Application Aware Resource Management For Cloud Platforms PDF Download

Modern distributed server applications are hosted on enterprise or cloud data centers that provide computing, storage, and networking capabilities to these applications. These applications are built using the implicit assumption that the underlying servers will be stable and normally available, barring for occasional faults. In many emerging scenarios, however, data centers and clouds only provide transient, rather than continuous, availability of their servers. Transiency in modern distributed systems arises in many contexts, such as green data centers powered using renewable intermittent sources, and cloud platforms that provide lower-cost transient servers which can be unilaterally revoked by the cloud operator. Transient computing resources are increasingly important, and existing fault-tolerance and resource management techniques are inadequate for transient servers because applications typically assume continuous resource availability. This thesis presents research in distributed systems design that treats transiency as a first-class design principle. I show that combining transiency-specific fault-tolerance mechanisms with resource management policies to suit application characteristics and requirements, can yield significant cost and performance benefits. These mechanisms and policies have been implemented and prototyped as part of software systems, which allow a wide range of applications, such as interactive services and distributed data processing, to be deployed on transient servers, and can reduce cloud computing costs by up to 90\%. This thesis makes contributions to four areas of computer systems research: transiency-specific fault-tolerance, resource allocation, abstractions, and resource reclamation. For reducing the impact of transient server revocations, I develop two fault-tolerance techniques that are tailored to transient server characteristics and application requirements. For interactive applications, I build a derivative cloud platform that masks revocations by transparently moving application-state between servers of different types. Similarly, for distributed data processing applications, I investigate the use of application level periodic checkpointing to reduce the performance impact of server revocations. For managing and reducing the risk of server revocations, I investigate the use of server portfolios that allow transient resource allocation to be tailored to application requirements. Finally, I investigate how resource providers (such as cloud platforms) can provide transient resource availability without revocation, by looking into alternative resource reclamation techniques. I develop resource deflation, wherein a server's resources are fractionally reclaimed, allowing the application to continue execution albeit with fewer resources. Resource deflation generalizes revocation, and the deflation mechanisms and cluster-wide policies can yield both high cluster utilization and low application performance degradation.