSP problem is solved in rective nture; wheres the second pproch tht predicts I j is proctive, since it provisions the resources hed of time bsed on the estimted incoming IoT trffic. Moreover, the predicting pproch hs the dvntge of getting n verge of I j during configurtion intervl, s pposed to the monitoring pproch tht only obtins n instnce of I j during configurtion intervl. In the Section II we discuss how employ lerning methods to predict I j nd obtin the witing times of the fog nodes nd cloud servers. 5) SLA Violtion: To mesure the qulity of given service, we need to see wht percentge of IoT requests do not meet the dely threshold th (SLA violtions). We first need to check if verge dely budget of fog node j for service is greter thn the threshold th defined in SLA for service. Let us define binry vrible v j to indicte this: 1, if d j > th v j =, j, A. (25) 0, otherwise We define nother vrible tht mesures the SLA violtion (SLAV) of given service ccording to the defined QoS prmeters in the SLA. We denote by V % the percentge of IoT service dely smples of service tht do not meet the dely requirement. V % cn be clculted s follows V % = j λin j v j, A. (26) j λin j Note tht V % is mesured s weighted verge of v j, with λ in j s the weight. 6) 5G Slice Resource pcity: The mount of the resources llocted to the services nd trffic within the SO s slice should not exceed the cpcity of the slice; this pplies to ll the nodes nd the links of the slice. We first look t this constrints with respect to the nodes in the slice, i.e. compute, storge, nd memory resources within node of the slice: x P j < Xj P, j ; x S j < Xj S, j ; x M j < Xj M, j ; k < X P k, k ; (27) x S k < X S k, k ; (28) x M k < X M k, k. (29) Similrly, we lso hve constrints for cpcity of the links in slice: the mount of trffic routed over prticulr slice link should not exceed the mount of llocted bndwidth to tht slice link, which is expressed by λ j j(l rq ) Y jj, (j, j ) E, (30) λ out j (l rq + l rp ) Y jk, (j, k) E. (31) 7) Arrivl of Requests: Let Λ j denote the rrivl rte of instructions (in MIPS) to fog node j for service over the SO s slice. This is the rrivl rte of instructions of the incoming requests tht re ccepted for processing by fog node j tht is given by Λ j = L P λ in j. Similrly, the rrivl rte of instructions (in MIPS) to the cloud server k for service over slice the SO s slice, Λ k, cn be written s Λ k = LP λ in k, where λ in k is the incoming trffic rte to cloud server k for service over the SO s slice. L P is the required mount of processing for service per request. The mount of computing resources llocted to the fog nodes nd cloud servers for given service on the SO s slice must be bigger thn the incoming rrivl of processing requests for tht service (i.e. stbility constrints): x P j Λ j, j, A, (32) k Λ k, k, A. (33)
8) Witing Times: To get the witing times, We dopt commonly used M/M/c queueing system [8], [9], [10] model for fog node with n j processing units, ech with service rte µ j nd totl rrivl rte of Λ j (totl processing cpcity of fog node j will be Kj P = n j µ j ). To model wht frction of processing units ech service cn obtin, we ssume tht the processing units of fog node re llocted to the deployed services proportionl to their processing needs (L P ). or instnce, if the requests for service 1 need twice the mount of processing thn tht of the requests for service 2 (L P 1 = 2 L P 2 ), service 1 should receive twice the service rte compred to service 2. orrespondingly, we define f j, the frction of service rte tht service obtins t fog node j s: f j = x j L P x jl P. (34) Ech deployed service cn be seen s n M/M/c queueing system with service rte of f j Kj P = f j n j µ j, nd rrivl rte of Λ j (both in MIPS). Thus, the witing time for requests of service t fog node j will be w j = 1 f j µ j + f j K P j P Q j Λ, (35) j where P Q j is the probbility tht n rriving request to fog node j for service hs to wit in the queue. P.. Q is lso referred to s Erlng s formul nd is equl to such tht ρ j = P 0 j = [ nj 1 c=0 P Q j = (n jρ j ) nj n j! Λj f jk P j nd (n j ρ j ) c c! + (n jρ j ) nj n j! P 0 j 1 ρ j, (36) 1 1 ρ j ] 1. (37) Note tht the requests for different services hve different processing times. Nevertheless, s discussed before, in the definition of Λ j we ccount for the different processing times by the inclusion of L P. Similrly, cloud server k with n k processing units (i.e. servers), ech with service rte µ k nd totl rrivl rte of Λ k cn be seen s n M/M/c queueing system (totl processing cpcity of cloud server k will be K k P = n k µ k ). Therefore, similr to Eq. (35), w k, the witing time for requests of service t cloud server k, could be derived s w k = 1 f k µ k + P Q k f k K P k Λ k, A, k. (38) An eqution similr to Eq. (36) is defined for P Q k, the probbility of queueing t cloud server k. Note tht for simplicity, insted of modeling ech cloud server M/M/c queue, one my lso model the whole cloud s n M/M/ queueing system. 9) Dependency of ompute, Storge, nd Memory: The vribles for compute, storge, nd memory lloction re interrelted to ech other; when there is no compute resources llocted to node for prticulr service, there is no need to llocte corresponding storge nd memory resources to the node for tht service. These dependencies for fog nodes cn be formulted using the following constrints: x S 0, if x P j j = = 0 L S, j, A, (39), otherwise x M 0, if x P j j = = 0 L M, j, A, (40), otherwise nd for cloud servers using the following constrints: x S 0, if k k = = 0 L S, otherwise x M k =, k, A, (41) 0, if k = 0 L M,, otherwise k, A. (42) L S nd L M re the required minimum mount of storge nd memory for service, respectively. It is worth mentioning tht since cloud is usully regrded s hving mple resources, the constrints (41) nd (42) could be esily modified to llocte more storge nd memory resources for cloud servers when k > 0. II. LEARNING METHOD OR INOMING IOT TRAI A. Predicting Incoming IoT Trffic To predict the incoming trffic to Iot nodes, we employ the follow the regulrized leder (TRL) online lerning method [11]. Our method for trffic prediction is minly influenced by the demnd prediction method discussed in [12]. To ccount for the mount of incoming trffic in different configurtion intervls, we dd time vrible t to the vrible I j, mke it I j (t), the predicted incoming IoT trffic for service over the SO s slice t fog node j in the configurtion intervl t. Let Ij (t) be the ctul (e.g. mesured) incoming IoT trffic for service over the SO s slice t fog node j in time t. Let us define Ij mx = mx t [0,T end ](I j (t)), which is known to or cn be mesured by the SO bsed on historicl observtions. Hence, the predicted incoming trffic I j (t) will be in [0, Ij mx ]. We cn then divide the time into configurtion intervls so tht we hve countble number of I j (t) s. We define convex loss function to minimize the prediction error for the incoming IoT trffic I j (t) s E t (I j ) = I j (t) I j(t), j, A, t. (43) Therefore, loss minimiztion problem over ll time slots then cn be expressed s min E t (I j ). (44) 0 t T end The solution to this minimiztion problem is set of vlues for I j (t) [0, I mx j ] for every t [0, T end ].
TRL chooses the predicted incoming trffic vlue I j (t+1) tht minimizes the cumultive loss function over the previous t time slots plus regulrizer (to void overfitting), tht is I j (t + 1) = rg min I j [0,Ij mx][ s=0 t E s (I j ) + R(I j )]. (45) where I j reflects the possible predicted incoming IoT trffic rte for service to fog node j in intervl t + 1. R(.) is convex regulrizer. Since solving this optimiztion problem t ech reconfigurtion intervl is not fesible, we convert it to the following problem I j (t + 1) = rg min I j [0,I mx j ] h t(i j ), (46) where h t (.) is the surrogte loss function tht should ccurtely estimtes the originl loss function nd is efficient to evlute. We use the surrogte function chosen in [12] s h t (I j ) = [ t ( E s (I j )] I j + r t (I j ). (47) s=0 E s (I j ) is the grdient of E s (I j ) nd r t (.) is regulriztion function (similr to R(.)). The regulriztion function is chosen s r t (x) = 0.5 t s=0 ( 1 α s 1 α s 1 )(x x(s)) 2, (48) where α s = Θ( 1 s ) is e lerning rte in time slot s [12]. Specificlly, we will hve β α t = t, (49) s=0 ( E s(i j )) 2 nd since ( E s (I j )) 2 = 1, then α t = β t+1, where β is hyper prmeter bsed on the fetures nd dt (set to β = Ij mx ) [12]. After this conversion, eq. (46) becomes smooth minimiztion problem with no constrints. We then simply tke the derivtive of our chosen surrogte function h t (.) with respect to I j nd set it to be equl to 0 to obtin the I j (t) tht minimizes eq. (46). Tking derivtive is esy nd its solution will be the predicted incoming IoT trffic. B. Obtining Witing Times Now hving the predicted incoming IoT trffic, one cn lso predicts the witing times. As discussed before, the online lerning pproch tht predicts I j is proctive in nture, since bsed on predicted witing times (nd hence violtion rte) the SO provisions the resources hed of time bsed. Bsiclly, the predicted incoming IoT trffic cn be simply plugged in the equtions for witing times to get the predicted witing times. Note: We leve the investigtion of other choices for surrogte functions nd regulriztion functions s our future work. More noteworthy, one cn lso study the pplicbility of reinforcement lerning techniques for incoming IoT trffic prediction. III. TESTBED SETUP or obtining numericl evlutions nd show the vlidity nd performnce of our proposed frmework, we implement RUGAL in the following testbed. We hve 4 SDN-enbled HPE switches in our lbortory, which we use to mke tree topology: 1 switch s the root nd 3 switches s the lef nodes. We directly connect one computer to ech lef switch to ct s fog node. hence, we hve 3 fog nodes in our testbed. We lso connect Wii ccess points to the switches to enble the pssing trffic from the Wii-enbled IoT devices. The RUGAL runs s n pp (e.g. on the sme computer tht hs the root SDN controller) tht bsiclly communictes with the SDN controllers through Openlow protocol, to monitor the IoT trffic, deploy or relese fog service, or dimension the compute, storge or networking resources in the testbed. The RUGAL solves the QSP problem nd determine wht resources need to be deployed, relese, or dimensioned. The experiment will consist of 2 mjor components: (1) run few IoT pplictions (e.g. object recognition from cmer imges) whose trffic pss through these switches in our lb to some service tht is hosted in the cloud nd (2) nlyzing the performnce of RUGAL with regrds to service dely, dely violtions, nd resource usge. The switches, ccess points, nd computers in our lb ct s the edge of the network, where fog nodes locte nd operte. or the cloud, we cn use ny cloud service provider such s AWS, Google, or reserch experiment cloud computing environments such s hmeleon loud or loudlb. The cloud/fog services could be implemented s continers or VMs. or the gile service provisioning in the fog network, we cn use Docker continers tht cn be esily provisioned using kubernetes or OpenStck. kybernetes nd OpenStck gives us the flexibility nd gility for resource mngement (deploying, relesing or dimensioning of compute, storge, nd networking resources). To generte IoT trffic, we hve cn do the following: (1) use the IoT bords we lredy hve in our lb (e.g. Rspberry Pi nd Arduino tht hve cmer nd other sensors) to generte ctul trffic, (2) use the Iperf tool to generte constnt bit rte trffic flows, (3) combintion of both pproches: use the IoT bords to generte ctul trffic, nd if needed more, use the Iperf tool to generte more trffic (e.g. for other services or bckground trffic). REERENES [1] Amzon E2 pricing, 2018. [Avilble] https://ws.mzon.com/ec2/pricing/. [2] Amzon S3 pricing, 2018. [Avilble] https://ws.mzon.com/s3/pricing/. [3] S. H. Mortzvi, M. Slehe,. S. Gomes,. Phillips, nd E. de Lr, loudpth: multi-tier cloud computing frmework, in Proceedings of the Second AM/IEEE Symposium on Edge omputing, p. 20, AM, 2017. [4] R. S. Montero, E. Rojs, A. A. rrillo, nd I. M. Llorente, Extending the cloud to the network edge., IEEE omputer, vol. 50, no. 4, pp. 91 95, 2017. [5] A. Yousefpour, G. Ishigki, R. Gour, nd J. P. Jue, On reducing IoT service dely vi fog offloding, IEEE Internet of Things Journl, vol. 5, no. 2, pp. 998 1010, 2018.
[6] A. Yousefpour, A. Ptil, G. Ishigki, I. Kim, X. Wng, H.. nky, Q. Zhng, W. Xie, nd J. P. Jue, QoS-wre dynmic fog service provisioning, rxiv preprint rxiv:1802.00800, 2018. [7]. Witt, M. Bux, W. Gusew, nd U. Leser, Predictive performnce modeling for distributed computing using blck-box monitoring nd mchine lerning, rxiv preprint rxiv:1805.11877, 2018. [8] M. Ji, J. o, nd W. Ling, Optiml cloudlet plcement nd user to cloudlet lloction in wireless metropolitn re networks, IEEE Trnsctions on loud omputing, 2015. [9] Z. Zhou, J. eng, L. Tn, Y. He, nd J. Gong, An ir-ground integrtion pproch for mobile edge computing in iot, IEEE ommunictions Mgzine, vol. 56, no. 8, pp. 40 47, 2018. [10] L. Liu, X. Guo, Z. hng, nd T. Ristniemi, Joint optimiztion of energy nd dely for computtion offloding in cloudlet-ssisted mobile cloud computing, Wireless Networks, pp. 1 14, July 2018. [11] B. McMhn, ollow-the-regulrized-leder nd mirror descent: Equivlence theorems nd l1 regulriztion, in Proceedings of the ourteenth Interntionl onference on Artificil Intelligence nd Sttistics, pp. 525 533, 2011. [12] X. ei,. Liu, H. Xu, nd H. Jin, Adptive vnf scling nd flow routing with proctive demnd prediction, in IEEE INOOM 2018- IEEE onference on omputer ommunictions, pp. 486 494, IEEE, 2018.