Training of modern large neural networks (NN) requires a combination of parallelization strategies encompassing data, model, or optimizer sharding. When strategies increase in complexity, it becomes necessary for partitioning tools to be 1) expressive, allowing the composition of simpler strategies, and 2) predictable to estimate performance analytically. We present PartIR, our design for a NN partitioning system. PartIR is focused on an incremental approach to rewriting and is hardware-and-runtime agnostic. We present a simple but powerful API for composing sharding strategies and a simulator to validate them. The process is driven by high-level programmer-issued partitioning tactics, which can be both manual and automatic. Importantly, the tactics are specified separately from the model code, making them easy to change. We evaluate PartIR on several different models to demonstrate its predictability, expressibility, and ability to reach peak performance..

Many important problems involving molecular property prediction from 3D structures have limited data, posing a generalization challenge for neural networks. In this paper, we describe a pre-training technique based on denoising that achieves a new state-of-the-art in molecular property prediction by utilizing large datasets of 3D molecular structures at equilibrium to learn meaningful representations for downstream tasks. Relying on the well-known link between denoising autoencoders and score-matching, we show that the denoising objective corresponds to learning a molecular force field -- arising from approximating the Boltzmann distribution with a mixture of Gaussians -- directly from equilibrium structures. Our experiments demonstrate that using this pre-training objective significantly improves performance on multiple benchmarks, achieving a new state-of-the-art on the majority of targets in the widely used QM9 dataset. Our analysis then provides practical insights into the effects of different factors -- dataset sizes, model size and architecture, and the choice of upstream and downstream datasets -- on pre-training.

Many important problems involving molecular property prediction from 3D structures have limited data, posing a generalization challenge for neural networks. In this paper, we describe a pre-training technique based on denoising that achieves a new state-of-the-art in molecular property prediction by utilizing large datasets of 3D molecular structures at equilibrium to learn meaningful representations for downstream tasks. Relying on the well-known link between denoising autoencoders and score-matching, we show that the denoising objective corresponds to learning a molecular force field – arising from approximating the Boltzmann distribution with a mixture of Gaussians – directly from equilibrium structures. Our experiments demonstrate that using this pre-training objective significantly improves performance on multiple benchmarks, achieving a new state-of-the-art on the majority of targets in the widely used QM9 dataset. Our analysis then provides practical insights into the effects of different factors – dataset sizes, model size and architecture, and the choice of upstream and downstream datasets – on pre-training.

Large neural network models are commonly trained through a combination of advanced parallelism strategies in a single program, multiple data (SPMD) paradigm. For example, training large transformer models requires combining data, model, and pipeline partitioning; and optimizer sharding techniques. However, identifying efficient combinations for many model architectures and accelerator systems requires significant manual analysis. In this work, we present an automatic partitioner that identifies these combinations through a goal-oriented search. Our key findings are that a Monte Carlo Tree Search-based partitioner leveraging partition-specific compiler analysis directly into the search and guided goals matches expert-level strategies for various models.

We present evidence that learned density functional theory (``DFT'') force fields are ready for ground state catalyst discovery. Our key finding is that relaxation using forces from a learned potential yields structures with similar or lower energy to those relaxed using the RPBE functional in over 50\% of evaluated systems, despite the fact that the predicted forces differ significantly from the ground truth. This has the surprising implication that learned potentials may be ready for replacing DFT in challenging catalytic systems such as those found in the Open Catalyst 2020 dataset. Furthermore, we show that a force field trained on a locally harmonic energy surface with the same minima as a target DFT energy is also able to find lower or similar energy structures in over 50\% of cases. This ``Easy Potential'' converges in fewer steps than a standard model trained on true energies and forces, which further accelerates calculations. Its success illustrates a key point: learned potentials can locate energy minima even when the model has high force errors. The main requirement for structure optimisation is simply that the learned potential has the correct minima. Since learned potentials are fast and scale linearly with system size, our results open the possibility of quickly finding ground states for large systems.

In this paper we show that simple noise regularisation can be an effective way to address GNN oversmoothing. First we argue that regularisers addressing oversmoothing should both penalise node latent similarity and encourage meaningful node representations. From this observation we derive "Noisy Nodes", a simple technique in which we corrupt the input graph with noise, and add a noise correcting node-level loss. The diverse node level loss encourages latent node diversity, and the denoising objective encourages graph manifold learning. Our regulariser applies well-studied methods in simple, straightforward ways which allow even generic architectures to overcome oversmoothing and achieve state of the art results on quantum chemistry tasks, and improve results significantly on Open Graph Benchmark (OGB) datasets. Our results suggest Noisy Nodes can serve as a complementary building block in the GNN toolkit.

The rapid rise in demand for training large neural network architectures has brought into focus the need for partitioning strategies, for example by using data, model, or pipeline parallelism. Implementing these methods is increasingly supported through program primitives, but identifying efficient partitioning strategies requires expensive experimentation and expertise. We present the prototype of an automated partitioner that seamlessly integrates into existing compilers and existing user workflows. Our partitioner enables SPMD-style parallelism that encompasses data parallelism and parameter/activation sharding. Through a combination of inductive tactics and search in a platform-independent partitioning IR, automap can recover expert partitioning strategies such as Megatron sharding for transformer layers.

Adaptive instruction for online education can increase learning gains and decrease the work required of learners, instructors, and course designers. Reinforcement Learning (RL) is a promising tool for developing instructional policies, as RL models can learn complex relationships between course activities, learner actions, and educational outcomes. This paper demonstrates the first RL model to schedule educational activities in real time for a large online course through active learning. Our model learns to assign a sequence of course activities while maximizing learning gains and minimizing the number of items assigned. Using a controlled experiment with over 1,000 learners, we investigate how this scheduling policy affects learning gains, dropout rates, and qualitative learner feedback. We show that our model produces better learning gains using fewer educational activities than a linear assignment condition, and produces similar learning gains to a self-directed condition using fewer educational activities and with lower dropout rates.

Reinforcement learning (RL) tasks are challenging to implement, execute and test due to algorithmic instability, hyper-parameter sensitivity, and heterogeneous distributed communication patterns. We argue for the separation of logical component composition, backend graph definition, and distributed execution. To this end, we introduce RLgraph, a library for designing and executing reinforcement learning tasks in both static graph and define-by-run paradigms. The resulting implementations are robust, incrementally testable, and yield high performance across different deep learning frameworks and distributed backends.

Configuration spaces for computer systems can be challenging for traditional and automatic tuning strategies. Injecting task-specific knowledge into the tuner for a task may allow for more efficient exploration of candidate configurations. We apply this idea to the task of index set selection to accelerate database workloads. Index set selection has been amenable to recent applications of vanilla deep RL, but real deployments remain out of reach. In this paper, we explore how learning index selection can be enhanced with task-specific inductive biases, specifically by encoding these inductive biases in better action structures. Index selection-specific action representations arise when the problem is reformulated in terms of permutation learning and we rely on recent work for learning RL policies on permutations. Through this approach, we build an indexing agent that is able to achieve improved indexing and validate its behavior with task-specific statistics. Early experiments reveal that our agent can find configurations that are up to 40% smaller for the same levels of latency as compared with other approaches and indicate more intuitive indexing behavior.

Reinforcement learning frameworks have introduced abstractions to implement and execute algorithms at scale. They assume standardized simulator interfaces but are not concerned with identifying suitable task representations. We present Wield, a first-of-its kind system to facilitate task design for practical reinforcement learning. Through software primitives, Wield enables practitioners to decouple system-interface and deployment-specific configuration from state and action design. To guide experimentation, Wield further introduces a novel task design protocol and classification scheme centred around staged randomization to incrementally evaluate model capabilities.

Reinforcement learning (RL) tasks are challenging to implement, execute and test due to algorithmic instability, hyper-parameter sensitivity, and heterogeneous distributed communication patterns. We argue for the separation of logical component composition, backend graph definition, and distributed execution. To this end, we introduce RLgraph, a library for designing and executing high performance RL computation graphs in both static graph and define-by-run paradigms. The resulting implementations yield high performance across different deep learning frameworks and distributed backends.

Reinforcement learning approaches have long appealed to the data management community due to their ability to learn to control dynamic behavior from raw system performance. Recent successes in combining deep neural networks with reinforcement learning have sparked significant new interest in this domain. However, practical solutions remain elusive due to large training data requirements, algorithmic instability, and lack of standard tools. In this work, we introduce LIFT, an end-to-end software stack for applying deep reinforcement learning to data management tasks. While prior work has frequently explored applications in simulations, LIFT centers on utilizing human expertise to learn from demonstrations, thus lowering online training times. We further introduce TensorForce, a TensorFlow library for applied deep reinforcement learning exposing a unified declarative interface to common RL algorithms, thus providing a backend to LIFT. We demonstrate the utility of LIFT in two case studies in database compound indexing and resource management in stream processing. Results show LIFT controllers initialized from demonstrations can outperform human baselines and heuristics across latency metrics and space usage by up to 70%.

Due to their complexity, modern systems expose many configuration parameters which users must tune to maximize performance. Auto-tuning has emerged as an alternative in which a black-box optimizer iteratively evaluates configurations to find efficient ones. Unfortunately, for many systems, such as distributed systems, evaluating performance takes too long and the space of configurations is too large for the optimizer to converge within a reasonable time. We present BOAT, a framework which allows developers to build efficient bespoke auto-tuners for their system, in situations where generic auto-tuners fail. At BOAT's core is structured Bayesian optimization (SBO), a novel extension of the Bayesian optimization algorithm. SBO leverages contextual information provided by system developers, in the form of a probabilistic model of the system's behavior, to make informed decisions about which configurations to evaluate. In a case study, we tune the scheduling of a neural network computation on a heterogeneous cluster. Our auto-tuner converges within ten iterations. The optimized configurations outperform those found by generic auto-tuners in thirty iterations by up to 2X.

Today, web performance is primarily governed by round-trip latencies between end devices and cloud services. To improve performance, services need to minimize the delay of accessing data. In this paper, we propose a novel approach to low latency that relies on existing content delivery and web caching infrastructure. The main idea is to enable application-independent caching of query results and records with tunable consistency guarantees, in particular bounded staleness. Q uaestor (Query Store) employs two key concepts to incorporate both expiration-based and invalidation-based web caches: (1) an Expiring Bloom Filter data structure to indicate potentially stale data, and (2) statistically derived cache expiration times to maximize cache hit rates. Through a distributed query invalidation pipeline, changes to cached query results are detected in real-time. The proposed caching algorithms offer a new means for data-centric cloud services to trade latency against staleness bounds, e.g. in a database-as-a-service. Q uaestor is the core technology of the backend-as-a-service platform Baqend, a cloud service for low-latency websites. We provide empirical evidence for Q uaestor 's scalability and performance through both simulation and experiments. The results indicate that for read-heavy workloads, up to tenfold speed-ups can be achieved through Q uaestor 's caching.

Today, web performance is primarily governed by round-trip latencies between end devices and cloud services. To improve performance, services need to minimize the delay of accessing data. In this paper, we propose a novel approach to low latency that relies on existing content delivery and web caching infrastructure. The main idea is to enable application-independent caching of query results and records with tunable consistency guarantees, in particular bounded staleness. QUAESTOR (Query Store) employs two key concepts to incorporate both expiration-based and invalidation-based web caches: (1) an Expiring Bloom Filter data structure to indicate potentially stale data, and (2) statistically derived cache expiration times to maximize cache hit rates. Through a distributed query invalidation pipeline, changes to cached query results are detected in real-time. The proposed caching algorithms offer a new means for data-centric cloud services to trade latency against staleness bounds, e.g. in a database-as-a-service. QUAESTOR is the core technology of the backend-as-a-service platform Baqend, a cloud service for low-latency websites. We provide empirical evidence for QUAESTOR's scalability and performance through both simulation and experiments. The results indicate that for read-heavy workloads, up to tenfold speed-ups can be achieved through QUAESTOR's caching.

We present an optimizer which uses Bayesian optimization to tune the system parameters of distributed stochastic gradient descent (SGD). Given a specific context, our goal is to quickly find efficient configurations which appropriately balance the load between the available machines to minimize the average SGD iteration time. Our experiments consider setups with over thirty parameters. Traditional Bayesian optimization, which uses a Gaussian process as its model, is not well suited to such high dimensional domains. To reduce convergence time, we exploit the available structure. We design a probabilistic model which simulates the behavior of distributed SGD and use it within Bayesian optimization. Our model can exploit many runtime measurements for inference per evaluation of the objective function. Our experiments show that our resulting optimizer converges to efficient configurations within ten iterations, the optimized configurations outperform those found by generic optimizer in thirty iterations by up to 2X.

Consistency is a key requirement for data-intensive applications. Various schemes exist to express application requirements ranging from traditional transactional requirements to weaker isolation models that sacrifice data safety for performance. In recent years, the rise of interactive applications has brought with it a new generation of programming frameworks. These frameworks enable data management through simplified request APIs to web-based data storage. Foregoing strong consistency guarantees in favour of scalable performance, the task of ensuring application integrity is transferred to the application logic instead being ensured by a database management system [3]. Accepting anomalies leads to the requirement of compensation actions. For instance, a web shop has to issue cancellations if an inventory is oversold. For some applications, the cost of compensation could outweigh the performance penalty of running a strongly consistent service. Other applications like social networks might not require any compensatory action at all because typical anomalies like off-by-one errors on follow-counters do not affect user experience. For application developers, a strongly consistent view on the database is the most intuitive approach since in this case application code does not need to consider anomalies. Existing solutions for adaptive consistency rely heavily on manual modification of application code, static invariant analysis and profiling, putting the burden of managing multiple consistency levels on developers. In this proposal, I plan to outline a potential solution to adaptive consistency that omits consistency overhead from developers. While there has been some exploratory work into consistency SLAs [7], these approaches rely on closed form analytical models to deliver on the desired trade-offs. In my PhD work, I will investigate model-free online learning techniques that can flexibly learn SLAs without relying on explicitly encoded knowledge of the infrastructure.

Learning effective configurations in computer systems without hand-crafting models for every parameter is a long-standing problem. This paper investigates the use of deep reinforcement learning for runtime parameters of cloud databases under latency constraints. Cloud services serve up to thousands of concurrent requests per second and can adjust critical parameters by leveraging performance metrics. In this work, we use continuous deep reinforcement learning to learn optimal cache expirations for HTTP caching in content delivery networks. To this end, we introduce a technique for asynchronous experience management called delayed experience injection, which facilitates delayed reward and next-state computation in concurrent environments where measurements are not immediately available. Evaluation results show that our approach based on normalized advantage functions and asynchronous CPU-only training outperforms a statistical estimator.