Ressursside optimaalne varustamine töövoogudele pilves
Files
Date
2014
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Pilvearvutuse populaarsus on viimaste aastate jooksul märkmisväärselt kasvanud. Kasutades teenustele orienteeritud arhitektuure ning virtualiseerimist, võimaldab pilv varieeruva koormusega
skaleerumist eelkõige ettevõtetele suunatud programmidele. See on üks suuremaid põhjuseid miks töövoogusid pilve migreeritakse. Kuna iga töövoo osa vajab vastavalt kas rohkem või vähem ressursse, siis pilve poolt pakutud ressurssid peavad skaleeruma nii, et see ühtiks töövoo vajadustega. Resursside skaleerimist saab teha manuaalselt, eeldades et töökoormuse muutusperioodid on deterministlikud, või automaatselt, kui töövoos esineb ettearvamatuid koormuse tõuse ning langusi. Seni on esitatud mitmeid automaatse skaleerumise ideid. Mõned neist meetoditest proovivad ennustada, kui palju koormust võib esineda, samal ajal kui teised meetodid proovivad ressursse pakkuda alles koormuse kohale jõudmise ajal.
Mõlema meetodi puhul leidub aga vajadus strateegia järgi, mis tagaks, et ressursse varustataks optimaalselt ehk tuvastada, kui mitu serverit tuleb lisada või eemaldada süsteemist, et rahuldada koormuse nõudlus ning samal ajal minimiseerida ka kulu. Antud magistritöös esitatakse lineaaprogrammeerimisel põhinev meetod, mis arvestab peamisi tegureid skaleerimises nagu, kulu, konfiguratsiooni hind, masinate jõudlus, pilve mahtuvus ning ka iga töömasina kestvus. Antud andmete põhjal tagastatakse optimaalne kombinatsioon võimalikest instantsitüüpidest mis rahuldaks igat töövoo alamosa kõige paremini. Lisaks loodi ka simulatsioon antud mudeli testimiseks ning katsete jooksutamiseks. Tulemuste kohaselt on näha, et pakutud meetod vähendab töövoogude jooksutamise hinda pilves.
Cloud computing has gained significant popularity over past few years. Employing service-oriented architecture and resource virtualization technology, cloud provides the highest level of scalability for enterprise applications with variant load. This feature of cloud is the main attraction for migration of workflows to the cloud. Since each task of a workflow requires different processing power to perform its operation, at time of load variation it must scale in a manner fulfilling its specific requirements the most. Scaling can be done manually, provided that the load change periods are deterministic, or automatically, when there are unpredicted load spikes and slopes in the workload. A number of auto-scaling policies have been proposed so far. Some of these methods try to predict next incoming loads, while others tend to react to the incoming load at its arrival time and change the resource setup based on the real load rate rather than predicted one. However, in both methods there is need for an optimal resource provisioning policy that determines how many servers must be added to or removed from the system in order to fulfill the load while minimizing the cost. Current methods in this field take into account several of related parameters such as incoming workload, CPU usage of servers, network bandwidth, response time, processing power and cost of the servers. Nevertheless, none of them incorporates the life duration of a running server, the metric that can contribute to finding the most optimal policy. This parameter finds importance when the scaling algorithm tries to optimize the cost with employing a spectrum of various instance types featuring different processing powers and costs. In this paper, we will propose a generic LP(linear programming) model that takes into account all major factors involved in scaling including periodic cost, configuration cost and processing power of each instance type, instance count limit of clouds, and also life duration of each instance with customizable level of precision, and outputs an optimal combination of possible instance types suiting each task of a workflow the most. We created a simulation tool based on the proposed model and used 24-hour workload of ClarkNet ISP to conduct performance experiments. The results of experiments suggest that our optimal policy can minimize the cost of running a workflow in the cloud.
Cloud computing has gained significant popularity over past few years. Employing service-oriented architecture and resource virtualization technology, cloud provides the highest level of scalability for enterprise applications with variant load. This feature of cloud is the main attraction for migration of workflows to the cloud. Since each task of a workflow requires different processing power to perform its operation, at time of load variation it must scale in a manner fulfilling its specific requirements the most. Scaling can be done manually, provided that the load change periods are deterministic, or automatically, when there are unpredicted load spikes and slopes in the workload. A number of auto-scaling policies have been proposed so far. Some of these methods try to predict next incoming loads, while others tend to react to the incoming load at its arrival time and change the resource setup based on the real load rate rather than predicted one. However, in both methods there is need for an optimal resource provisioning policy that determines how many servers must be added to or removed from the system in order to fulfill the load while minimizing the cost. Current methods in this field take into account several of related parameters such as incoming workload, CPU usage of servers, network bandwidth, response time, processing power and cost of the servers. Nevertheless, none of them incorporates the life duration of a running server, the metric that can contribute to finding the most optimal policy. This parameter finds importance when the scaling algorithm tries to optimize the cost with employing a spectrum of various instance types featuring different processing powers and costs. In this paper, we will propose a generic LP(linear programming) model that takes into account all major factors involved in scaling including periodic cost, configuration cost and processing power of each instance type, instance count limit of clouds, and also life duration of each instance with customizable level of precision, and outputs an optimal combination of possible instance types suiting each task of a workflow the most. We created a simulation tool based on the proposed model and used 24-hour workload of ClarkNet ISP to conduct performance experiments. The results of experiments suggest that our optimal policy can minimize the cost of running a workflow in the cloud.