Performance Guarantee Mechanism for Multi-Tenancy SaaS Service Based on Kalman Filtering

Transcription

1 BULGARIA ACADEMY OF SCIECES CYBERETICS AD IFORMATIO TECHOLOGIES Volume 5, o 3 Sofa 205 Prnt ISS: ; Onlne ISS: DOI: 0.55/cat Performance Guarantee Mechansm for Mult-Tenancy SaaS Servce Based on Kalman Flterng Lu Yanpe, 2, L Chunln,Yang Zhyong, Chen Yuxuan 3, Xu Lun Department of Computer Scence, Wuhan Unversty of Technology, Wuhan, Chna 2 Informaton Engneerng College, Henan Insttute of Scence and Technology, Xnxang, Chna 3 Department of Mcroelectroncs and Sold-State Electroncs, Unversty of Electroncs Scence and Technology of Chna, Chengdu, Chna Emal: lupeo23@gmal.com Abstract: Ths paper proposes a specal System Archtecture for Mult-tenancy SaaS Servce (SAMSS), whch studes the performance securty ssues at the busness logc layer and data processng layer respectvely. The Kalman flterng Admsson Control algorthm (KAC) and the Greedy Copy Management algorthm (GCM) are proposed. At the busness logc layer, Kalman flterng admsson control algorthm s presented. It uses a Kalman flter to conduct the dynamc evaluaton for the CPU resource for mult-tenancy SaaS servce and reduces the unnecessary performance expenses caused by drect measurement of CPU resources. At the data processng layer, the Greedy Copy Management algorthm (GCM) s presented. It changes the copy placement as a K-parttonng set parttonng problem and adopts a greedy strategy to reduce the number of tmes for creatng a data copy. Fnally, the expermental analyss and results prove the feasblty and effcency of the algorthms proposed. Keywords: Kalman flterng, busness logc layer, data processng layer, multtenancy, SaaS.. Introducton Software as a Servce (SaaS) s a way of software deployment. Tenants of the order servces have authorzed ts end users to access the software through the networ on demand, where the end users share the applcatons and data. Mult-tenancy enables the concurrent users from dfferent tenants to share the same nfrastructure 50

2 resources, for the sae of cost reducng and ncome ncreasng, whch s one of the ey features of SaaS servce under the crcumstances of large-scale cloud computng [-3]. Performance guarantee for mult-tenancy SaaS servce has been focused on n recent studes. Flash crowd [4] s a large number of tenants requestng SaaS servce smultaneously, causng the server to overload and the server buffer to be completely consumed. Large numbers of request pacets have been dscarded. Due to the computng resource lmts, the accepted tenant requests must wat long for the allocated computng resources, that causes a great mpact on the delay performance, the tenants tend to gve up the servce because of a pacet loss or too long watng tme, resultng n great loss. In addton, wth the rapd ncrease of SaaS servce tenants and the random appearance of tenants Flash crowd, the data resources are facng great load and stress, whch means that the data resources wll become the bottlenec of the whole SaaS servce. You need to adapt to the load fluctuaton by dynamc changng the number of servers.the strategy of admsson control and copy management s the ey to the system performance guarantee mechansm for mult-tenancy SaaS servce. As for the admsson control strategy, García D F et al. [5] pont out the requrements that should be met when the QoS control mechansm of servce provder server wors n B2B envronment, puttng forward the QoS control mechansm wth a montor and a controller. The servce qualty control algorthm maes schedules accordng to the user s prorty. The QoS control mechansm ensures the system performance under crcumstances of overload. Z h a n g et al. [6] propose a resource consumpton estmate model n a multtenancy applcaton envronment, based on the maxmum sze of tenants wth gven nodes, fgured out by the heurstc algorthm. But the present researches can hardly assess the tme-varyng resource state of SaaS servces effcently and the precson of the method depends on long tme, hgh qualty samples as nputs, whch s easly affected by outlers caused by resource competton, thus generatng errors n the admsson control mechansm [7]. In addton, the current copy management strategy mostly adopts the method of creatng copy numbers and fxed tmng [8], so t s dffcult to adapt to SaaS servce by the dynamc change of the number of tenants, such as creatng too many copes, for there wll be waste of storage space. On the other hand, the system performance cannot be mproved effcently. In addton, a frequent mgraton of the copy data wll consume the networ bandwdth, thus reducng the performance of the system. In the vewpont of the problems above mentoned, ths paper studes ts performance guarantee respectvely at the layer of busness logc and data processng. At the busness logc layer, Kalman flterng [9] was used to reflect the resource usage and surplus stuaton on dfferent servers n tme for CPU resources dynamc assessment for mult-tenancy SaaS servce. It provdes the bass for tenant s admsson control mechansm. KAC, the Kalman flterng admsson control algorthm was proposed. At the data processng layer, the system archtecture s adapted to the dynamc changes of the load by dynamcally adustng the load dstrbuton among every copy and the placement of a copy between the nodes. The optmal weghted rotaton schedulng algorthm was used to realze the 5

3 dstrbuton of the load. Also, the system archtecture classfed the copy placement nto a K-parttonng [0] set partton problem and utlzes the greedy strategy to reduce the creaton tmes. GCM, the algorthm of greedy copy management was proposed. Ths paper s dvded nto fve sectons. Secton outlnes the system archtecture SAMSS for mult-tenancy SaaS servce; Secton 2 dscusses the performance guarantee mechansm at the busness logc layer; Secton 3 descrbes the performance guarantee mechansm of the data management layer; Secton 4 presents an expermental analyss for the performance guarantee technology. Fnally, there s a summary of the paper. 2. System archtecture for mult-tenancy SaaS servce The System Archtecture for Mult-tenancy SaaS Servce (SAMSS) dscussed n ths paper s made up of the busness logc layer and data processng layer. A full servce request process starts from the busness logc layer, and t may need access to the data processng layer many tmes durng the process untl t generates a complete response and returns the response to the tenant. The busness logc layer s manly ntended to provde a certan degree of performance guarantee for tenants. The busness logc layer of SAMSS manly conssts of resource dynamc evaluaton components and an access control unt. The resource dynamc assessment components manly use a Kalman flter to mae resource consumpton calculaton for the mult-tenant SaaS servce, so that the system resource usage and the remanng ones can be obtaned tmely to provde a bass for udgment at the next step of access control. The access control unt manages the access request of the tenants n the whole system to prevent the system from beng overloaded and decdes whether to accept the new tenant requests accordng to the usage of the server resource. The data processng layer s manly ntended to provde hgh avalablty of data protecton to tenants. SAMSS mantans multple copes of data dstrbuted on dfferent nodes for each tenant. In SAMSS the relablty and access effcency of each tenant s data can be mproved through the use of multple nodes, so a request dspatcher used for load balancng between nodes s supposed to be added. 2.. The performance guarantee mechansm of the busness logc layer Ths strategy uses a Kalman flter to do the dynamc assessment of CPU resources for mult-tenant SaaS servces and reduces the drect measurement of CPU resources, when t needs to nect a probe and causes unnecessary performance overhead. Besdes, the mult-tenant strategy access control s desgned by usng the method of resource allocaton and resource reservaton and t avods the system s overload caused by flash crowd tenants The dynamc evaluaton of mult-tenant CPU resources based on a Kalman flter Kalman flter has been wdely used and studed n the felds of automatc control and auxlary navgaton, and ts man characterstc s that t can use a form of 52

4 approxmately optmal estmaton based on an observable value to estmate unobservable value, and can update the former observed value as the new observed value comes out, therefore t s more sutable for onlne assessment of the tmevaryng resource state. Kalman flter provdes a general method to estmate the unobservable value x n dscrete tme ponts. The state x at pont can be defned as a lnear stochastc dfference equaton: () x = + Ax w. The test value z at pont s defned as (2) z = H x + v, where: A s the state transton matrx from the pont to the -th pont; w s the process error; Q s the covarance matrx; H s the transton matrx from x to z ; v s the observaton error; R s the covarance matrx. Mult-tenant SaaS servce request process may nvolve a varety of resources. Those wth hgher resource utlzaton rate are called bottlenec resources. For dfferent types of servces, the bottlenec resources may be dfferent, but the bottlenec resources for mult-tenant SaaS servce are most lely to be the CPU resources. Therefore, n ths paper we manly consder the dynamc assessment of the CPU resources. For mult-tenant SaaS servces, the prerequste condton for dynamc assessment s collectng onlne server log nformaton, ncludng the tenant s throughput amount and CPU utlzaton rate of the server. Durng the servce, runtme nformaton s montored and recorded at a fxed tme nterval, ths nterval s called a montorng wndow. For convenence of dscusson, to a server whch s the host of tenants, we gve the followng symbols and ther meanngs: T ndcates the sze of the montorng wndow; ndcates the number of transactons that the -th( )tenant completes n the montorng wndow; U CPU ndcates the average CPU utlzaton rate of the server n the montorng wndow; S ndcates the average servce tme of all the -th tenant s transactons (namely, average CPU tme occuped by all transactons). Accordng to the Utlzaton Law [], the resource utlzaton rate s equal to the throughput amount multpled by the servce tme, the equaton obtaned s as follows: (3) U T S. CPU = Due to the dffculty to measure the servce tme S accurately, we use C to represent ts approxmaton, so as to obtan the calculaton equaton of the resource utlzaton U rate approxmaton: CPU 53

5 C (4) U CPU =. T A statstcal analyss method can be used to solvec, for such ndrect approxmaton soluton of the problem, the error of U CPU and U CPU s one of the typcal ndcators for measurng accuracy. ext, our goal s to wor out how to reduce the error of the real servce tme S and the approxmate servce tme C. Frstly, we model the unobservable state x nto an -dmensonal vector x = ( C, C2, L, C ) whch contans tenants average servce tme of transactons, and t ndcates each tenant s average servce tme of transactons at pont. Then, accordng to Formula (3), we model the observed total CPU utlzaton rate z and we obtan: C (5) z = + v, T where ndcates the montored throughput rate of tenant ; the transton matrx 2 H from x to z s defned as (,, L, ). T T T The Kalman flter algorthm maes teraton assessment on the servce tme at the end of each montorng wndow, the ntal value ncludng the state ntal value ˆx and the ntal error covarance matrx. The teraton process s as follows: 0 forward proecton of the state of x (6) x ˆ ˆ = Ax ; 2 calculate forwards the covarance matrx of the state pror estmate error P T (7) P AP A + Q ; 54 = 3 calculate the Kalman gan K T T (8) K = P H ( H P H + R ) ; update the state of x accordng to the observed varable z (9) x ˆ = ˆ + ( ˆ x K z H x ) ; calculate the covarance matrx P of the estmate error after the state has been updated (0) P = ( I K H ) P. In the process of teraton, step whch modfes the state of x s the ey to update the estmated value, ths equaton can be smplfed n the form of x new = x old + K e, whch s to say, that K can be regarded as the weght matrx of the

6 modfed x, and the error e and the correspondng weght are used to correct the data x old At the same tme, n the calculaton of step, t s also necessary to consder to set the valuaton range of the average servce tme for each tenant affar, namely non-negatve and less than a certan upper bound: 0 C < u, where u s the upper bound of the estmated value. In ths paper we set the range of state x as: ( 0,( μ u + ( μ) x )), where μ = 0. 9 and u = U CPUT /. Then the calculaton at step s revsed as: ) tral ) ) () x = x + K ( z H x ), ) ) tral (2) x = mn( x, ( μu + ( μ) x )). When there s any tenant requestng to be allowed to enter the system, the system calculates the resource consumpton by usng the mult-tenancy CPU resource dynamc assessment method based on Kalman flterng. The usage and the rest of the system resource can be nown n tme, provdng the bass for the next tenant s admsson control mechansm, so that that the delay, due to mang a decson by montorng the response tme, can be avoded, reducng the unnecessary performance expenses caused by drect measurement of CPU resources whch needs an nected probe Admsson control mechansm In the proposed system framewor for mult-tenancy SaaS servce, there are three dependent servces represented, respectvely standng for tenant s three nds of servces, where there s a competton for servce S3 between (S, S2, S3) and (S4, S5, S3), whch s very common n the mult-tenancy SaaS servce system. To llustrate our control strategy, we deal wth the n requests from tenant t, and only consderng the CPU resource as an example. When a tenant request arrves, the system wll begn to mplement the admsson control algorthm, whch ncludes the allocaton part and the reserve part of resources. In the allocaton part, f the remanng resources n the server where there are servce nstances of S, S2, S3 are greater than the specfed mnmum remanng resources of each server, the algorthm wll frstly wor out the acceptable request number n. Once a new request s accepted, the system wll recalculate the remanng resources of R S, R S2, R S3 by usng a mult-tenancy CPU resource dynamc assessment method based on Kalman flterng, and at the same tme, t wll record ts onlne tenants numbers accordng to the request category. If the remanng resources n the servers of S, S2, S3 are less than the mnmum remanng resources, ths means t has attaned a crtcal load state. If the tenants are stll a lot, we can only accept a lmted number of accessed requests from a part of the senor tenants (tenant classfcaton: senor tenants, ntermedate tenants and prmary tenants). Ths effcently mproves the mpact the system has on the performance of senor tenants wth heavy load. In order to further mprove the system s ablty to handle the request, we set a buffer queue n the controller, to store temporarly n the cache queue a certan number of access requests reected due to overload protecton. Once the system resources have a rest, these requests wll be processed n tme. We call t the Kalman flterng Admsson Control algorthm (KAC). 55

7 2.2. The performance guarantee mechansm of the data processng layer Wth the rapd ncrease of SaaS servce tenants and the random appearance of tenants Flash crowd, the data resources are under great load and pressure, whch means that the data resources wll become the bottlenec of the whole SaaS servce; hence a performance guarantee mechansm s requred for the layer of data processng. Wthn the data processng layer, SAMSS wll allocate each data copy to each node and use the data-copyng algorthm accordng to [2] to ensure the consstency of each copy. The pont of ths secton s to see out how to adust the allocaton of loads among the copes and the placement of copes among nodes to guarantee the performance ndcators at the data processng layer on the bass of ensurng the consstency of each copy. At present, the settng of the load dstrbuton strategy [3, 4] and the copy placement strategy [5-7] are facng the followng problems. Settng the load dstrbuton strategy, under the premse of fxed copy placement, how to allocate the read only load to each node has become a ey subproblem for the performance safeguard mechansm. The dffculty of ths subproblem s how to mprove the overall performance of the system under the restrcton of the state by balancng the load of each node as much as possble. The settng of a copy placement strategy, the dramatc change of the load maes the wor of balancng each node only by load dstrbuton dffcult, hence the system must adust the mappng of copes to nodes and the gross of data nodes to have the load balanced. In ths process, the system must mnmze the usage of resources and the moves of copes, whch ncrease the dffculty of settng of the copy placement strategy. Ths secton focuses manly on studyng the above two problems. In order to facltate the dscusson, we lst the followng symbols and ther meanngs: S the set of data nodes, of whch s denoted by s, and [, M] ; T the tenants collecton, of whch s denoted by t, and [, ] ; L the system s overall load vector, namely L =< L, L, L, L > ; 2 SF, the Stretch Factor ndcator of the request of the tenant t when performed on s ; R T T T T the copy number of tenant t, and R max, R mn, R default, respectvely represent the maxmum, the mnmum and the default value of R T ; K the vector the server uses; when =, t means that the server s n a runnng state, and when = 0 ust the opposte; A the placed matrx of a copy of the data, namely 56

8 A A = M A,, L O L A, M M A, M ; f server s mantans a copy of the data of tenant t then A, =, otherwse A, = 0 ; A 0 the generated placed matrx of copy n the process of the last copy s adustment; λ the load dstrbuton matrx, namely λ, L λ, M λ = M O M, λ, L λ, M where λ, represents the request rate when the load of tenant t s allocated to the R server node s, furthermore, λ, represents the request rate when the read load of W tenant t s allocated to the server node s, and λ, represents the request rate when the load of tenant t s allocated to the server node s. SF max the elongaton factor threshold of the data query request; U max the bggest resource utlzaton threshold of each node; U target the overall goal resource utlzaton of the system; U upperthreshold the floatng range of the upper lmt of the overall resource utlzaton of the system; U lowerthreshold the floatng range of the lower lmt of the overall resource utlzaton of the system Load dstrbuton In order to accurately analyze the load dstrbuton problems, t s necessary to defne the metrcs of the load balancng degree, and establsh the performance model of data nodes, whch s used to estmate the effect of adustment before settng the real resource management strateges Tang lessons from prevous wors, ths paper uses a queue havng multple types of M/G// PS requests as the performance model of each data node. 57

9 Accordng to the queung theory results, the requestng elongaton factor performed on node s s (3) SF, = =, W W R R U ( λ D + λ D ),, = W R where node U represents the resource utlzaton rate, and 0 U < ; D and D respectvely represent the wrte data rate and read data rate of tenant t. Ths paper taes the load nclnaton rate as the cluster node load mbalance degree of metrcs, the formal defnton s as follows: M (4) Slope =. = U Thus the features of the load nclnaton rate could be concluded:. When a node s resource utlzaton rate s hgh, ts contrbuton to the load nclnaton rate wll greatly mprove. When a node s resource utlzaton rate approaches 00%, the system load nclnaton rate approaches nfnty. 2. When each data node s at the same resource utlzaton rate, the load nclnaton wll be mnmum. Based on the above performance model and the metrcs, we could further form the load balancng problem nto the followng formula: M (5) mnslope( λ ) = = W W R R λ D + λ D 58 ( ),,, where the reasonable value range of read requests dstrbuton s [, ], [, ], M R R R R R M λ = L,, f A = 0, then λ = 0,, 0 λ L, ; the = W, W L A =, reasonable value range of the wrte requests dstrbuton s λ =,, whch 0, A = 0, also requres that the load summaton of each node dstrbuted must be less than ts W W R R processng power, namely [ λ, D + λ, D ] <. = The above ssues belong to a specal nd of the load dstrbuton problem; the deal state s the system to be able to get the load nformaton of each node n a data copy n the dstrbuton of each tenant request. For each tenant t request, we dstrbute t to the node whch has the smallest queue length n the node collecton of the data copy where t s stored. But t s the proper state to be able to get the data and the load nformaton of each node n the copy n tme; we call ths algorthm the Shortest Queue Frst (SQF). In an actual stuaton, when each tenant request s dstrbuted n the system, the load nformaton of each node can only be obtaned perodcally. Then the Weghted Round-Robn schedulng algorthm s adopted. We use the loadng ntensty that each tenant t s assgned to the node s =

10 as the weght λ, of the Weghted Round-Robn schedulng algorthm. Ths algorthm s called the Optmal Weghted Round-Robn (OWRR). The load dstrbuton matrx obtaned from ths algorthm s called the Optmal Balanced Load (OBL) The copy placement When the load fluctuaton s relatvely severe, smply changng the dstrbuton of load wll mae t dffcult to accomplsh the whole goal of ths secton that s mnmzng the consumpton of resources on condton that each performance ndex of the tenant s assured. To further mplement the target, the dynamc adustment to the copy placement s needed. Specfcally, two nds of performance management methods of the load dstrbuton and the copy placement need to be consdered smultaneously. The performance management goal on the data processng layer can be formally descrbed as follows: M (6) mn f( λ, A) = mn g( λ, A) = A A A {0,}, 0, = where the number of data copy of each tenant has the followng range [, ], M = T A, [ Rmn, R T max ] ; and t s requred that only nodes n the movng state can f A, > 0 = place the data copy, namely [, M ], = ; the reasonable 0 f = A, 0 = W, W L A =, value range of read requests s λ =, ; and the performance ndex for 0, A = 0, each tenant s bounded, namely SF W W R R max. ( λ D + λ D ) =,, Through classfyng the problem above mentoned nto a K-parttonng set partton problem, we can draw the concluson that t s a P-hard problem, so t s dffcult to get the optmal soluton. Ths paper uses a greedy strategy to try to fnd an optmal soluton. We call t the GCM algorthm. GCM s composed of the node provdng algorthm and the copy placement algorthm. The node provdng algorthm adapts to changes n the overall ntensty of load by dynamcally adustng the number of nodes n the system. Specfcally, when shrnng the number of nodes, the algorthm adopts the strategy of deletng the node wth the lowest load frst under the Optmal Balanced Load (OBL) scheme. Snce deletng any nodes may mae the amount of certan tenant data copy less than T the node R mn, t s reasonable to move t to some reserved nodes rather than smply throw t away. Also, t s necessary to move the copy to the node wth the lowest load frst under the OBL scheme when a copy movement s requred. 59

11 In essence, the copy placement algorthm s a nd of a greedy algorthm. Every tme when the copy that needs adustment s selected, the copy placement algorthm frst chooses to completely reduce the extent of the copy that has an unbalanced load. The tme complexty of ths algorthm s O(n), where n represents the number of nodes. The copy placement algorthm s used for the adustment of the copy placement to adapt to the change of the proporton of each tenant n the system when the load vares. Frst of all, the copy placement algorthm calls the node provdng algorthm to adust the node volume. Subsequently, t eeps an eye on the load balance between nodes. 3. Expermental analyss Ths secton dscussed the effect of the mult-tenant SaaS servce performance guarantee mechansm separately on the busness logc layer and data processng layer. 3.. The busness logc layer The expermental envronment s a typcal three-ter archtecture system, ncludng the Tomcat server whch deploys the servces S, S2, S3, S4 and a number of database servers. S, S2, S3, S4 are deployed on the same server, the Apache JMeter 2.4 s run to smulate a tenant s load, and the average response tme, throughput, and the other data s obtaned through Graph Result lsteners. All systems are run on Wndows Server In order to verfy the effect of the performance guarantee mechansm at the busness logc layer, we perform multtenant CPU resources dynamc evaluaton strategy based on a Kalman flter, and the admsson control mechansm s used on the bass of the evaluaton. The CPU overload threshold s set to 90% accordng to the experence. Fgs and 2 show the CPU utlzaton and the total CPU utlzaton of the servces S, S2, S3, S4, separately wthout an admsson control mechansm and wth the use of an access control mechansm when the tenants number rapdly fluctuate. CPU utlzaton/% 00% 90% 80% 70% 60% 50% 40% 30% 20% 0% 0% tme/ m n ser v ce S ser v ce S2 ser v ce S3 ser v ce S4 total CPU utlzaton Fg.. Wthout an admsson control mechansm 60

12 CPU utlzaton/% 00% 90% 80% 70% 60% 50% 40% 30% 20% 0% 0% tme/ mn ser v ce S ser v ce S2 ser v ce S3 ser v ce S4 total CPU utlzaton Fg. 2. Wth an admsson control mechansm As shown n Fg., wthout the use of an admsson control mechansm, wth the rapd ncrease of the number of tenants and the tenant request servces S, S2, S3, S4 at the same tme, leadng to the server s overload, the server s CPU total utlzaton rate s close to 00%, whch heavly affects the response tme. Therefore, the tenants need to wat for a long tme. Even some of the tenants request pacet s drectly dscarded, whch serously affects the user experence. Fg. 2 shows that due to the use of an admsson control mechansm, wth the same hgh load, the server s CPU total utlzaton rate remans below 90%. Although t s n the case of a Flash crowd, t stll affects the response tme of the request, but the total CPU utlzaton remans n a reasonable scope, promotng the user experence The data processng layer In order to evaluate the effect of the performance guarantee mechansm on the data processng layer, we adopt the method of smulaton analyss to verfy whether GCM can adust to the dynamc changes of the load through the load dstrbuton and the copy adustment. We use the wdely used and authortatve smulaton tool CloudSm. The parameters of the smulator confguraton are as follows: the number of tenants s 500; the number of copes of each tenant s 4; the number of nodes s 40. In the smulaton process, we gradually adust the overall load ntensty to reflect the change of node average resource utlzaton. We compare the algorthm of OWRR that we proposed, the classcal algorthm of Shaped Round-Robn (SRR), Defct Round Robn (DRR) wth the deal algorthm of the Shortest Queue Frst (SQF). The comparson results are shown n Fg. 3. r esponse t me/ ms resource utlzaton SQF OWRR DRR SRR Fg. 3. The performance comparson of four nds of load dstrbuton algorthms 6

13 We can see from Fg. 3 that SQF s an deal state. Wth the ncrease of the resource utlzaton, there s almost no change n the response tme. The performance of OWRR s far better than SRR and DRR, the reason for whch s that there s no overall coordnaton of each tenant load n SRR and DRR, thus leadng to a stuaton where some nodes are under too heavy loads whereas some nodes loads are too lght. The performance of processng the tenant requests wll be greatly reduced when the load of the node s too heavy. The results show that the OWRR algorthm has a better performance. It s dffcult to deal wth the wde margn of fluctuatons of the overall load ntensty by usng the load dstrbuton alone. Then we tae the dynamc adustment of the copy placement nto consderaton to verfy whether GCM can guarantee the performance ndex of each tenant database on condton that the utlzaton rato of system resources s qute hgh. Tang nto account the fact that the system performance wll be nfluenced durng the process of copy adustment, we are loong forward to the mnmum tmes of copy adustment. Copy adustment can be refned nto creatng a copy and deletng a copy. The cost of creatng a copy s much hgher than deletng one, therefore, n the subsequent experments, we use the tmes of creatng a copy as one of the man evaluaton ndces. Based on the approxmate lnear relatonshp between the system load and the response tme, we concluded from the experment that when the tenant number was 000, the system load reached ts lmt. Then we tested the stuaton where the tenant number ranged from 800 to 500 to determne the effcency of the safeguard mechansm n the system data processng layer wth lght load and heavy load. Fg. 4 shows the performance of the algorthm of GCM under the load drver. t he number of t enant s tme/ m n (a) resource utlzaton 00% 90% 80% 70% 60% 50% 40% 30% 20% 0% 0% tme/ m n (b) the tmes of t he copy cr eat on tme/ m n t he aver age r esponse tme/ ms tme/ m n (c) (d) Fg. 4. The performance of the algorthm GCM under the load drver 62

14 Fg. 4 (a) shows the curve chart wth an ndependent varable of tme when the number of tenants ranges from 800 up to 500. Fg. 4(b) shows the change of the average resource utlzaton of nodes as tme vares. Fg. 4(c) shows the change of the tmes of the copy creaton. Fg. 4(d) shows the change of the average response tme of all tenants as tme vares. We can see from the dagrams for the load above gven, that GCM can well guarantee the resource utlzaton on the overall platform and assure that the tmes of creatng a copy are wthn a reasonable range. A concluson can be drawn that our safeguard mechansm at the data processng layer has played a good role. 4. Concluson The present paper proposed the system archtecture SAMSS for mult-tenancy SaaS servce, studed SaaS servce performance guarantee mechansm respectvely at the busness logc layer and the data processng layer. For a mult-tenant request processng process at the busness logc layer ths paper, based on the two aspects of dynamc evaluaton of CPU resources and admsson control mechansm, puts forward KAC the Kalman flterng admsson control algorthm that has used a Kalman flter to mae a dynamc assessment for CPU resources of mult-tenancy SaaS servce and response on dfferent server resource usage n tme, whch would avod unnecessary performance overhead resultng from an nected probe by drect measurement of the CPU resources. For mult-tenant data processng ths paper pertnently studes the two aspects of load allocaton and a copy, the optmal weghted rotaton schedulng algorthm s used to get the dstrbuton of the load between each copy. Ths paper brngs up the busness management approach, the GCM, whch loads the balancng ssues under the premse of a fxed copy placement formalzed as an optmzaton problem, and puts forward to get the optmal load dstrbuton of the load balancng algorthm. The expermental results show that the GCM can adapt to the dynamc changes of the load through fewer copes of adustment and guarantee the performance of the tenant database whle mantanng hgh node resource utlzaton. Acnowledgements: Ths wor was supported by the atonal atural Scence Foundaton under Grants (o , o 67075), the Program for the Hgh-end Talents of Hube Provnce, Key atural Scence Foundaton of Hube Provnce (o. 204CFA050), and the Open Fund of the State Key Laboratory of Software Development Envronment (SKLSDE). Any opnons, fndngs, and conclusons belong to the authors and do not necessarly reflect the vews of the above agences. References. M e t z n e r, R., A. M e t z g e r, F. L e y m a n n et al. Varablty Modelng to Support Customzaton and Deployment of Mult-Tenant-Aware Software as a Servce Applcatons. In: Proc. of 2009 ICSE Worshop on Prncples of Engneerng Servce Orented Systems. IEEE Computer Socety, 2009, pp Z h a n g, F., J. C h e n, H. C h e n et al. Cloudvsor: Retrofttng Protecton of Vrtual Machnes n Mult-Tenant Cloud wth ested Vrtualzaton. In: Proc. of 23rd ACM Symposum on Operatng Systems Prncples. ACM, 20, pp

15 3. Z h a n g, Q., L. C h e n g, R. B o u t a b a. Cloud Computng: State-of-the-Art and Research Challenges. Journal of Internet Servces and Applcatons, 200, pp Lu, F., B. L, L. Zhong et al. Flash Crowd n P2P Lve Streamng Systems: Fundamental Characterstcs and Desgn Implcatons. Parallel and Dstrbuted Systems, IEEE Transactons on, Vol. 23, 202, o 7, pp G a r c í a, D. F., J. G a r c í a, J. E n t r a l g o et al. A qos Control Mechansm to Provde Servce Dfferentaton and Overload Protecton to Internet Scalable Servers. Servces Computng, IEEE Transactons on, Vol. 2, 2009, o, pp Z h a n g, Y, Z h h u W a n g, B o G a o et al. An Effectve Heurstc for On-Lne Tenant Placement Problem n SaaS. 200 eee Internatonal Conference on. IEEE, 200, pp Cherasova, L., K. Ozonat,. M et al. Automated Anomaly Detecton and Performance Modelng of Enterprse Applcatons. ACM Transactons on Computer Systems, Vol. 27, 2009, o 3, pp Shvacho, K., H. Kuang, S. Rada et al. The Hadoop Dstrbuted Fle System. Mass Storage Systems and Technologes. In: Proc. of 200 IEEE 26th Symposum on IEEE, 200, pp K a l m a n, R. E. A ew Approach to Lnear Flterng and Predcton Problems. Journal of Basc Engneerng, Vol. 82, 960, o, pp B o v e r, A., I. K u r o v a. Posson Convergence n the Restrcted Parttonng Problem. Random Structures & Algorthms, Vol. 30, 2007, o 4, pp L a z o w s a, E. D., J. Z a h o r a n, G. S. G r a h a m et al. Quanttatve System Performance: Computer System Analyss Usng Queueng etwor Models. Prentce-Hall, Inc., B a n s a l, S., S. S h a r m a, I. T r v e d. A Dynamc Replca Placement Algorthm to Enhance Multple Falures Capablty n Dstrbuted System. Internatonal Journal of Dstrbuted and Parallel Systems (IJDPS), Vol. 27, 20, o Y o o n, J. H., J. H. C h o, K. J. L e e et al. A Fully Dstrbuted Replca Allocaton Scheme for an Opportunstc etwor. Wreless etwors, 203, pp L, B., S. Song, I. Bezaova et al. Energy-Aware Replca Selecton for Data-Intensve Servces n Cloud. Modelng, Analyss & Smulaton of Computer and Telecommuncaton Systems. In: Proc. of 202 IEEE 20th Internatonal Symposum on. IEEE, 202, pp H u a n g, X., Y. P e n g. A ovel Replca Placement Strategy for Data Center etwor. In: Proc. of 202 Internatonal Conference on Informaton Technology and Management Scence Proceedngs. Berln Hedelberg, Sprnger, 203, pp T u, M., I. L. Y e n. Dstrbuted Replca Placement Algorthms for Correlated Data. The Journal of Supercomputng, 203, pp Méndez Muñoz, V., G. Amorós Vcente, F. García Carballera et al. Emergent Algorthms for Replca Locaton and Selecton n Data Grd. Future Generaton Computer Systems, Vol. 26,200, o 7, pp