Modelng and Analyss of D Servce Dfferentaton on e-commerce Servers Xaobo Zhou, Unversty of Colorado, Colorado Sprng, CO zbo@cs.uccs.edu Janbn We and Cheng-Zhong Xu Wayne State Unversty, Detrot, Mchgan {bwe,czxu}@wayne.edu More nfo s avalable at http://cc.eng.wayne.edu
Outlne Servce Dfferentaton Slowdown Metrc n e-transactons -D Servce Dff Modelng Alg : Proportonal Slowdown Dff. Alg : Optmal Allocaton Algorthm Performance Evaluaton Conclusons C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers
Servce Dfferentaton Provde dfferent QoS levels to dff. requests SD vs Best-effort, same servce to all Support clent-aware QoS adaptve applcatons Create ncentves for dfferentated prcng Protect servers from DDoS attack!! C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 3
Servce Dff. n Network QoS control n core swtches IntServ (995): tght control for ndvdual flows DffServ(998): loose control for classes Proportonal DffServ: relatve per-class QoS guarantee (Dovrols, et al 999) Network alone s not suffcent to support end-toend servce dff. C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 4
Server Perspectves for SD On a Web content hostng ste Treat clents of dfferent content provders dfferently (Almeda, et al 998) On a streamng or meda-rch ste Adapt to varous devces & access patterns (Chandra, et al 000, Zhou, et al 003) On an ndscrmnate Web ste Control behavors of aggressve clents for farness (Zhu et al 00) On an E-commerce ste Gve hgher prorty to sessons of buyers than vstors n admsson control (Cherksova, et al 00, Chen & Mohapatra 00) Accept/Reect Gradual perf. degradaton C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 5
D Sesson-based Servce Dff. Inter-sesson Dfferent customers exhbt varous navgaton patterns. Accordng to Menasce 00 Heavy buyer class: 5% and 56% of them completed Occasonal buyer & vstor class: 95% Intra-sesson A sesson has a number of states: browsng, search, select, check-out, etc States have dfferent probabltes to transt to the fnal buy state, and cost dfferent resource usage on the average C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 6
Customer Behavors (Menasce, et al 000) Heavy Buyer Occasonal Buyer C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 7
A -D Servce Dff. Model Inter-sesson vs Intra-sesson Each dmenson meets the basc PCF propertes: Predctablty: schedules must be consstent, ndependent of varatons of the class workloads Controllablty: controllable parameters to adust qualty factors between classes Farness: lower classes not be overcompromsed, especally when workload s hgh C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 8
Performance Metrc Slowdown vs Response Tme Slowdown = Queung tme/servce tme Requests have dfferent servce tme; users tend to tolerate long delays for large requests Slowdown translates more drectly to userperceved system load (Bender 88, Harchol-Balter 0, Rska 0, Zhu et al 0) C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 9
Proportonal Slowdown Qualty spacng between states be proportonal to ther pre-defned dfferental weghts α and β. S S S S,,,, = = α α β β α β Sesson β States n the same sesson α α3 α β C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 0
Assumptons m n M/M/ queues wth GPS schedulng Posson sesson arrval process of each customer class and arrval rates of dfferent classes are ndependent Posson arrval rate of requests n each state λ Requests from a customer are dependent, but Requests from dfferent customers are ndependent. Number of re-vsts are small C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers
Processng Rate Allocaton Let d be the demanded resource by requests n state of sesson ; Denote c the allocated resource to the requests Resource constrant Slowdown of a state m n = = c C s = c λ d λ d C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers
Alg : Proportonal-Share Alloc Obectves: S S,, α, = and = S, β α PS allocaton rate: S β ~ λ = αβλd PS allocaton guarantees PCF schedules for requests of dfferent classes. But, t may not be optmal wth respect to an overall resource utlzaton. C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 3
Optmzng -d Servce Dff A D resource allocaton problem Mnmze Subect to m n αβs = = m n = = c C Slowdown of a request and S > 0 s, = λ d, m m n = =, n ( C λ d ) α β λ d = =, α β λ d, C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 4
Propertes of Allocaton Dff. weght of a class ncreases, ts qualty factor ncreases at the cost of others. Qualty factor of a class drops wth the ncrease of ts arrval rate Protecton of DoS attack Load varaton of a hgher-weghted class causes a bgger qualty change of others. Any load shft from a hgher-weghted class to a lower-weghted class leads to an ncrease of the qualty of each class. C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 5
C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 6 Proportonal-sharng Property The qualty spacng of requests n each dmenson s square-root proportonal to ther pre-defned dff. weghts.,,, v v S S α α λ λ =,,,, d d S S β β =,,,, d d S S β α β α λ λ =
Smulaton Model TPC-W e-commerce workload C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 7
Smulaton Parameters Heavy Buyer vs Occasonal Buyer Ther sesson arrval rates are :9 (Menasce 00) Sesson dff. weghts are : 4, based on ther vst ratos Sx States Pay, Add-to-Cart, Select, Search, Browse, and Entry State dff. weghts are 5:4:3::: C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 8
System Slowdown System Slowdown 0 8 6 4 0 no sevce dff proportonal optmal no sevce dff (expected) proportonal (expected) optmal (expected) 0% 0% 40% 60% 80% 00% Server Load (%) C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 9
D Servce Dff. n Slowdown Slowdown.5.5 0.5 0 Slowdown.5 Occasonal buyer Heavy buyer 0.5 Entry state Search state Browse state Select state Add state Buy state Entry state Search state Browse state Select state Add state Buy state 0 Occasonal buyer Heavy buyer Optmal Allocaton C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 0
Mcroscopc Vew of Servce Dff 8 6 4 0 seldom buy busy buy 8 0% 7 home search 6 browse 5 select cart 4 pay Inter-sesson SD (search state) 3 0 50% 0% 80% -0 30 80 30 80 50% 0 60 0 80 Intra-sesson SD (class B) 80% C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers
Long-term vew of SD Slowdown 50 40 30 0 0 0 Inter-sesson servce dfferentaton seldom buy busy buy seldom buy (calculated) busy buy (calculated) Slowdown 5 0 5 0 5 0 Intra-sesson servce dff home search select cart pay 0% 0% 0% 30% 40% 50% 60% 70% 80% 90% Server Load -0% 0% 30% 50% 70% 90% Server Load C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers
Impact of Sesson Arrval Rate.5 Class A Class B Slowdown.5 0.5 0 :7 :8 :9 :3 :4 :5 :6 9: 8: 7: 6: 5: 4: 3: : : : Rato of sesson arrval rate between class A and class B As the rato ncreases, the qualty spacng narrows When the rato goes up to 7:, the predctablty of nter-sesson SD becomes volated C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 3
Impact of Sesson Weght 3.5 Class A Class B Slowdown.5 0.5 0 4: 3: :3 :4 Rato of sesson weght between class A and class B 0:5 9:6 8:7 7:8 No dfferentaton volaton untl the sesson weght raton goes down to :4. The degree of volaton ncreases as the rato contnues to drop. C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 4
Conclusons D Servce Dfferentaton Model for e-commerce transactons Heavy Buyers vs Vstors Dfferentate Check-out from Browsng Proportonal Slowdown Slowdown versus Queung delay On-gong Work Implementaton of processng rate alloc C. Xu @ Wayne State Unv. Servce Dff on e-commerce Servers 5
Modelng and Analyss of D Servce Dff. on e-commerce Servers Thanks More nfo s avalable at http://cc.eng.wayne.edu