Automatic Scaling in Cloud Computing

Abstract

This dissertation thesis deals with automatic scaling in cloud computing, mainly focusing on the performance of interactive workloads, that is web servers and services, running in an elastic cloud environment. In the rst part of the thesis, the possibility of forecasting the daily curve of workload is evaluated using long-range seasonal techniques of statistical time series analysis. The accuracy is high enough to enable either green computing or lling the unused capacity with batch jobs, hence the need for long-range forecasts. The second part focuses on simulations of automatic scaling, which is necessary for the interactive workload to actually free up space when it is not being utilized at peak capacity. Cloud users are mostly scared of letting a machine control their servers, which is why realistic simulations are needed. We have explored two methods, event-driven simulation and queuetheoretic models. During work on the rst, we have extended the widely-used CloudSim simulation package to be able to dynamically scale the simulation setup at run time and have corrected its engine using knowledge from queueing theory. Our own simulator then relies solely on theoretical models, making it much more precise and much faster than the more general CloudSim. The tools from the two parts together constitute the theoretical foundation which, once implemented in practice, can help leverage cloud technology to actually increase the e ciency of data center hardware. In particular, the main contributions of the dissertation thesis are as follows: 1. New methodology for forecasting time series of web server load and its validation 2. Extension of the often-used simulator CloudSim for interactive load and increasing the accuracy of its output 3. Design and implementation of a fast and accurate simulator of automatic scaling using queueing theory

Tato dizerta cn pr ace se zab yv a cloud computingem, konkr etn e se zam e ruje na v ykon interaktivn z at e ze, nap r klad webov ych server u a slu zeb, kter e b e z v elastick em cloudov em prost red . V prvn c asti pr ace je zhodnocena mo znost p redpov d an denn k rivky z at e ze pomoc metod statistick e anal yzy casov ych rad se sez onn m prvkem a dlouh ym dosahem. P resnost je dostate cn e vysok a, aby umo znila bu d set ren energi nebo vypl nov an nevyu zit e kapacity d avkov ymi ulohami, jejich z doba b ehu je hlavn m d uvodem pro pot rebu dlouhodob e p redpov edi. Druh a c ast se zam e ruje na simulace automatick eho sk alov an , kter e je nutn e, aby interaktivn z at e z skute cn e uvolnila prostor, pokud nen vyt e zov ana na plnou kapacitu. U zivatel e cloud u se p rev a zn e boj nechat stroj, aby ovl adal jejich servery, a pr av e proto jsou pot reba realistick e simulace. Prozkoumali jsme dv e metody, konkr etn e simulaci s prom enn ym casov ym krokem r zen ym ud alostmi a modely z teorie hromadn e obsluhy. B ehem pr ace na prvn z t echto metod jsme roz s rili siroce pou z van y simula cn bal k CloudSim o mo znost dynamicky sk alovat simulovan y syst em za b ehu a opravili jsme jeho j adro za pomoci znalost z teorie hromadn e obsluhy. N a s vlastn simul ator se pak spol eh a pouze na teoretick e modely, co z ho cin p resn ej s m a mnohem rychlej s m ne zli obecn ej s CloudSim. N astroje z obou c ast pr ace tvo r dohromady teoretick y z aklad, kter y, pokud bude implementov an v praxi, pom u ze vyu z t technologii cloudu tak, aby se skute cn e zv y sila efektivita vyu zit hardwaru datov ych center. Hlavn p r nosy t eto dizerta cn pr ace jsou n asleduj c : 1. Stanoven metodologie pro p redpov d an casov ych rad z at e ze webov ych server u a jej validace 2. Roz s ren casto citovan eho simul atoru CloudSim o mo znost simulace interaktivn z at e ze a zp resn en jeho v ysledk u 3. N avrh a implementace rychl eho a p resn eho simul atoru automatick eho sk alov an vyu z vaj c ho teorii hromadn e obsluhy