arXiv subjects

Philippas Tsigas

Publications and source records attributed to Philippas Tsigas.

At least 19 recordsLinked to original sources

arXiv2024

How to Relax Instantly: Elastic Relaxation of Concurrent Data Structures

The sequential semantics of many concurrent data structures, such as stacks and queues, inevitably lead to memory contention in parallel environments, thus limiting scalability. Semantic relaxation has the potential to address this issue, increasing the parallelism at the expense of weakened semantics. Although prior research has shown that improved performance can be attained by relaxing concurrent data structure semantics, there is no one-size-fits-all relaxation that adequately addresses the varying needs of dynamic executions. In this paper, we first introduce the concept of elastic relaxation and consequently present the Lateral structure, which is an algorithmic component capable of supporting the design of elastically relaxed concurrent data structures. Using the Lateral , we design novel elastically relaxed, lock-free queues and stacks capable of reconfiguring relaxation during run time. We establish linearizability and define upper bounds for relaxation errors in our designs. Experimental evaluations show that our elastic designs hold up against state-of-the-art statically relaxed designs, while also swiftly managing trade-offs between relaxation and operational latency. We also outline how to use the Lateral to design elastically relaxed lock-free counters and deques.

cs.DS↗

arXiv2023

Relaxing Concurrent Data-structure Semantics for Increasing Performance: A Multi-structure 2D Design Framework

There has been a significant amount of work in the literature proposing semantic relaxation of concurrent data structures for improving scalability and performance. By relaxing the semantics of a data structure, a bigger design space, that allows weaker synchronization and more useful parallelism, is unveiled. Investigating new data structure designs, capable of trading semantics for achieving better performance in a monotonic way, is a major challenge in the area. We algorithmically address this challenge in this paper.To address this challenge, we present an efficient lock-free, concurrent data structure design framework for out-of-order semantic relaxation. Our framework introduces a new two dimensional algorithmic design, that uses multiple instances of an implementation of the given data structure. The first dimension of our design is the number of data structure instances onto which operations are spread to, in order to achieve increased parallelism through disjoint memory accesses. The second dimension is the number of consecutive operations of a single thread that can stay at the same data structure instance in order to benefit from data locality. Our design can flexibly explore this two-dimensional space to achieve the property of monotonically increasing throughput performance via relaxing concurrent data structure semantics within a tight deterministic relaxation bound, as we prove in the paper. We show how our framework can instantiate lock-free out-of-order queues, stacks, counters and dequeues. The experimental evaluation shows that our two-dimensional data structures: i) significantly outperform the respected previous proposed ones with respect to scalability and throughput performance and ii) monotonically increase throughput as relaxation increases.

cs.DS↗

arXiv2022

STRETCH: Virtual Shared-Nothing Parallelism for Scalable and Elastic Stream Processing

Stream processing applications extract value from raw data through Directed Acyclic Graphs of data analysis tasks. Shared-nothing (SN) parallelism is the de-facto standard to scale stream processing applications. Given an application, SN parallelism instantiates several copies of each analysis task, making each instance responsible for a dedicated portion of the overall analysis, and relies on dedicated queues to exchange data among connected instances. On the one hand, SN parallelism can scale the execution of applications both up and out since threads can run task instances within and across processes/nodes. On the other hand, its lack of sharing can cause unnecessary overheads and hinder the scaling up when threads operate on data that could be jointly accessed in shared memory. This trade-off motivated us in studying a way for stream processing applications to leverage shared memory and boost the scale up (before the scale out) while adhering to the widely-adopted and SN-based APIs for stream processing applications. We introduce STRETCH, a framework that maximizes the scale up and offers instantaneous elastic reconfigurations (without state transfer) for stream processing applications. We propose the concept of Virtual Shared-Nothing (VSN) parallelism and elasticity and provide formal definitions and correctness proofs for the semantics of the analysis tasks supported by STRETCH, showing they extend the ones found in common Stream Processing Engines. We also provide a fully implemented prototype and show that STRETCH's performance exceeds that of state-of-the-art frameworks such as Apache Flink and offers, to the best of our knowledge, unprecedented ultra-fast reconfigurations, taking less than 40 ms even when provisioning tens of new task instances.

cs.DC↗

arXiv2021

Performance Modeling and Vertical Autoscaling of Stream Joins

Streaming analysis is widely used in cloud as well as edge infrastructures. In these contexts, fine-grained application performance can be based on accurate modeling of streaming operators. This is especially beneficial for computationally expensive operators like adaptive stream joins that, being very sensitive to rate-varying data streams, would otherwise require costly frequent monitoring. We propose a dynamic model for the processing throughput and latency of adaptive stream joins that run with different parallelism degrees. The model is presented with progressive complexity, from a centralized non-deterministic up to a deterministic parallel stream join, describing how throughput and latency dynamics are influenced by various configuration parameters. The model is catalytic for understanding the behavior of stream joins against different system deployments, as we show with our model-based autoscaling methodology to change the parallelism degree of stream joins during the execution. Our thorough evaluation, for a broad spectrum of parameter, confirms the model can reliably predict throughput and latency metrics with a fairly high accuracy, with the median error in estimation ranging from approximately 0.1% to 6.5%, even for an overloaded system. Furthermore, we show that our model allows to efficiently control adaptive stream joins by estimating the needed resources solely based on the observed input load. In particular, we show it can be employed to enable efficient autoscaling, even when big changes in the input load happen frequently (in the realm of seconds).

cs.DC↗

arXiv2021

Consistent Lock-free Parallel Stochastic Gradient Descent for Fast and Stable Convergence

Stochastic gradient descent (SGD) is an essential element in Machine Learning (ML) algorithms. Asynchronous parallel shared-memory SGD (AsyncSGD), including synchronization-free algorithms, e.g. HOGWILD!, have received interest in certain contexts, due to reduced overhead compared to synchronous parallelization. Despite that they induce staleness and inconsistency, they have shown speedup for problems satisfying smooth, strongly convex targets, and gradient sparsity. Recent works take important steps towards understanding the potential of parallel SGD for problems not conforming to these strong assumptions, in particular for deep learning (DL). There is however a gap in current literature in understanding when AsyncSGD algorithms are useful in practice, and in particular how mechanisms for synchronization and consistency play a role. We focus on the impact of consistency-preserving non-blocking synchronization in SGD convergence, and in sensitivity to hyper-parameter tuning. We propose Leashed-SGD, an extensible algorithmic framework of consistency-preserving implementations of AsyncSGD, employing lock-free synchronization, effectively balancing throughput and latency. We argue analytically about the dynamics of the algorithms, memory consumption, the threads' progress over time, and the expected contention. We provide a comprehensive empirical evaluation, validating the analytical claims, benchmarking the proposed Leashed-SGD framework, and comparing to baselines for training multilayer perceptrons (MLP) and convolutional neural networks (CNN). We observe the crucial impact of contention, staleness and consistency and show how Leashed-SGD provides significant improvements in stability as well as wall-clock time to convergence (from 20-80% up to 4x improvements) compared to the standard lock-based AsyncSGD algorithm and HOGWILD!, while reducing the overall memory footprint.

cs.DC↗

arXiv2019

MindTheStep-AsyncPSGD: Adaptive Asynchronous Parallel Stochastic Gradient Descent

Stochastic Gradient Descent (SGD) is very useful in optimization problems with high-dimensional non-convex target functions, and hence constitutes an important component of several Machine Learning and Data Analytics methods. Recently there have been significant works on understanding the parallelism inherent to SGD, and its convergence properties. Asynchronous, parallel SGD (AsyncPSGD) has received particular attention, due to observed performance benefits. On the other hand, asynchrony implies inherent challenges in understanding the execution of the algorithm and its convergence, stemming from the fact that the contribution of a thread might be based on an old (stale) view of the state. In this work we aim to deepen the understanding of AsyncPSGD in order to increase the statistical efficiency in the presence of stale gradients. We propose new models for capturing the nature of the staleness distribution in a practical setting. Using the proposed models, we derive a staleness-adaptive SGD framework, MindTheStep-AsyncPSGD, for adapting the step size in an online-fashion, which provably reduces the negative impact of asynchrony. Moreover, we provide general convergence time bounds for a wide class of staleness-adaptive step size strategies for convex target functions. We also provide a detailed empirical study, showing how our approach implies faster convergence for deep learning applications.

cs.DC↗

arXiv2018

Lock-Free Search Data Structures: Throughput Modelling with Poisson Processes

This paper considers the modelling and the analysis of the performance of lock-free concurrent search data structures. Our analysis considers such lock-free data structures that are utilized through a sequence of operations which are generated with a memoryless and stationary access pattern. Our main contribution is a new way of analysing lock-free search data structures: our execution model matches with the behavior that we observe in practice and achieves good throughput predictions. Search data structures are formed of linked basic blocks, usually referred as nodes, that can be accessed by two kinds of events, characterized by their latencies; (i) CAS events originated as a result of modifications of the search data structures (ii) Read events originated during traversals. This type of data structures are usually designed to accommodate a large number of data nodes, which makes the occurrence of an event on a given node rare at any given time. The throughput is defined by the number of events per operation in conjunction with the factors that impact the latencies of these events. We frame these impacting factors under capacity and coherence cache misses. In this context, we model the events as Poisson processes that we can merge and split to estimate the latencies of the events based on the interleaving of events from different threads, and in turn estimate the throughput. We have validated our analysis on several fundamental lock-free search data structures such as linked lists, hash tables, skip lists and binary trees.

cs.DS↗

arXiv2018

D2.2 White-box methodologies, programming abstractions and libraries

This deliverable reports the results of white-box methodologies and early results of the first prototype of libraries and programming abstractions as available by project month 18 by Work Package 2 (WP2). It reports i) the latest results of Task 2.2 on white-box methodologies, programming abstractions and libraries for developing energy-efficient data structures and algorithms and ii) the improved results of Task 2.1 on investigating and modeling the trade-off between energy and performance of concurrent data structures and algorithms. The work has been conducted on two main EXCESS platforms: Intel platforms with recent Intel multicore CPUs and Movidius Myriad1 platform. Regarding white-box methodologies, we have devised new relaxed cache-oblivious models and proposed a new power model for Myriad1 platform and an energy model for lock-free queues on CPU platforms. For Myriad1 platform, the im- proved model now considers both computation and data movement cost as well as architecture and application properties. The model has been evaluated with a set of micro-benchmarks and application benchmarks. For Intel platforms, we have generalized the model for concurrent queues on CPU platforms to offer more flexibility according to the workers calling the data structure (parallel section sizes of enqueuers and dequeuers are decoupled). Regarding programming abstractions and libraries, we have continued investigat- ing the trade-offs between energy consumption and performance of data structures such as concurrent queues and concurrent search trees based on the early results of Task 2.1.The preliminary results show that our concurrent trees are faster and more energy efficient than the state-of-the-art on commodity HPC and embedded platforms.

cs.DC↗

arXiv2018

D2.1 Models for energy consumption of data structures and algorithms

This deliverable reports our early energy models for data structures and algorithms based on both micro-benchmarks and concurrent algorithms. It reports the early results of Task 2.1 on investigating and modeling the trade-off between energy and performance in concurrent data structures and algorithms, which forms the basis for the whole work package 2 (WP2). The work has been conducted on the two main EXCESS platforms: (1) Intel platform with recent Intel multi-core CPUs and (2) Movidius embedded platform.

cs.DC↗

arXiv2018

D2.3 Power models, energy models and libraries for energy-efficient concurrent data structures and algorithms

This deliverable reports the results of the power models, energy models and libraries for energy-efficient concurrent data structures and algorithms as available by project month 30 of Work Package 2 (WP2). It reports i) the latest results of Task 2.2-2.4 on providing programming abstractions and libraries for developing energy-efficient data structures and algorithms and ii) the improved results of Task 2.1 on investigating and modeling the trade-off between energy and performance of concurrent data structures and algorithms. The work has been conducted on two main EXCESS platforms: Intel platforms with recent Intel multicore CPUs and Movidius Myriad platforms.

cs.DC↗

arXiv2018

D2.4 Report on the final prototype of programming abstractions for energy-efficient inter-process communication

Work package 2 (WP2) aims to develop libraries for energy-efficient inter-process communication and data sharing on the EXCESS platforms. The Deliverable D2.4 reports on the final prototype of programming abstractions for energy-efficient inter- process communication. Section 1 is the updated overview of the prototype of programming abstraction and devised power/energy models. The Section 2-6 contain the latest results of the four studies: i) GreenBST, a energy-efficient and concurrent search tree (cf. Section 2) ii) Customization methodology for implementation of streaming aggregation in embedded systems (cf. Section 3) iii) Energy Model on CPU for Lock-free Data-structures in Dynamic Environments (cf. Section 4.10) iv) A General and Validated Energy Complexity Model for Multithreaded Algorithms (cf. Section 5)

cs.DC↗

arXiv2016

How Lock-free Data Structures Perform in Dynamic Environments: Models and Analyses

In this paper we present two analytical frameworks for calculating the performance of lock-free data structures. Lock-free data structures are based on retry loops and are called by application-specific routines. In contrast to previous work, we consider in this paper lock-free data structures in dynamic environments. The size of each of the retry loops, and the size of the application routines invoked in between, are not constant but may change dynamically. The new frameworks follow two different approaches. The first framework, the simplest one, is based on queuing theory. It introduces an average-based approach that facilitates a more coarse-grained analysis, with the benefit of being ignorant of size distributions. Because of this independence from the distribution nature it covers a set of complicated designs. The second approach, instantiated with an exponential distribution for the size of the application routines, uses Markov chains, and is tighter because it constructs stochastically the execution, step by step. Both frameworks provide a performance estimate which is close to what we observe in practice. We have validated our analysis on (i) several fundamental lock-free data structures such as stacks, queues, deques and counters, some of them employing helping mechanisms, and (ii) synthetic tests covering a wide range of possible lock-free designs. We show the applicability of our results by introducing new back-off mechanisms, tested in application contexts, and by designing an efficient memory management scheme that typical lock-free algorithms can utilize.

cs.DC↗

arXiv2016

Efficient data streaming multiway aggregation through concurrent algorithmic designs and new abstract data types

Data streaming relies on continuous queries to process unbounded streams of data in a real-time fashion. It is commonly demanding in computation capacity, given that the relevant applications involve very large volumes of data. Data structures act as articulation points and maintain the state of data streaming operators, potentially supporting high parallelism and balancing the work between them. Prompted by this fact, in this work we study and analyze parallelization needs of these articulation points, focusing on the problem of streaming multiway aggregation, where large data volumes are received from multiple input streams. The analysis of the parallelization needs, as well as of the use and limitations of existing aggregate designs and their data structures, leads us to identify needs for proper shared objects that can achieve low-latency and high throughput multiway aggregation. We present the requirements of such objects as abstract data types and we provide efficient lock-free linearizable algorithmic implementations of them, along with new multiway aggregate algorithmic designs that leverage them, supporting both deterministic order-sensitive and order-insensitive aggregate functions. Furthermore, we point out future directions that open through these contributions. The paper includes an extensive experimental study, based on a variety of aggregation continuous queries on two large datasets extracted from SoundCloud, a music social network, and from a Smart Grid network. In all the experiments, the proposed data structures and the enhanced aggregate operators improved the processing performance significantly, up to one order of magnitude, in terms of both throughput and latency, over the commonly-used techniques based on queues.

cs.DS↗

arXiv2015

Shared-object System Equilibria: Delay and Throughput Analysis

We consider shared-object systems that require their threads to fulfill the system jobs by first acquiring sequentially the objects needed for the jobs and then holding on to them until the job completion. Such systems are in the core of a variety of shared-resource allocation and synchronization systems. This work opens a new perspective to study the expected job delay and throughput analytically, given the possible set of jobs that may join the system dynamically. We identify the system dependencies that cause contention among the threads as they try to acquire the job objects. We use these observations to define the shared-object system equilibria. We note that the system is in equilibrium whenever the rate in which jobs arrive at the system matches the job completion rate. These equilibria consider not only the job delay but also the job throughput, as well as the time in which each thread blocks other threads in order to complete its job. We then further study in detail the thread work cycles and, by using a graph representation of the problem, we are able to propose procedures for finding and estimating equilibria, i.e., discovering the job delay and throughput, as well as the blocking time. To the best of our knowledge, this is a new perspective, that can provide better analytical tools for the problem, in order to estimate performance measures similar to ones that can be acquired through experimentation on working systems and simulations, e.g., as job delay and throughput in (distributed) shared-object systems.

cs.DC↗

arXiv2015

Analyzing the Performance of Lock-Free Data Structures: A Conflict-based Model

This paper considers the modeling and the analysis of the performance of lock-free concurrent data structures. Lock-free designs employ an optimistic conflict control mechanism, allowing several processes to access the shared data object at the same time. They guarantee that at least one concurrent operation finishes in a finite number of its own steps regardless of the state of the operations. Our analysis considers such lock-free data structures that can be represented as linear combinations of fixed size retry loops. Our main contribution is a new way of modeling and analyzing a general class of lock-free algorithms, achieving predictions of throughput that are close to what we observe in practice. We emphasize two kinds of conflicts that shape the performance: (i) hardware conflicts, due to concurrent calls to atomic primitives; (ii) logical conflicts, caused by simultaneous operations on the shared data structure. We show how to deal with these hardware and logical conflicts separately, and how to combine them, so as to calculate the throughput of lock-free algorithms. We propose also a common framework that enables a fair comparison between lock-free implementations by covering the whole contention domain, together with a better understanding of the performance impacting factors. This part of our analysis comes with a method for calculating a good back-off strategy to finely tune the performance of a lock-free algorithm. Our experimental results, based on a set of widely used concurrent data structures and on abstract lock-free designs, show that our analysis follows closely the actual code behavior.

cs.DS↗

arXiv2015

The Lock-free $k$-LSM Relaxed Priority Queue

Priority queues are data structures which store keys in an ordered fashion to allow efficient access to the minimal (maximal) key. Priority queues are essential for many applications, e.g., Dijkstra's single-source shortest path algorithm, branch-and-bound algorithms, and prioritized schedulers. Efficient multiprocessor computing requires implementations of basic data structures that can be used concurrently and scale to large numbers of threads and cores. Lock-free data structures promise superior scalability by avoiding blocking synchronization primitives, but the \emph{delete-min} operation is an inherent scalability bottleneck in concurrent priority queues. Recent work has focused on alleviating this obstacle either by batching operations, or by relaxing the requirements to the \emph{delete-min} operation. We present a new, lock-free priority queue that relaxes the \emph{delete-min} operation so that it is allowed to delete \emph{any} of the $ρ+1$ smallest keys, where $ρ$ is a runtime configurable parameter. Additionally, the behavior is identical to a non-relaxed priority queue for items added and removed by the same thread. The priority queue is built from a logarithmic number of sorted arrays in a way similar to log-structured merge-trees. We experimentally compare our priority queue to recent state-of-the-art lock-free priority queues, both with relaxed and non-relaxed semantics, showing high performance and good scalability of our approach.

cs.DS↗

arXiv2014

Efficient Lock-free Binary Search Trees

In this paper we present a novel algorithm for concurrent lock-free internal binary search trees (BST) and implement a Set abstract data type (ADT) based on that. We show that in the presented lock-free BST algorithm the amortized step complexity of each set operation - {\sc Add}, {\sc Remove} and {\sc Contains} - is $O(H(n) + c)$, where, $H(n)$ is the height of BST with $n$ number of nodes and $c$ is the contention during the execution. Our algorithm adapts to contention measures according to read-write load. If the situation is read-heavy, the operations avoid helping pending concurrent {\sc Remove} operations during traversal, and, adapt to interval contention. However, for write-heavy situations we let an operation help pending {\sc Remove}, even though it is not obstructed, and so adapt to tighter point contention. It uses single-word compare-and-swap (\texttt{CAS}) operations. We show that our algorithm has improved disjoint-access-parallelism compared to similar existing algorithms. We prove that the presented algorithm is linearizable. To the best of our knowledge this is the first algorithm for any concurrent tree data structure in which the modify operations are performed with an additive term of contention measure.

cs.DC↗

arXiv2014

Self-stabilizing TDMA Algorithms for Wireless Ad-hoc Networks without External Reference

Time division multiple access (TDMA) is a method for sharing communication media. In wireless communications, TDMA algorithms often divide the radio time into timeslots of uniform size, $ξ$, and then combine them into frames of uniform size, $τ$. We consider TDMA algorithms that allocate at least one timeslot in every frame to every node. Given a maximal node degree, $δ$, and no access to external references for collision detection, time or position, we consider the problem of collision-free self-stabilizing TDMA algorithms that use constant frame size. We demonstrate that this problem has no solution when the frame size is $τ< \max\{2δ,χ_2\}$, where $χ_2$ is the chromatic number for distance-$2$ vertex coloring. As a complement to this lower bound, we focus on proving the existence of collision-free self-stabilizing TDMA algorithms that use constant frame size of $τ$. We consider basic settings (no hardware support for collision detection and no prior clock synchronization), and the collision of concurrent transmissions from transmitters that are at most two hops apart. In the context of self-stabilizing systems that have no external reference, we are the first to study this problem (to the best of our knowledge), and use simulations to show convergence even with computation time uncertainties.

cs.DC↗