Learning Clouds

点击翻译按钮获取中文摘要

Cloud computing is witnessing an unprecedented popularity with many organizations turning to cloud-hosted solutions for their applications and even for operating entire computing infrastructures. Those organizations who shy away from this transition are starting to deploy private clouds to offer similar capabilities in-house, presupposing they would retain better control. The benefits of cloud computing are multifaceted: Cost reduction, avoiding redundancy, and global data sharing, to just name a few. Clouds are widely touted for elasticity, endless scalability, security, and cost reductions by cutting IT personnel and conserving resources. However, many of these promises are fallacies, as the proper resource management algorithms and control mechanisms are not well understood. Existing cloud management frameworks and therefore the clouds they enable require the cloud user and developer to specify resource usage needs, an art that is often more based on black magic than on sound principles. Worse yet, users fear the worst case scenario and do not want their services and applications to under-perform, so massive resource over-provisioning is the answer and the norm rather than the exception. It remains poorly understood how much compute, memory, disk, and I/O capabilities an application system really requires to function. Decisions are ad hoc, more driven by budget than by true needs. Also, application systems do not always experience the same level of utilization, so at times expensive resources remain idle This is where Learning Clouds sets in and what we are trying to remedy in our research. Learning Clouds treat the cloud with its hosted services and applications as an autonomous system that self-manages, optimally allocating its resources in an online fashion guided via machine learning, rebalancing scarce resources on demand. The net outcome will be ease of application deployment as users are freed from making hard resource allocation decisions, better application performance as applications are assigned resources dynamically on an as-needed basis, and cost reductions as physical resources are allocated optimally and remain better utilized. Learning Clouds are based on novel reactive and proactive control techniques that strive to make better decisions as more information becomes available by leverages online convex optimization and anytime algorithm concepts.