Brief Announcement: Object Replication Degree Customization for High Availability∗

Object replication is commonly employed to enhance the availability of data-intensive services. As far we we know, existing availability-oriented replication schemes are oblivious to object request popularities when determining object replication degrees. However, many large-scale data-intensive applications contain objects with highly skewed data object request popularity distributions. Such nonuniform popularities may be exploited to improve system availability under a given space constraint. Intuitively, using more replicas for popular objects (less for unpopular ones) can increase the overall expected service availability while keeping the total space cost unchanged. Here by expected service availability, we mean the proportion of successful service requests (whose requested objects are available) in all requests. Using a simplistic model, we propose a novel replication degree policy that maximizes the availability for systems with known object popularities and sizes. Specifically, we find that the optimal replication degree for each object i should be linear in log ri si , where ri is the normalized object popularity and si is the normalized object size. Our evaluation uses applications driven by large real-life system data object request traces and failure traces. Results show that our proposed customization can achieve significant system availability increase over uniform replication.