Skip to main content
eScholarship
Open Access Publications from the University of California

Efficient and Secure Management of Warehouse-Scale Computers

  • Author(s): Islam, Mohammad Atiqul
  • Advisor(s): Ren, Shaolei
  • et al.
Creative Commons Attribution 4.0 International Public License
Abstract

Warehouse-scale computers or data centers are booming both in numbers and sizes. Consequently, data centers have been receiving major research attention in the recent years. However, prior literature primarily focuses on Google-type hyper-scale data centers and overlook the important segment of the multi-tenant colocation data centers where multiple tenants rent power and space for their physical servers while the data center operator manages the non-IT infrastructure like the power and cooling. Multi-tenant data centers are widely used across various industry sectors and hence efficient management of multi-tenant data centers is crucial.

However, many existing efficient operation approaches cannot be applied in multi-tenant data centers because the IT-equipment (e.g., servers) are owned by different tenants and therefore the data center operator has no direct control over them. In this dissertation research, I propose market-based techniques for coordination between the tenants and the operator towards efficient data center operation. Specifically, I propose an incentive framework that pays tenants for energy reduction such that the operator's overall cost is minimized. I also propose a novel market design that allows tenants to temporarily acquire additional capacity from other tenants' unused capacity for performance boosts.

Further, the criticality of the hosted services makes data centers a prime target for attacks. While data center cyber-security has been extensively studied, the equally important security aspect - data center physical security - remained unchecked. In this dissertation, I identify that an adversary disguised as a tenant in a multi-tenant data center can launch power attacks to create overloads in the power infrastructure. However, launching power attacks requires careful timing. Specifically, an attacker needs to estimate other tenants’ power consumption to time its malicious load injection to create overloads. I identify the existence of multiple side channels that can assist in attacker’s timing. I show that there exists a thermal side channel due to server heat recirculation, an acoustic side channel due to server fan noise, and a voltage side channel due to Ohm’s Law that can reveal the benign tenants' power consumption to an attacker. I also discuss the merits and challenges of possible countermeasures against these attacks.

Main Content
Current View