On November 26 last year I successfully defended my dissertation at Karlsruhe Institute of Technology. The full text has been published through the KIT library with the DOI 10.5445/IR/1000127719.
If you find any mistakes or have corrections of any kind, please email me at my name at this domain. I will add them to the list of corrections and remarks.
Abstract
This dissertation focuses on three fundamental problem families in big data systems, for which we develop communication-efficient probabilistic algorithms. In the first part, we consider various selection problems, in the second, weighted sampling problems, and the third, checking of basic operations in big data systems. This is motivated by a growing need for communication efficiency as the network and its usage increasingly dominate supercomputer system cost and energy consumption as well as running times of distributed applications. Surprisingly few communication-efficient algorithms are known for fundamental big data problems, and we close several of these gaps.
We first consider different selection problems, starting with selecting the element with rank k from a large, distributed input. There, we show how this can be achieved without assuming a random distribution of the input by redesigning the pivot selection step. Next, we show that selection from locally sorted sequences—also known as multisequence selection—becomes considerably easier if we allow the precise rank of the output to vary in some range. We then describe how this can be used to construct a bulk priority queue, which we later use to construct an algorithm for weighted reservoir sampling. Lastly, we consider finding the globally most frequent objects as well as those whose associated values add up to the highest sums with a sample-based approach.
In the weighted sampling chapter, we begin by giving new construction algorithms for a classical data structure for weighted sampling, alias tables. After presenting the first linear-time construction algorithm with only constant memory overhead, we proceed to parallelise this algorithm for shared memory, obtaining the first parallel construction algorithm for alias tables. We then show how to approach the problem in distributed memory with a two-level algorithm. Next, we present an output-sensitive algorithm for weighted sampling with replacement, meaning that it requires time linear in the number of unique items in the sample. This algorithm works sequentially as well as in shared and distributed memory and is the first such algorithm in all three categories. We subsequently adapt it to weighted sampling without replacement by combining it with an estimator for the number of unique items in a sample with replacement. Poisson sampling, a generalisation of Bernoulli sampling to weighted items, can be reduced to integer sorting, and we show how an existing approach can be parallelised. For sampling from data streams, we adapt a sequential approach and show how it can be parallelised in a mini-batch model using our bulk priority queue introduced in the selection chapter. The chapter is concluded by an extensive evaluation of our alias table construction algorithms, our output-sensitive algorithm for weighted sampling with replacement, and our weighted reservoir sampling algorithm.
To probabilistically verify the correctness of distributed algorithms, we propose checkers for fundamental operations of big data systems. We show that checking numerous operations can be reduced onto two ‘core’ checkers, namely checking aggregations and whether one sequence is a permutation of another. While multiple approaches exist for the latter and can be easily parallelised, our sum aggregation checker is a novel application of the same data structure that underpins counting Bloom filters and count-min sketches. We implemented both checkers in Thrill, a big data framework. Experiments with deliberately introduced errors confirm the detection accuracy predicted by theory, even when using real-world hash functions with non-ideal properties. Scaling experiments on a supercomputer show that our checkers have very little runtime overhead, in the area of 2 % for near-certainty in the correctness of the result.