Caching and Distributed Storage

A simple task of storing a database or transferring it to a different point via a communication channel turns far more complex as the size of the database grows large. Limited bandwidth available for transmission plays a central role in this predicament. In two broad contexts, Content Distribution Networks (CDN) and Distributed Storage Systems (DSS), the adverse effect of the growing size of the database on the transmission bandwidth can be mitigated by exploiting additional storage units. Characterizing the optimal tradeoff between the transmission bandwidth and the storage size is the central quest to numerous works in the recent literature, including this thesis.

In a DSS, individual servers fail routinely and must be replicated by downloading data from the remaining servers, a task referred to as the repair process. To render this process of repairing failed servers more straightforward and efficient, various forms of redundancy can be introduced in the system. One of the benchmarks by which the reliability of a DSS is measured is availability, which refers to the number of disjoint sets of servers that can help to repair any failed server. We study the interaction of this parameter with the amount of traffic generated during the repair process (the repair bandwidth) and the storage size. In particular, we propose a novel DSS architecture which can achieve much smaller repair bandwidth for the same availability, compared to the state of the art.

In the context of CDNs, the network can be highly congested during certain hours of the day and almost idle at other times. This variability of traffic can be reduced by utilizing local storage units that prefetch the data while the network is idle. This approach is referred to as caching. In this thesis we analyze a CDN that has access to independent data from various content providers. We characterize the best caching strategy in terms of the aggregate peak traffic under the constraint that coding across contents from different libraries is prohibited. Furthermore we prove that under certain set of conditions this restriction is without loss of optimality.

From Massive Parallelization to Quantum Computing: Seven Novel Approaches to Query Optimization

Immanuel Trummer

The goal of query optimization is to map a declarative query (describing data to generate) to a query plan (describing how to generate the data) with optimal execution cost. Query optimization is required to support declarative query interfaces. It is a core problem in the area of database systems and has received tremendous attention in the research community, starting with an initial publication in 1979. In this thesis, we revisit the query optimization problem. This visit is motivated by several developments that change the context of query optimization. That change is not reflected in prior literature. First, advances in query execution platforms and processing techniques have changed the context of query optimization. Novel provisioning models and processing techniques such as Cloud computing, crowdsourcing, or approximate processing allow to trade between different execution cost metrics (e.g., execution time versus monetary execution fees in case of Cloud computing). This makes it necessary to compare alternative execution plans according to multiple cost metrics in query optimization. While this is a common scenario nowadays, the literature on query optimization with multiple cost metrics (a generalization of the classical problem variant with one execution cost metric) is surprisingly sparse. While prior methods take hours to optimize even moderately sized queries when considering multiple cost metrics, we propose a multitude of approaches to make query optimization in such scenarios practical. A second development that we address in this thesis is the availability of novel software and hardware platforms that can be exploited for optimization. We will show that integer programming solvers, massively parallel clusters (which nowadays are commonly used for query execution), and adiabatic quantum annealers enable us to solve query optimization problem instances that are far beyond the capabilities of prior approaches. In summary, we propose seven novel approaches to query optimization that significantly increase the size of the problem instances that can be addressed (measured by the query size and by the number of considered execution cost metrics). Those novel approaches can be classified into three broad categories: moving query optimization before run time to relax constraints on optimization time, trading optimization time for relaxed optimality guarantees (leading to approximation schemes, incremental algorithms, and randomized algorithms for query optimization with multiple cost metrics), and reducing optimization time by leveraging novel software and hardware platforms (integer programming solvers, massively parallel clusters, and adiabatic quantum annealers). Those approaches are novel since they address novel problem variants of query optimization, introduced in this thesis, since they are novel for their respective problem variant (e.g., we propose the first randomized algorithm for query optimization with multiple cost metrics), or because they have never been used for optimization problems in the database domain (e.g., this is the first time that quantum computing is used to solve a database-specific optimization problem).

EPFL2016

Rack-Scale Memory Pooling for Datacenters

Stanko Novakovic

The rise of web-scale services has led to a staggering growth in user data on the Internet. To transform such a vast raw data into valuable information for the user and provide quality assurances, it is important to minimize access latency and enable in-memory processing. For more than a decade, the only practical way to accommodate for ever-growing data in memory has been to scale out server resources, which has led to the emergence of large-scale datacenters and distributed non-relational databases (NoSQL). Such horizontal scaling of resources translates to an increasing number of servers that participate in processing individual user requests. Typically, each user request results in hundreds of independent queries targeting different NoSQL nodes - servers, and the larger the number of servers involved, the higher the fan-out. To complete a single user request, all of the queries associated with that request have to complete first, and thus, the slowest query determines the completion time. Because of skewed popularity distributions and resource contention, the more servers we have, the harder it is to achieve high throughput and facilitate server utilization, without violating service level objectives. This thesis proposes rack-scale memory pooling (RSMP), a new scaling technique for future datacenters that reduces networking overheads and improves the performance of core datacenter software. RSMP is an approach to building larger, rack-scale capacity units for datacenters through specialized fabric interconnects with support for one-sided operations, and using them, in lieu of conventional servers (e.g. 1U), to scale out. We define an RSMP unit to be a server rack connecting 10s to 100s of servers to a secondary network enabling direct, low-latency access to the global memory of the rack. We, then, propose a new RSMP design - Scale-Out NUMA that leverages integration and a NUMA fabric to bridge the gap between local and remote memory to only 5× difference in access latency. Finally, we show how RSMP impacts NoSQL data serving, a key datacenter service used by most web-scale applications today. We show that using fewer larger data shards leads to less load imbalance and higher effective throughput, without violating applications¿ service level objectives. For example, by using Scale-Out NUMA, RSMP improves the throughput of a key-value store up to 8.2× over a traditional scale-out deployment.

EPFL2017

Single-server Multi-message Private Information Retrieval with Side Information

Michael Christoph Gastpar, Su Li

We study the problem of single-server multi-message private information retrieval with side information. One user wants to recover N out of K independent messages which are stored at a single server. The user initially possesses a subset of M messages as side information. The goal of the user is to download the N demand messages while not leaking any information about the indices of these messages to the server. In this paper, we characterize the minimum number of required transmissions for linear codes. Moreover, we present an optimal coding scheme which enables the user to download the demand messages and preserves the privacy of their indices by using linear codes. Additionally, we show that the trivial MDS coding scheme with K - M transmissions is optimal if N > M or N-2 + N >= K - M. This means if one wishes to privately download more than the square-root of the number of files in the database, then one must effectively download the full database (minus the side information), irrespective of the amount of side information one has available.