Investigation of power attack to data center

• Main Author: Koh, Kiat
• Other Authors: Lu Rongxing
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/64662
• Institution: Nanyang Technological University

Due to the increased popularity of cloud computing services, there is a need to increase the amount of servers within the data centers. Thus, the rapid server deployment in data centers has caused their power distribution and cooling systems to approach peak capacity. However, it is very expensive to upgrade the power infrastructures of data centers and the related cost is commonly in hundreds of millions of dollars. Thus, power oversubscription, whereby more servers are added to the current power systems capacity, is now commonly used as a method to defer the costly upgrades. This is based on the assumption that peak power consumption of the server racks do not peak simultaneously with normal workload. A previous study from Google also indicates that this assumption is valid. However, power oversubscription induces a hidden risk in which there is a possibility that power consumption of the servers will exceed the power system capacity, causing the circuit breakers to trip and lead to undesired power outages. This can be due to malicious workloads that can generate power spikes on servers at the same time. Such malicious workload can be defined as power attacks. The power outrage would bring about great financial loss to organizations due to data center downtime. Hence, there is a need to survey on existing power attacks on data centers and investigate possible approaches to mitigate the power attacks. In this paper, several power attack vectors are surveyed, namely eDDoS attacks, cache misses and insider attacks. To conduct these attacks, an adversary would need a huge network of cloud users or an insider. Thus, preventive measures using cryptography for authentication and access control measures which include the use of honeywords and Shamir’s secret sharing scheme is discussed. In this paper, a new (2, 3) secret sharing scheme is proposed in which shares are periodically updated to prevent adversaries from using old shares from a previously compromised server. The properties of this proposed scheme is then analysed.