Perform

17
CS 640 1 Network Performance Measurement and Analysis Outline Measurement Tools and Techniques Workload generation Analysis Basic statistics Queuing models Simulation

description

Network performance measurements

Transcript of Perform

CS 640 1Network Performance Measurement and AnalysisOutlineMeasurementTools and Techniues!orkload "enerationAnalysis#asic statistics$ueuin" modelsSimulationCS 640 %Measurement and Analysis O&er&iew'Si(e) com*le+ity and di&ersity of the ,nternet makes it &ery difficult to understand cause-effect relationshi*s'Measurement is necessary for understandin" current system .eha&ior and how new systems will .eha&e/0ow) when) where) what do we measure1'Measurement is meanin"less without careful analysis/Analysis of data "athered from networks is uite different from work done in other disci*lines' Measurement2analysis ena.les models to .e .uilt which can .e used to effecti&ely de&elo* and e&aluate new techniues/Statistical models/$ueuin" models/Simulation models CS 640 34eterminin" What to Measure'#efore any measurements can take *lace one must determine what to measure'There are many commonly used network *erformance characteristics/5atency/Throu"h*ut/6es*onse time/Arri&al rate/7tili(ation/#andwidth/5oss/6outin"/6elia.ilityCS 640 4Measurement ,ntroduction',nternet measurement is done to either analy(e2characteri(e network *henomena or to test new tools) *rotocols) systems) etc8'Measurin" ,nternet *erformance is easier said than done/!hat does 9*erformance: mean1/!orkload ;what and where youority of ,nternet traffic'Trace-.ased workloads/Ca*ture traces and re*lay them/#lack-.o+ method'Synthetic workloads/A.straction of actual o*eration/May not ca*ture all as*ects of workload'Analytic workloads/Attem*t to model workload *recisely/Hery difficultCS 640 IS76FJ !e. !orkload Fenerator'Scala.le 76l Fenerator/Analytic workload "enerator/#ased on 1% em*irically deri&ed distri.utions of !e. .rowsin" .eha&iror/J+*licit) *arameteri(ed models/Ca*tures 9hea&y-tailed: ;hi"hly &aria.le= *ro*erties of !e. workloads/!idely used'S76FJ com*onents@/Statistical distri.ution "enerator/0y*er Te+t Transfer Protocol ;0TTP= reuest "eneratorCS 640 K!orkload characteristics ca*tured in S76FJCharacteristic Com*onent Model System ,m*actDile Si(e #ase file - .ody 5o"normal Dile System L#ase file - tail Pareto LJm.edded file 5o"normal LSin"le file1 5o"normal LSin"le file % 5o"normal L6euest Si(e #ody 5o"normal Network LTail Pareto L4ocument Po*ularity Ai*fCaches) .uffersTem*oral 5ocality 5o"normal Caches) .uffersODD Times Pareto LJm.edded 6eferences Pareto ON Times LSession 5en"ths ,n&erse Faussian Connection times#D JD1 JD% Off time SD Off time #D JD1CS 640 10S76FJ ArchitectureSURGE Client SystemSURGE Client SystemSURGE Client SystemLANON/OFF ThreadON/OFF ThreadON/OFF Thread Web Server SystemCS 640 11S76FJ and SPJC!e.K6 e+ercise ser&ers &ery differently Sur"eSPJC!e.K6-505101520253035400 200 400 600Packets per SecondPercent CPU UtilizationSPECWeb96SURGECS 640 1%Analy(in" Measured 4ata'Analy(in" measured data in networks is ty*ically done usin" statistical methods/Selectin" a**ro*riate analysis method;s= is critical'A&era"in"'4is*ersion ;&aria.ility='Correlations'6e"ression analysis'4istri.utional analysis'Dreuency analysis'Princi*al-com*onent analysis'Cluster analysis'Jach form of analysis has stren"ths and weaknessesCS 640 13Self-Similar Nature of Network Traffric'!8 5eland) M8 Tau) !8 !illin"er) 48 !ilson) On the Self-Similar Nature of Ethernet Traffic, IEEE/ACM TON) 1KK48/#aker Award winner'H8 Pa+son) S8 Dloyd) Wie-Area Traffic!The "ailure of #oi$$on Moeling, IEEE/ACM TON) 1KK?8 'M8 Cro&ella) A8 #esta&ros) Self-Similarit% in Worl Wie We& Traffic! Evience an #o$$i&le Cau$e$, IEEE/ACM TON) 1KKG8 CS 640 14$ueuin" Models'One of the key modelin" techniues for com*uter systems in "eneral/Hast literature on ueuin" theory/Nicely suited for network analysis/Prof8 Mary Hernon is our local e+*ert'Fenerally) ueuin" systems deal with a situation where >o.s ;of which there are many= wait in line for a resource ;of which there are few=/$ueuin" theory can ena.le us to determine res*onse time/J+am*les1CS 640 1?$ueuin" Models contd8'J+am*le@*ackets arri&in" at a router / how can we determine how lon" it takes for *ackets to .e forwarded .y the router1'Characteristics necessary to s*ecify a ueuin" system/Arri&al *rocess/Ser&ice time distri.ution/Num.er of ser&ers/System ca*acity ;num.er of .uffers=/Po*ulation si(e/Ser&ice disci*line/Mendal notation@A2S2m2#2M2S4'6es*onse time C waitin" time B ser&ice time'Dor sta.ility) mean arri&al rate must .e less than mean ser&ice rateCS 640 165ittleo.s in system C arri&al rate L mean res*onse time /Treats a system as a .lack .o+/A**lies whene&er num.er of >o.s enterin" the system euals num.er of >o.s lea&in" the system'No >o.s created or lost inside system/Can .e e+tended to include systems with finite .uffers'J+am*le@A&era"e forwardin" time in a router is 100 microseconds) ,2O rate for *ackets is 100k8!hat is the mean num.er of *ackets .uffered in the router1CS 640 1GSimulation Models'Simulation is one of the most common2im*ortant methods of analysis2modelin"/Ty*ically an a.straction of the system under consideration/Can *ro&ide si"nificant insi"ht to system