Demand for low-latency and high-bandwidth data transfer between GPUs has driven the development of multi-GPU nodes. Physical constraints on the manufacture and integration of such systems has yielded heterogeneous intra-node interconnects, where not all devices are connected equally. The next generation of supercomputing platforms are expected to feature AMD CPUs and GPUs. This work characterizes the extent to which interconnect heterogeneity is visible through GPU programming APIs on a system with four AMD MI250x GPUs, and provides several insights for users of such systems.
Augmented reality (AR) is one of emerging applications in modern multimedia systems research. Due to intensive time-consuming computations for AR visualization in mobile devices, quality-aware real-time computing under delay constraints is essentially required. Inspired by Lyapunov optimization framework, this paper proposes a time-average quality maximization method for the AR visualization under delay considerations.
The Internet is responsible for accelerating growth in several fields such as digital media, healthcare, the military. Furthermore, the Internet was founded on the principle of allowing clients to communicating with servers. However, serverless computing is one such field that tries to break free from this paradigm. Event-driven compute services allow users to build more agile applications using capacity provisioning and a pay-for-value billing model. This paper provides a formal account of the research contributions in the field of Serverless computing.
This document presents implementations of fundamental convolutional neural network (CNN) layers on the Manticore cluster-based many-core architecture and discusses their characteristics and trade-offs.
We propose a novel proactive composition framework of wireless energy services in a crowdsourced IoT environment. We define a new model for energy services and requests that includes providers' and consumers' mobility patterns and energy usage behavior. The proposed composition approach leverages the mobility and energy usage behavior to generate energy services and requests proactively. Preliminary experimental results demonstrate the effectiveness of generating proactive energy requests and composing proactive services.
MPI accomplishes portable, standardized message-passing between processes by exposing a standard API that hides the implementation of the underlying mechanism for message passing. Until now, checkpointing an MPI program required knowledge of these underlying mechanisms. Through the addition of a proxy, we demonstrate that MPI programs can be checkpointed and restarted regardless of the MPI implementation utilized. Further, proxies may enable MPI programs to be checkpointed on one MPI implementation, and restarted on another.
We discuss the issue of what we call {\em incentive mismatch}, a fundamental problem with public blockchains supported by economic incentives. This is an open problem, but one potential solution is to make application portable. Portability is desirable for applications on private blockchains. Then, we present examples of middleware designs that enable application portability and, in particular, support migration between blockchains.
It is undeniable that most developers today are building distributed applications. However, most of these applications are developed by composing existing systems together through unspecified APIs exposed to the application developer. Systems are not going away: they solve a particular problem and most applications today need to rely on several of these systems working in concert. Given this, we propose a research direction where higher-level languages with well defined semantics target underlying systems infrastructure as a middle-ground.
We present a linearizable, non-blocking $k$-ary search tree ($k$-ST) that supports fast searches and range queries. Our algorithm uses single-word compare-and-swap (CAS) operations, and tolerates any number of crash failures. Performance experiments show that, for workloads containing small range queries, our $k$-ST significantly outperforms other algorithms which support these operations, and rivals the performance of a leading concurrent skip-list, which provides range queries that cannot always be linearized.
In the world of supercomputers, the large number of processors requires to minimize the inefficiencies of parallelization, which appear as a sequential part of the program from the point of view of Amdahl's law. The recently suggested new figure of merit is applied to the recently presented supercomputer, and the timeline of "Top 500" supercomputers is scrutinized using the metric. It is demonstrated, that in addition to the computing performance and power consumption, the new supercomputer is also excellent in the efficiency of parallelization. Based on the suggested merit, a "Moore-law" like observation is derived for the timeline of parallelization efficacy of supercomputers.
Numerous administration tools and techniques require near real time vision of the activity occurring on a distributed filesystem. The changelog facility provided by Lustre to address this need suffers limitations in terms of scalability and flexibility. We have been working on reducing those limitations by enhancing Lustre itself and developing external tools such as Lustre ChangeLog Aggregate and Publish (LCAP) proxy. Beyond the ability to distribute changelog processing, this effort aims at opening new prospectives by making the changelog stream simpler to leverage for various purposes.
Cloud for Gaming refers to the use of cloud computing technologies to build large-scale gaming infrastructures, with the goal of improving scalability and responsiveness, improve the user's experience and enable new business models.
A family of gossiping algorithms depending on a parameter permutation is introduced, formalized, and discussed. Several of its members are analyzed and their asymptotic behaviour is revealed, including a member whose model and performance closely follows the one of hardware pipelined processors. This similarity is exposed. An optimizing algorithm is finally proposed and discussed as a general strategy to increase the performance of the base algorithms.
We show that there is no deterministic local algorithm (constant-time distributed graph algorithm) that finds a $(7-ε)$-approximation of a minimum dominating set on planar graphs, for any positive constant $ε$. In prior work, the best lower bound on the approximation ratio has been $5-ε$; there is also an upper bound of $52$.
This paper describes recent developments with the licas (lightweight Internet-based communication for autonomic services) software package. In particular, it describes how the architecture and functionality have changed from the first version release. The autonomous nature of the system is focused on, which requires independent behaviour and metadata descriptions of each service. The system has now also been ported to the Java mobile environment. Then some open questions or problems will be discussed in the areas of metadata consistency, security and trust. Finally, some solutions to these problems will also be suggested.
We found that a reliability model commonly used to estimate Mean-Time-To-Data-Loss (MTTDL), while suitable for modeling RAID 0 and RAID 5, fails to accurately model systems having a fault-tolerance greater than 1. Therefore, to model the reliability of RAID 6, Triple-Replication, or k-of-n systems requires an alternate technique. In this paper, we explore some alternatives, and evaluate their efficacy by comparing their predictions to simulations. Our main result is a new formula which more accurately models storage system reliability.
The MapReduce framework has been generating a lot of interest in a wide range of areas. It has been widely adopted in industry and has been used to solve a number of non-trivial problems in academia. Putting MapReduce on strong theoretical foundations is crucial in understanding its capabilities. This work links MapReduce to the BSP model of computation, underlining the relevance of BSP to modern parallel algorithm design and defining a subclass of BSP algorithms that can be efficiently implemented in MapReduce.
Joan Feigenbaum, Brighten Godfrey, Aurojit Panda
et al.
Many modern network designs incorporate "failover" paths into routers' forwarding tables. We initiate the theoretical study of the conditions under which such resilient routing tables can guarantee delivery of packets.