A Computation Migration Approach to Elasticity of Cloud Computing

Introduction

Cloud computing is emerging as an important paradigm shift in how computing demands are being met in future. It is at an all-time high and actively transforming the role of IT in businesses in recent years. Different people may have different understanding and definitions on cloud computing. But in essence, it can be seen as a model for Internet-wide access to a shared pool of hardware (processors, storage, etc) and software that are configurable on demand and presented as pay-per-use utility services, primarily Software-as-a-Service (SaaS), Platform-as-a-Service (PaaS) and Infrastructure-as-a-Service (IaaS). Cloud computing poses both opportunities and challenges. Opportunities center on whatever savings (in upfront expenditure, running costs, energy, etc) coming from improved resource utilization and deployment ease.

Enterprises face the problem of high management cost when running their own data centers. It is difficult to estimate the adequate scale of servers—underestimation leads to overloaded conditions while over-provisioning just means wasted investment. On-demand scalability or elasticity is an essential attribute of cloud computing that eases this trouble and reduces the total cost of ownership. With cloud architecture, computing resources can be rapidly provisioned (scale-out) and released (scale-in) with minimal management effort or service provider interaction (Mell & Grance, 2009). Such an elastic infrastructure marks the beauty of cloud computing for enhancing IT delivery’s efficiency and cost-effectiveness. Challenges however include issues of data privacy, security, performance, availability and interoperability. Despite the compelling benefits of cost and flexibility, these areas need significant improvement for cloud computing to go beyond being a fledgling technology.

While enterprises, especially SMBs, are getting on the cloud bandwagon, many may still dread to become a cloud convert. Performance is increasingly becoming their paramount concern ahead of deploying their applications on clouds because it may have direct impact on quality of service and revenue. When applications move from an on-premises platform to a cloud platform, the possible increased latency could overshadow the great promise of clouds. The effect of latency on revenue has been evidenced by various key websites’ analyses: Amazon found every 100 ms delay cost them 1% drop in sales; Google and Yahoo! got 20% and 5-9% fewer traffic from an extra load time of 500 ms and 400 ms respectively (Stefanov, 2009). The impact on latency-sensitive applications can be even more dramatic. TABB Group estimates that if a broker’s e-trading platform is 5 ms behind the competition, it could lose at least 1% of its flow, analogous to a revenue loss of $4 million per millisecond (Willy, 2008). Todd Hoff listed nine sources of latency in his article (Hoff, 2009); and among them, service dependency and geographical distribution tend to be the aggravated factors adding to extra latency when applications go hosted on clouds.

This is an expected aftereffect of adopting the cloud model which abstracts away network infrastructure details. While the abstraction can be a useful concept, it tends to weaken relationships between software components and their associated data. For example, when scaling out, or if compute cycles and storage are to be bought at their preferential rates from different cloud providers, an application may go distributed and involve more hops over the Internet for interaction, resulting in higher user-perceived latency. Therefore, any techniques that can effectively trim latency down are an important contribution to cloud computing. Collocation of computation and data access is one approach, and can be dynamically achieved by computation migration (CM)—an umbrella term of all the concepts behind moving an active computational component from one node to another with its execution state preserved. Besides improving data access locality, CM can also facilitate dynamic load distribution, fault resilience (migrating processes out of nodes with partial failures), and improved system administration (e.g. for server consolidation, graceful shutdown of nodes). It is a useful approach to on-demand cloud provisioning. With virtual machine (VM) migration technology, for instance, one can pool or shrink computing resources as desired by moving VM instances in or out over a cluster of nodes or even a wide area network.