We present a work in progress report on applying deforestation to distributed, dynamic dataflow programming models. We propose a novel algorithm, dynamic path contraction, that applies and reverses optimizations to a distributed dataflow application as the program executes. With this algorithm, data and control flow is tracked by the runtime system used to identify potential optimizations as the system is running. We demonstrate and present preliminary results regarding this technique on an actor-based distributed programming model, Lasp, implemented on the Erlang virtual machine.
The task-based dataflow programming model has emerged as an alternative to the process-centric programming model for extreme-scale applications. However, load balancing is still a challenge in task-based dataflow runtimes. In this paper, we present extensions to the PaR-SEC runtime to demonstrate that distributed work stealing is an effective load-balancing method for task-based dataflow runtimes. In contrast to shared-memory work stealing, we find that each process should consider future tasks and the expected waiting time for execution when determining whether to steal. We demonstrate the ef…
We design and implement Megaphone, a data migration mechanism for stateful distributed dataflow engines with latency objectives. When compared to existing migration mechanisms, Megaphone has the following differentiating characteristics: (i) migrations can be subdivided to a configurable granularity to avoid latency spikes, and (ii) migrations can be prepared ahead of time to avoid runtime coordination. Megaphone is implemented as a library on an unmodified timely dataflow implementation, and provides an operator interface compatible with its existing APIs. We evaluate Megaphone on established…
Given a large graph, a graph sample determines a subgraph with similar characteristics for certain metrics of the original graph. The samples are much smaller thereby accelerating and simplifying the analysis and visualization of large graphs. We focus on the implementation of distributed graph sampling for Big Data frameworks and in-memory dataflow systems such as Apache Spark or Apache Flink. We evaluate the scalability of the new implementations and analyze to what degree the sampling approaches preserve certain graph metrics compared to the original graph. The latter analysis also uses com…
Analyzing large datasets with distributed dataflow systems requires the use of clusters. Public cloud providers offer a large variety and quantity of resources that can be used for such clusters. However, picking the appropriate resources in both type and number can often be challenging, as the selected configuration needs to match a distributed dataflow job's resource demands and access patterns. A good cluster configuration avoids hardware bottlenecks and maximizes resource utilization, avoiding costly overprovisioning. We propose a collaborative approach for finding optimal cluster configur…
We present ST2, an end-to-end solution to analyze distributed dataflows in an online setting. It is powered by Timely Dataflow, a low-latency, distributed data-parallel dataflow computational framework, and expands on its predecessor SnailTrail 1, a system to run online critical path analysis on program activity graphs derived from dataflow execution traces. ST2 connects to a running Timely computation, creates the program activity graph representation, and runs multiple analyses on top of it. Analyses include aggregate metrics, progress and temporal invariant checking, and graph pattern match…
Researchers and practitioners in the field of reinforcement learning (RL) frequently leverage parallel computation, which has led to a plethora of new algorithms and systems in the last few years. In this paper, we re-examine the challenges posed by distributed RL and try to view it through the lens of an old idea: distributed dataflow. We show that viewing RL as a dataflow problem leads to highly composable and performant implementations. We propose RLlib Flow, a hybrid actor-dataflow programming model for distributed RL, and validate its practicality by porting the full suite of algorithms i…
Distributed dataflow systems like Spark and Flink enable the use of clusters for scalable data analytics. While runtime prediction models can be used to initially select appropriate cluster resources given target runtimes, the actual runtime performance of dataflow jobs depends on several factors and varies over time. Yet, in many situations, dynamic scaling can be used to meet formulated runtime targets despite significant performance variance. This paper presents Enel, a novel dynamic scaling approach that uses message propagation on an attributed graph to model dataflow jobs and, thus, allo…
We present the design of a new large scale orchestration layer for accelerators. Our system, Pathways, is explicitly designed to enable exploration of new systems and ML research ideas, while retaining state of the art performance for current models. Pathways uses a sharded dataflow graph of asynchronous operators that consume and produce futures, and efficiently gang-schedules heterogeneous parallel computations on thousands of accelerators while coordinating data transfers over their dedicated interconnects. Pathways makes use of a novel asynchronous distributed dataflow design that lets the…
Distributed dataflow systems enable the use of clusters for scalable data analytics. However, selecting appropriate cluster resources for a processing job is often not straightforward. Performance models trained on historical executions of a concrete job are helpful in such situations, yet they are usually bound to a specific job execution context (e.g. node type, software versions, job parameters) due to the few considered input parameters. Even in case of slight context changes, such supportive models need to be retrained and cannot benefit from historical execution data from related context…
Swim extends the actor model to support applications composed of linked distributed actors that continuously analyze boundless streams of events from millions of sources, to respond in-sync with the real-world. Swim builds a running application from streaming events, creating a distributed dataflow graph of linked, stateful, concurrent streaming actors that is overlaid on a mesh of runtime instances. Streaming actors are vertices in the dataflow graph that concurrently analyze new events and modify their states. A link is an edge in the graph and is a URI binding to an actor's streaming API. T…
Imperial College London（帝国理工学院）；Huawei Technologies Co., Ltd.（华为技术有限公司）；Huawei R&D United Kingdom（华为英国研发中心）；Hong Kong University of Science and Technology（香港科技大学）
Reinforcement learning (RL) trains many agents, which is resource-intensive and must scale to large GPU clusters. Different RL training algorithms offer different opportunities for distributing and parallelising the computation. Yet, current distributed RL systems tie the definition of RL algorithms to their distributed execution: they hard-code particular distribution strategies and only accelerate specific parts of the computation (e.g. policy network updates) on GPU workers. Fundamentally, current systems lack abstractions that decouple RL algorithms from their execution. We describe MindSp…
Prediction serving systems are designed to provide large volumes of low-latency inferences machine learning models. These systems mix data processing and computationally intensive model inference and benefit from multiple heterogeneous processors and distributed computing resources. In this paper, we argue that a familiar dataflow API is well-suited to this latency-sensitive task, and amenable to optimization even with unmodified black-box ML models. We present the design of Cloudflow, a system that provides this API and realizes it on an autoscaling serverless backend. Cloudflow transparently…
This paper tries to reduce the effort of learning, deploying, and integrating several frameworks for the development of e-Science applications that combine simulations with High-Performance Data Analytics (HPDA). We propose a way to extend task-based management systems to support continuous input and output data to enable the combination of task-based workflows and dataflows (Hybrid Workflows from now on) using a single programming model. Hence, developers can build complex Data Science workflows with different approaches depending on the requirements. To illustrate the capabilities of Hybrid …
We present Tierkreis, a higher-order dataflow graph program representation and runtime designed for compositional, quantum-classical hybrid algorithms. The design of the system is motivated by the remote nature of quantum computers, the need for hybrid algorithms to involve cloud and distributed computing, and the long-running nature of these algorithms. The graph-based representation reflects how designers reason about and visualise algorithms, and allows automatic parallelism and asynchronicity. A strong, static type system and higher-order semantics allow for high expressivity and compositi…
Distributed dataflow systems enable data-parallel processing of large datasets on clusters. Public cloud providers offer a large variety and quantity of resources that can be used for such clusters. Yet, selecting appropriate cloud resources for dataflow jobs - that neither lead to bottlenecks nor to low resource utilization - is often challenging, even for expert users such as data engineers. We present C3O, a collaborative system for optimizing data processing cluster configurations in public clouds based on shared historical runtime data. The shared data is utilized for predicting the runti…
Distributed dataflow systems like Apache Flink and Apache Spark simplify processing large amounts of data on clusters in a data-parallel manner. However, choosing suitable cluster resources for distributed dataflow jobs in both type and number is difficult, especially for users who do not have access to previous performance metrics. One approach to overcoming this issue is to have users share runtime metrics to train context-aware performance models that help find a suitable configuration for the job at hand. A problem when sharing runtime data instead of trained models or model parameters is …
Edge computing has been emerging as a popular scenario for model inference. However, the inference performance on edge devices (e.g., Multi-Core DSP, FGPA, etc.) suffers from inefficiency due to the lack of highly optimized inference frameworks. Previous model inference frameworks are mainly developed in an operator-centric way, which provides insufficient acceleration to edge-based inference. Besides, the operator-centric framework incurs significant costs for continuous development and maintenance. In this paper, we propose Xenos, which can automatically conduct dataflow-centric optimization…
The Density Matrix Renormalization Group (DMRG) algorithm is a powerful tool for solving eigenvalue problems to model quantum systems. DMRG relies on tensor contractions and dense linear algebra to compute properties of condensed matter physics systems. However, its efficient parallel implementation is challenging due to limited concurrency, large memory footprint, and tensor sparsity. We mitigate these problems by implementing two new parallel approaches that handle block sparsity arising in DMRG, via Cyclops, a distributed memory tensor contraction library. We benchmark their performance on …
We provide a simple new randomized contraction approach to the global minimum cut problem for simple undirected graphs. The contractions exploit 2-out edge sampling from each vertex rather than the standard uniform edge sampling. We demonstrate the power of our new approach by obtaining better algorithms for sequential, distributed, and parallel models of computation. Our end results include the following randomized algorithms for computing edge connectivity with high probability: -- Two sequential algorithms with complexities $O(m \log n)$ and $O(m+n \log^3 n)$. These improve on a long line o…
Miller and Reif's FOCS'85 classic and fundamental tree contraction algorithm is a broadly applicable technique for the parallel solution of a large number of tree problems. Additionally it is also used as an algorithmic design technique for a large number of parallel graph algorithms. In all previously explored models of computation, however, tree contractions have only been achieved in $Ω(\log n)$ rounds of parallel run time. In this work, we not only introduce a generalized tree contraction method but also show it can be computed highly efficiently in $O(1/ε^3)$ rounds in the Adaptive Massiv…
Simulation of quantum systems is challenging due to the exponential size of the state space. Tensor networks provide a systematically improvable approximation for quantum states. 2D tensor networks such as Projected Entangled Pair States (PEPS) are well-suited for key classes of physical systems and quantum circuits. However, direct contraction of PEPS networks has exponential cost, while approximate algorithms require computations with large tensors. We propose new scalable algorithms and software abstractions for PEPS-based methods, accelerating the bottleneck operation of contraction and re…
Tensor computations are increasingly prevalent numerical techniques in data science, but pose unique challenges for high-performance implementation. We provide novel algorithms and systems infrastructure which enable efficient parallel implementation of algorithms for tensor completion with generalized loss functions. Specifically, we consider alternating minimization, coordinate minimization, and a quasi-Newton (generalized Gauss-Newton) method. By extending the Cyclops library, we implement all of these methods in high-level Python syntax. To make possible tensor completion for very sparse t…
