On-path Cloudlet Pricing for Low Latency Application Provisioning

Transcription

1 On-path Cloulet Pricing for Low Latency Application Provisioning Argyrios G. Tasiopoulos, Onur Ascigil, Ioannis Psaras, Stavros Toumpis, George Pavlou Dept. of Electronic an Electrical Engineering, University College Lonon Dept. of Informatics, Athens University of Economics an Business {argyrios.tasiopoulos, o.ascigil, i.psaras, Abstract Clou computing has been tremenously successful in proviing a commercial infrastructure for hosting computationally intensive applications. Nevertheless, an increasing number of Low Latency Applications (LLAs) notably in the entertainment, IoT, an automative omains require response times much smaller than the supporte ones by the typical client-to-clou network moel. Cloulets have been introuce as ata centres in a box, for bringing computing resources closer to the en users. As a result, LLAs can take avantage of cloulets to improve their Quality-of-Service (QoS) by reucing the unerlying response times between their users an application instances location. In this work, we stuy the emerging market of stateful LLAs provisioning over geo-istribute thir-party cloulets. We assume that cloulets offer their resources in the form of Virtual Machines (VMs) via collocate markets. Forwaring requests for LLAs interact with cloulet markets for performing on-path an on-eman resource provisioning. We introuce a pricing scheme where users pay a fixe price for each time unit of their engagement to an LLA instance. Our evaluation on realistic topologies an application requests emonstrate the merits of on-eman provisioning when accompanie by a payas-you-go pricing scheme. I. INTRODUCTION Clou computing is the prominent cost-efficient infrastructure for elivering computationally emaning applications to en-users. Clous abunance of resources enables them to elastically cope with changes in applications eman in a scalable manner via on-eman computation. That is, Clous exploit economies of scale for ecreasing their running cost [7] while eploying usage-base pricing mechanisms for controlling the eman for their computing resources. However, an increasing number of applications (like augmente reality, automotive, health monitoring etc.) require low response times, renering the current clou-base infrastructure unfit for the purpose of application provisioning. We refer to these as Low Latency Applications (LLAs) since they strongly rely on the latency of the network for achieving a satisfying Quality-of- Service (QoS). Cloulets have been propose as ata centres in a box that can be eploye throughout the Internet, for bringing computing resources closer to the en users [12]. That is, cloulets have the potential of improving the network conitions by ecreasing the Roun-Trip-Time (RTT) between en-users an application instances when hoste at cloulets points-ofpresence. Clearly, cloulets can serve as an alternative com /18/$31. c 218 IEEE Fig. 1: On-path application provisioning setting puting infrastructure, that augments as well as complements the centralise provisioning at the clou, for enabling LLAs. Nevertheless, cloulets are incapable of proviing the essentially bounless elasticity of the clou, meaning that occasionally the eman for resources at a given cloulet can excee their availability. That is, resource allocation mechanisms have to prioritise the requests that benefit the most when serve by a cloulet. Especially in the case of stateful LLA provisioning, resource allocation is a challenging task since once a request is accepte an a user is engage to an application instance, the establishe application session cannot be isrupte for reallocating the involve resources. The reason is that session isruptions can cause a user to quit the LLA, which is consiere the worst possible outcome in terms of QoS. In this work we investigate the emerging market of provisioning heavily stateful LLAs over thir party cloulets. In our context, heavily stateful LLAs are efine as the ones that cannot be migrate to another cloulet location, ue to the size of their runtime ata generate by their users, without seriously amaging their QoS [4]. We consier that computation is available either at the ege or mile-tier locations of the network, in the form of cloulets, an/or at istant clous/ata centres (Fig. 1). We argue that application provisioning over cloulets is expecte to take place in a ecentralise an uncoorinate environment. Given that cloulet resources are limite, the key challenge is to create a market that operates on a per-request basis for offering the finest possible resource allocation granularity. We aim to provie answers to the funamental questions of: i) how shoul cloulet resources be allocate over time to ifferent applications/services? an

2 ii) how much shoul a cloulet charge an application? Here, we present a ecentralise pricing mechanism that answers both questions, while aressing the challenges of ynamic application provisioning over an infrastructure of thir-party cloulets. We introuce a pricing mechanism that associates each cloulet with a price for on-path, on-eman, uncoorinate LLA provisioning. In our context, cloulets offer their resources in the form of Virtual Machines (VMs) via collocate markets. As LLA requests are forware towars a efault execution location, i.e., at the Clou, they interact with onpath cloulet markets. If the price of a market is below the gain the LLA woul have if it is serve, an available VM is allocate to serve the request; otherwise, the request is rejecte an continues its journey towars the Clou. For example in Fig. 1, the requests of User 1 (User 2) attempt to provision resources on-path at Cloulet 1 (Cloulet 2) followe by Cloulet 3 before getting serve by the Clou as the last available provisioning infrastructure. Our esign aresses explicitly the heavily stateful requirement of seamless users engagement to LLA instances while setting the price of each cloulet base on its resource utilisation. To this en, the main technical contributions of this paper are as follows: 1) We stuy the emerging market of heavily stateful LLAs provisioning over inepenent cloulets from an economics point of view. 2) We evelop a pricing mechanism for on-path, on-eman LLA provisioning, taking into account the cloulets computing resource limitations an the en users requirement for seamless LLA engagement. 3) We evaluate the merits of our mechanism in an ISP topology uner trace-base application requests. II. DESIGN RATIONALE & SYSTEM MODEL A. Design Rationale We aim to esign a market mechanism tailore to the provisioning of LLAs over an uncoorinate cloulet infrastructure. Our esign has to aress explicitly the application provisioning challenges of i) cloulets resource iscovery, since cloulets points-of-presence are expecte to excee by far the number of clous, an ii) the cloulets limite elasticity of resources. Given that application requests are forware in the network towars a istant ata centre, we argue for both on-path an on-eman application provisioning. That is, by applying onpath provisioning, there is no nee for a centralise resource iscovery process since resources are iscovere in real-time opportunistically. Furthermore, ue to on-eman provisioning, cloulets o not waste resources since each allocate VM serves an application request. We argue that these esign choices enable the most promising an incrementally eployable conitions for the problem of application provisioning we tackle; proviing an uncoorinate solution. However, oneman provisioning has to be accompanie by a pricing scheme that controls the eman at each cloulet, an this nees to happen in a way that the requests that have the highest provisioning gain at a specific location are given priority. Clou resources are offere in terms of remote instances, i.e., virtual machines (VMs), with eicate CPU, memory, an storage resources, as an Infrastructure-as-a-Service (IaaS) [2]. For example as a clou resource provier, Amazon offers instances uner three pricing schemes: 1) Reserve instances guarantee the long-term availability of an instance, i.e., for more than a year, by charging fixe usage-base prices. 2) On-eman instances guarantee the short-term availability of an instance, i.e., for an hour, by also charging fixe usage-base prices. 3) Spot instances create an auction-base market for spare instances. In etail, users can use spot instances only if their bis excee a spot price while spot prices are upate every 5 minutes; meaning that the availability of instances is not guarantee. Offering cloulet resources as reserve an/or on-eman instances fails to capture eman fluctuations uner the boune instances elasticity. Furthermore, spot instances are suitable for interruptible jobs, since a task is execute as long as its bi excees the current spot price; otherwise, the task is suspene. Despite some efforts in strategically biing spot instances for uninterrupte tasks to the clou [16], we suggest that a cloulet pricing plan shoul inclue the applications seamless execution in its inherent characteristics, especially in the case of heavily stateful applications. In this work we argue that cloulets pricing schemes shoul follow a pay-as-you-go structure in terms of application user engagement uration. In other wors, the LLA proviers shoul be charge base on the time their users occupy a cloulet instance. B. System Moel We consier a set S {1, 2,..., S} of LLAs an a set D {1, 2,..., D} of cloulets. We assume that each LLA s S can serve its users eman at a istant clou/ata centre, enote by D s, whose capacity is sufficient for serving the total number of LLA requests it receives. Application users are connecte to the network via a set of P {1, 2,..., P } access points. We assume that the network conitions to a cloulet are access-point specific. That is, a user connecte to access point p while requesting an LLA s S experiences a QoS improvement u s,p, in terms of network conitions, for each secon that the en user remains engage to an LLA instance at cloulet D, as oppose to the efault execution clou D s. The notation followe in throughout the paper is presente in Table I. III. PAY-AS-YOU-GO MECHANISM In this section, we escribe the iniviual components of the propose pay-as-you-go mechanism, namely i) the oneman provisioning component, an ii) the utilisation-base price erivation component. Given a price assignment at each cloulet, the on-eman provisioning component efines the real-time interaction of LLA requests with the collocate

3 System Moel S Set of LLAs. D,D s Set of cloulets, efault ata centre of LLA s. P Set of requests access points. u s,p Per secon QoS gain of LLA s at cloulet for request from access point p, in terms of network conition. Charging Mechanism π Price of cloulet. b s,p Bi of request prouce at access point p for LLA s at cloulet. ρ min, ρ max Target minimum an maximum utilisation of resources at. π Price ecrease step. TABLE I: Notation Fig. 2: On-Deman Provisioning Charging Scheme Overview cloulets markets, as the requests are forware in the network. On the other han, the price erivation component is responsible for eriving a price for each cloulet given the resource utilisation achieve at each cloulet. A. On-Deman Provisioning An overview of the on-eman provisioning scheme is epicte in Fig. 2, where we consier an en user request of LLA s S that arrives at the market of cloulet D (step a). The interaction of the request with each market is characterise by the following properties: A1) The request is for a single VM. A2) The request is associate with a bi, b s,p, that expresses the user s willingness to pay for a VM for each engagement time-unit of the user to an LLA s instance at when her access point is p. A3) The request is not queue at the market, i.e., the request is either serve or rejecte immeiately. A4) The VM allocation time overhea has no impact on the LLA s QoS. In more etail, let the price for the use of a VM at cloulet be π. Upon the arrival of the request the market operates accoring to the following rules: R1 If there are no available VMs, the request is rejecte. R2 Else, if there are available VMs: If b s,p π, a VM is allocate to serve the request at price π for each time unit of user engagement. Else, if b s,p < π, the request is rejecte. Assuming b s,p π an the existence of available VMs, the allocation is taking place an the cloulet starts a timer for keeping track of the user s engagement uration (step b). After the session completion/application termination, the cloulet informs the market about the engagement uration (step c). Then the market verifies the uration an converts it to a bill that equals the engagement uration times the agree per time unit service price π (step ). Finally, the bill is sent to the corresponing LLA provier which eventually pays the cloulet provier for its service. 1 Next, we explain how the request bis as well as the cloulet prices are erive. B. Request Bis an Cloulet Prices erivation Let u s,p be the per-time-unit QoS gain of LLA s when a request is coming from access point p to get serve at cloulet, instea of its efault clou. Biing truthfulness is efine in the following straightforwar way: Definition 1. A bi b s,p for a request arriving at cloulet from access point p for LLA s is truthful iff b s,p := u s,p, where u s,p is the per-time-unit QoS gain of the requeste LLA at. In other wors, a bi is truthful when it equals the actual gain of the serve application. Then, given the cloulet s price, π, the net utility rate of the involve request is efine as: Definition 2. The net utility rate of class p LLA s from bi b s,p an QoS gain u s,p is: { u s,p π, if b s,p π, net utility rate =, otherwise. Proposition 1. Uner rules R1-R2, the request bis are truthful. Proof. We investigate the following possible cases: If u s,p π then the bier woul win the item with a truthful bi as well as an overbi, i.e., b s,p u s,p. If u s,p π > b s,p unerbiing returns a net utility rate equal to, as oppose to a truthful bi for which the net utility rate is non-negative. If π > u s,p then the bier woul loose the item with a truthful bi as well as an unerbi, i.e., u s,p b s,p. If b s,p > π > u s,p, then overbiing woul return a negative net utility rate, as oppose to a truthful bi for which the net utility rate is. From the previous cases, it is clear that truthful biing, i.e., b s,p = u s,p, is the ominant strategy of the involve request. Because the bis must be truthful for a rational user, the strategy of their eployment is straightforwar, i.e., assigning a price b s,p = u s,p to each request. Therefore, a request has to simply be associate to its involve QoS per time unit gain at a given cloulet. Lastly with respect to cloulet price erivation, we assume that each cloulet perioically (every w secons) observes the utilisation ρ ol of its resources that was achieve at the 1 The technical etails of the explaine process, relate to the security, verification, etc., are beyon the scope of this paper although orthogonal to its contribution.

4 previously set price π ol. Then, given a minimum target utilization ρ min the price is ajuste accoring to: π ol π, if ρ ol < ρ min, π = π ol + π, if ρ ol > ρ max, π ol, otherwise. where π > is the price ecreasing increment. Note that as a price ecreases (increases) the number of requests that can be serve at cloulet rises (rops) with the result of increase (ecrease) in the utilisation of s resources. IV. PERFORMANCE EVALUATION In this section, we emonstrate s performance via simulations in an ISP topology. We implement the an the baseline approaches, escribe next, by extening the Icarus caching simulator for application provisioning problems [11]. Application categories for LLAs: The QoS of the LLAs is expresse as a ecreasing function of the en users perceive latency to each cloulet [5], [15]. Similarly to [8] for abstract resource allocation gains, we consier that each LLA is characterise by a ecreasing QoS function of the latency x having the general form: ( umin ( u(x) = + 1 u )( min 1 x l ) 1 ) min α u max (1) u max u max l max The constants u max (u min ) represents the maximum (minimum) QoS that the application user can achieve at the minimum (maximum) latency l min (l max ), i.e., u(l min ) = u max an u(l max ) = u min. We set u max = 1 an l min = 5 ms 2 when is not specifie ifferently. Moreover, the function u( ) is convex for < α 1; that is, we set α =.2 since LLAs QoS is expecte to be more sensitive to latency changes closer to l min. Base on Eq. 1, we create ten LLA categories, each associate to a QoS function that moels a certain sensitivity of the LLA to latency. For each LLA category we assign a ifferent u min value from the set of {, 1, 2,..., 9}. In this way, istinct application categories have iverse QoS gains from being provisione at a cloulet, varying, for example, from 1, for u min = 9 an less-latency sensitive LLAs, to 1, for u min = an latency-critical LLAs, when provisione at the ege, i.e., x = l min. We consier ten LLA categories to be sufficient for the purpose of our evaluation since this number is comparable to the currently consiere types in the context of IoT [3], [1] an Tactile Internet [6]. Lastly, we assume that users remain engage to their LLA instance on average for 1 minute while we execute the following experiments for the uration of three hours in our simulator. LLA Realistic Requests Generation: Each user selects one of the 1 LLA categories base on the sequence of request arrivals in Google s cluster ataset 3. In etail, we associate 2 With recent avances in LTE technology, mobile operators reporte hanset-to-base-station latencies aroun 2 msec (RTT of 5 ms), see: 3 Available at each LLA category to a ParentID fiel, that ientifies the service, of the 1 most popular services in the ataset accounting for more than 2K requests. Then by selecting ranom time intervals in the perio of the seven hours that the ataset covers, users request LLA categories base on the sequence of ParentID fiels that arrive into Google s cluster. ISP Topology & Cloulets Deployment: We evaluate uner the Tiscali topology of the Rocketfuel ataset [13] that contains 24 noes an 44 links. Among the 24 noes, we esignate as hosts the 8 noes that present a egree of one, i.e., they are connecte to only a single noe of the topology. After that, we place the Clou at the noe with the highest closeness centrality with respect to all hosts, i.e., shortest average istance to all hosts. Then we set each host to generate one request per secon forware to the Clou via the shortest path in terms of latency. We place a cloulet to each noe of the topology that belong to at least a single shortest path from a host to the Clou. That is, we have 79 cloulets all over the topology for reucing the Clou RTT latency that on average is 155 ms to each host. The efault number of available VMs per cloulet is 2, while we set the minimum an maximum target utilisation of each cloulet to be equal to ρ min =.85 an ρ max =.95 respectively. Comparison approaches: We compare against the following approaches: : Cloulets assign their VMs to application instances accoring to their eman that is consiere known in avance, e.g., an equal number of VMs is assigne to each LLA in the case of an uniform requests istribution. Least Frequently Use () provisioning: Cloulets perioically assign their VMs to LLA instances, prioritising the ones with the highest observe eman, i.e., popularity. In this case, VMs are allocate first to the most popular application, an the process continues with secon most popular application, an so on, until all the VMs at a cloulet are allocate. The number of VMs assigne to each application is etermine accoring to the number of both the requests serve an rejecte. Provisioning: Cloulets perioically assign their VMs to LLA instances by prioritising the ones with the highest QoS gain. In particular, this strategy iterates through the applications sorte in ecreasing orer of QoS gain, an assigns a number of unallocate VMs to the current application. This results with VMs being assigne to applications that are willing to pay the most. The number of VMs assigne to each application again epens on its observe eman. The Static approach is a form of proactive provisioning, because it relies on prior knowlege of the requests eman for each application, while an Self-tuning approaches perform reactive provisioning, as each cloulet perioically reassigns their VM instances to iniviual LLAs base on the request patterns receive uring the most recent observation perio. We set the efault length of the observation perio to 5 minutes, i.e., 3 secons. Note that we consier a single LLA per LLA category in orer to give an avantage to

5 Average QoS Gain Number of VMs (a) QoS Percent Ile Time Number of VMs (b) Ile Time Percentage Revenue (X $) 1.6x x x1 8 1x1 8 8x1 7 6x1 7 4x1 7 2x Number of VMs (c) Revenue Fig. 3: The impact of the number of VMs on the performance of the provisioning mechanisms. the above-mentione approaches against s oneman provisioning, otherwise their performance coul be arbitrarily eteriorate. For example by increasing the number of LLAs per category from 1 to 1, the utilisation of each cloulet woul ecrease to the 1% of the currently achieve utilisation for the Static, Self-tuning, an provisioning with a negative impact on the achieve QoS gain as well as cloulets revenue. Lastly, in orer to perform a straightforwar an fair comparison against, we assume a charging mechanism, ientical to, where each cloulet is associate to a price that equals the lowest QoS gain over all the accepte LLA request classes at a given location. In that way, we avoi creating application-specific prices that challenge the requests biing truthfulness since they woul pay ifferent prices for ientical VMs. A. Impact of the Number of VMs In Fig. 3, we epict the performance of the LLA provisioning approaches for an increasing number of VMs per cloulet for the metrics of: i) users average QoS gain (u(x)), ii) percentage of ile time experience by the VMs of the cloulets, an iii) the total revenue obtaine by all cloulets. User requests serve by the back-en Clou experience the maximum latency (i.e., l max ), an thus obtain zero gain. The revenue obtaine by a cloulet is compute as the prouct of the engagement time of the serve users at each VM an the per-secon usage price of the VM etermine by a cloulet. For the mechanism, the per-secon usage price of the VMs at a cloulet is erive perioically base on the ifference between the observe an the target utilisation, as escribe in Section III-A. For the rest of the strategies, we assume a secon-price auction mechanism, where the winning biers pay for the next highest bi, an the bis are set to the QoS gain of the users as in a true value biing. As shown in Fig. 3a, the mechanism can achieve a higher average QoS gain than the other strategies. Specifically, the mechanism is outperforme since it ignores the utilisation of resources when provisioning the LLAs. As a result, a larger portion of the requests are serve by the Clou without obtaining any QoS gain. On the other han, an Static approaches provision LLAs purely base on utilisation, without consiering the QoS gain of the users, resulting in lower average QoS gain. Our QoS finings are in accorance with the average ile time percentage of the resources, as it is emonstrate in Fig. 3b. As expecte, Static provisioning presents a higher utilisation of resources, i.e., lower ile time, compare to an Self-tuning by exploiting the knowlege of LLAs upcoming eman. However, since Static,, an Selftuning approaches operate by assigning VMs to LLA instances perioically, they are outperforme by the mechanism that provisions requests as soon as they arrive, in an on-eman manner. As a result of achieving higher QoS gain an utilisation of resources, the mechanism obtains the highest revenue, as epicte in Fig. 3c. As the number of VMs increases, the ile time of the strategies raises by aroun 15%. At the same time the average QoS gain of the users presents an upwar tren, which leas to higher bis for resources followe by greater revenues. We observe that the approach obtains higher revenue than the an nearly the same revenue as Static provisioning. B. Impact of Observation Perio Length In the experiments presente earlier, we use a efault observation perio length of 3 secons (5 minutes). In Fig. 4, we emonstrate the impact of this parameter on QoS gain (left plot) an ile time percentage (right plot) of the provisioning mechanisms. First of all, we observe that the performances of both an Self-tuning approaches are increasing function of the observation perio uration before being stabilise at 2 secons. This is relate to the fact that an Self-tuning rely on the estimation of LLAs eman, that is inaccurate for short observation urations in the trace of requests we are using since it fails to capture the eman fluctuations. In terms of QoS gain, the Self-tuning mechanism achieves a remarkable increase of 33% in the uration interval to 2 secons. At the same time, the mechanism achieves nearly 5% increase in resource utilisation since the ile time percentage rops from 4%, for a uration of secons, to 2%, for 2 secons. The mechanism is affecte the least by the observation perio factor while constantly achieving the highest QoS gain an lowest ile time percentage compare to the rest of approaches. Our evaluation emonstrates that the mechanism can provision resources with high utilisation while pro-

6 Average QoS Gain Observation Perio Length (sec) Percent Ile Time Observation Perio Length (sec) Fig. 4: Impact of observation perio length on QoS gain of users (left) an average ile times of VMs (right). viing high QoS gains for the en-users. The price erivation mechanism of our on-eman provisioning mechanism leas to higher revenues than the provisioning mechanisms that perioically allocate the VMs to LLAs base on the observe eman. V. RELATED WORK Cloulets have been propose in [12] with the initial purpose of acting as a surrogate infrastructure where mobile evices can offloa intensive tasks as a response to their computing an battery limitations. Specifically, [1] aresses the problem of application-specific task offloaing over the fog infrastructure, i.e., a task requires the corresponing virtual instance of the application at a cloulet before getting offloae. However, the authors only consier stateless applications, while ignoring completely the economic aspects of such a problem. In [9], the authors propose a self-tuning application provisioning mechanism using a combinatorial auction mechanism, where application proviers bi perioically for a specific number of VMs organise in ifferent execution zones. Nevertheless, the presente mechanism relies on precise preictions about the future eman of an application an takes place in a centralise way that challenges the scalability of such a scheme. Closer to our setting, [14] iscusses LLA provisioning over thir-party cloulets in an economic context for lightly stateful applications, which can be migrate between cloulets without severely egraing the involve QoS for mobile users (as a result of the han-offs between ifferent cells). Similar to spot pricing mechanisms (e.g., [16], our approach can react to transient changes in eman, but without either suspening applications or collecting (i.e., queuing) user bis to execute an auction. VI. CONCLUSIONS An increasing number of network applications with low latency requirements rener clou provisioning unfit for purpose. In this paper, we investigate the emerging market where heavily stateful low latency applications lease thirparty cloulet resources for provisioning their instances. This is a promising research irection for bringing application instances closer to their users in orer to improve the involve perceive latency. Along these lines, we introuce a charging scheme as an on-path, on-eman market base provisioning mechanism. In the propose scheme, cloulets are associate with a single price that is ajuste base on their target resource utilisation. Then LLAs requests forware in the network interact on-path with collocate markets to the cloulets, provisioning on-eman an application instance to the first cloulet with a price lower than their QoS gain. Our mechanism aress explicitly the challenges of heavily stateful applications whose instances cannot be suspene an/or migrate to other cloulet locations, without impairing irreversibly the perceive QoS. The avantages of such an approach were emonstrate against a variety of proactive an reactive provisioning techniques in realistic topologies. ACKNOWLEDGMENT This work has been supporte by the EC H22 ICN22 project (GA no ), the EPSRC INSP fellowship (EP/M3787/1), an UK EPSRC KCN project (EP/L2612/1). REFERENCES [1] O. Ascigil, T. K. Phan, A. G. Tasiopoulos, V. Sourlas, I. Psaras, an G. Pavlou. On uncoorinate service placement in Ege-Clous. In Clou Computing Technology an Science (ClouCom), 217 IEEE International Conference on, pages IEEE, 217. [2] S. Bharwaj, L. Jain, an S. Jain. Clou computing: A stuy of Infrastructure as a Service (IaaS). International Journal of engineering an information Technology, 2(1):6 63, 21. [3] F. Bonomi, R. Milito, J. Zhu, an S. Aepalli. Fog computing an its role in the Internet of Things. In Proceeings of the first eition of the MCC workshop on Mobile clou computing, pages ACM, 212. [4] C. Clark, K. Fraser, S. Han, J. G. Hansen, E. Jul, C. Limpach, I. Pratt, an A. Warfiel. Live migration of virtual machines. In Proceeings of the 2n Conference on Symposium on Networke Systems Design & Implementation-Volume 2, pages USENIX Association, 25. [5] M. Claypool an K. Claypool. Latency an player actions in online games. Communications of the ACM, 49(11):4 45, 26. [6] G. P. Fettweis. The tactile internet: Applications an challenges. IEEE Vehicular Technology Magazine, 9(1):64 7, 214. [7] A. Greenberg, J. Hamilton, D. A. Maltz, an P. Patel. The cost of a clou: research problems in ata center networks. ACM SIGCOMM computer communication review, 39(1):68 73, 28. [8] H. Izakian, A. Abraham, an B. T. Laani. An auction metho for resource allocation in computational gris. Future Generation Computer Systems, 26(2): , 21. [9] R. Lana, M. Charalambies, R. G. Clegg, D. Griffin, an M. Rio. Self-tuning service provisioning for ecentralize Clou applications. IEEE Transactions on Network an Service Management, 13(2): , 216. [1] I. Lee an K. Lee. The internet of things (iot): Applications, investments, an challenges for enterprises. Business Horizons, 58(4):431 44, 215. [11] L. Saino, I. Psaras, an G. Pavlou. Icarus: a caching simulator for information centric networking (icn). In Proceeings of the 7th International ICST conference on Simulation Tools an Techniques, pages ICST (Institute for Computer Sciences, Social-Informatics an Telecommunications Engineering), 214. [12] M. Satyanarayanan, P. Bahl, R. Caceres, an N. Davies. The case for VM-base Cloulets in mobile computing. IEEE pervasive Computing, 8(4), 29. [13] N. Spring, R. Mahajan, an D. Wetherall. Measuring ISP topologies with Rocketfuel. ACM SIGCOMM Computer Communication Review, 32(4): , 22. [14] A. G. Tasiopoulos, O. Ascigil, I. Psaras, an G. Pavlou. Ege-MAP: Auction markets for ege resource provisioning. In IEEE 19th Worl of Wireless Mobile an Multimeia Networks (WoWMoM), 218. [15] B. Zhang, N. Mor, J. Kolb, D. S. Chan, K. Lutz, E. Allman, J. Wawrzynek, E. A. Lee, an J. Kubiatowicz. The Clou is not enough: Saving IoT from the clou. In HotStorage, 215. [16] L. Zheng, C. Joe-Wong, C. W. Tan, M. Chiang, an X. Wang. How to bi the clou. In ACM SIGCOMM Computer Communication Review, volume 45, pages ACM, 215.