Memory disaggregation separates compute (CPU) and main memory resources into disjoint physical units to enable elastic and independent scaling. Connected via high-speed RDMA-enabled networks, compute nodes can directly access remote memory. This setting often requires complex protocols with many network roundtrips as memory nodes have near-zero compute power. In this paper, we design a scalable distributed inverted list index for disaggregated memory architectures. An inverted list index maps a set of terms to lists of documents that contain this term. Current solutions either partition the index horizontally or vertically with severe limitations in the disaggregated memory setting due to data and access skew, high network latency, or out-of-memory errors. Our method partitions lists into fixed-size blocks and spreads them across the memory nodes to balance skewed accesses. Block-based list processing keeps the memory footprint of compute nodes low and masks latency by interleaving remote accesses with expensive list operations. In addition, we propose efficient updates with optimistic concurrency control and read-write conflict detection. Our experiments confirm the efficiency and scalability of our method.

Given two collections of sets, the set similarity join reports all pairs of sets that are within a given distance threshold. State-of-the-art solutions employ an inverted list index and several heuristics to compute the join result efficiently. Prefix-based solutions benefit from infrequent set elements, known as tokens, and spend considerable time scanning long lists if the token frequency is not sufficiently skewed. Partition-based methods are less sensitive to the token distribution but suffer from a significantly larger memory footprint, limiting their applicability as the threshold or the set sizes grow. Solutions from the domain of metric-based similarity search are designed to reduce the overall number of distance computations. Generic metric techniques cannot compete with state-of-the-art similarity joins tailored to sets, which in turn do not exploit metric filter opportunities.We propose MetricJoin, the first exact set similarity join technique that leverages the metric properties of set distance functions. In contrast to its competitors, MetricJoin is robust, i.e., datasets with different characteristics can be joined efficiently in terms of runtime and memory. Our algorithm embeds sets in vector space, organizes long inverted lists in spatial indexes, and employs an effective metric filter to prune unqualified sets. MetricJoin requires only linear space in the collection size and substantially reduces the number of sets that must be considered. In our performance studies, MetricJoin outperforms state-of-the-art solutions by up to an order of magnitude in runtime and generates up to five orders of magnitude fewer candidates.

Expressive text encoders such as RNNs and Transformer Networks have been at the center of NLP models in recent work.Most of the effort has focused on sentence-level tasks, capturing the dependencies between words in a single sentence, or pairs of sentences.However, certain tasks, such as argumentation mining, require accounting for longer texts and complicated structural dependencies between them.Deep structured prediction is a general framework to combine the complementary strengths of expressive neural encoders and structured inference for highly structured domains.Nevertheless, when the need arises to go beyond sentences, most work relies on combining the output scores of independently trained classifiers.One of the main reasons for this is that constrained inference comes at a high computational cost.In this paper, we explore the use of randomized inference to alleviate this concern and show that we can efficiently leverage deep structured prediction and expressive neural encoders for a set of tasks involving complicated argumentative structures.