Article Preview
TopIntroduction
The rate at which data is generated by modern day applications is growing exponentially and is unpredictable with respect to time. Amount of information contained in the data are also disruptive for predictive data processing and analytics functionality. This data needs to be analyzed timely for building business intelligence leading to better decisions. Hence, these applications depend heavily on cloud computing paradigm. Using pay-per-use mode, cloud computing provides autoscaling services to manage the sporadic resource requirement of modern day applications. As industries desire reduced time to market for their applications, they adopt to cloud. Cloud operates via DC (datacenters), has huge computing and storage resources, resulting in heavy electric power consumption. This leads to higher operational cost for the CSPs (cloud service provider) and eventually the same will be passed to the customers. Hence, the current focus of researchers is to identify efficient DC management leading to reduced power consumption while delivering high quality service to the customers. This presents several research challenges including resource management (Manvi, & Shyam, 2014). These challenges need to be addressed after considering multiple parameters related to application types, cloud services and networks of DC.
Application types could be multi-tier web applications, big data processing applications, machine learning and multimedia applications, and high-performance applications. Applications such as video surveillance, real time gaming, real time streaming and augmented reality have intercommunication tasks1 which are too sensitive towards latency and also need more computational power. For example, VMs hosting database tasks and hosting application tasks need intercommunication in three-tier web applications. Similarly, VMs that hosts object tracker task in a video surveillance system that uses multiple cameras covering multiple areas, need inter task communication. As these applications are sporadic resource requester by nature, they depend heavily on elastic capability. Typically, these applications having group of jobs (could be long-duration jobs needing autoscaling), execute their intercommunicating or non-intercommunicating tasks using VMs (Virtual Machines instance) in a DC. Autoscaling service operates on an adjustable capacity (minimum and maximum number of VM), (Boucher Jr, et al, 2018; Barclay, 2016) within which the tasks need to be executed. Without violating this capacity range, addition or deletion of VM instances shall be carried out, upon meeting the resource threshold value specified by the user.
Figure 1. Sample VM and its placement using FCFS
Virtual Machine Placement (VMP) (Lopez-Pires & Baran, 2015) is the process of selecting suitable Physical Machine (PM) to place the requested VM. VMP taking into account the auto-scalability of the applications along with intercommunication of theirs tasks is a challenging cloud problem.
Consider an example of three application requests that are currently (at time T0) executing in a DC. Let ncR1vm1 and ncR3vm1 be the VMs of first and third application respectively, which are non-intercommunication type. Let the second application have two inter-communicating VMs (cR2vm1 and cR2vm2) that needs autoscaling. Assume that at time T1, fourth and fifth requests arrive to the DC. While fourth is a fresh non-intercommunicating type request (ncR4vm1), the fifth request (cR2vm3) is an additional VM of the second request that needs inter communication with cR2vm1 and cR2vm2. Further, let these tasks follow the communication pattern represented by arrows, as shown in Figure 1(a). Assuming that we have two PMs to place these tasks, using First Come First Serve (FCFS) VMP algorithm along with load balancing, placements of VMs at T0 resembles to that of Figure 1(b). Placement of tasks at T1 is shown Figure 1(c). If all the tasks are placed onto single PM itself, the other PM could be switched off to reduce the overall power consumption. Else placing all the intercommunicating tasks (cR2vm1, cR2vm2 and cR2vm3) on to the same PM or on to a network neighborhood machines will reduce the power consumption used for communication, as network latency among these tasks is minimal.