The Priority-based Flow Control (PFC) protocol is adopted to guarantee zero packet loss in many high-performance data centers. PFC, however, can induce deadlocks and severe cases cause the entire network be blocked. Existing solutions have focused on deadlock avoidance; unfortunately, they are not foolproof. Therefore, detection a necessity. We propose ITSY, novel system that correctly detects resolves entirely plane. It works with any topologies routing algorithms. Unique ITSY use of...

Traffic aggregates in cloud data center networks are by and large buffered transmitted simple physical FIFO queues. Despite the crucial role they play, a well-known problem of queues is that unable to provide precise bandwidth guarantees. This leads range negative impacts spanning application layer, transport link layer.

We consider extending decision support facilities toward large-scale sophisticated networks, upon which multidimensional attributes are associated variety of entities, forming the so-called Multidimensional heterogeneous network. network has become important tool for modelling information meanwhile, OLAP (Online Analytical Processing) proven to beeffective on relational data, however, it is computationally an enormous challenge manage and analyse effective making operations. In this paper,...

We quantify the threat of network adversaries to inducing \emph{network overload} through \emph{routing attacks}, where a subset nodes are hijacked by an adversary. develop routing attacks on for two objectives related overload: \emph{no-loss throughput minimization} and \emph{loss maximization}. The first objective attempts identify attack that minimizes network's is guaranteed survive. polynomial-time algorithm can output optimal in multi-hop networks with global information topology,...

An optical circuit-switched network core has the potential to overcome inherent challenges of a conventional electrical packet-switched today's compute clusters. As circuit switches (OCS) directly handle photon beams without any optical-electrical-optical (O/E/O) conversion and packet processing, OCS-based cores have following desirable properties: a) agnostic data-rate, b) negligible/zero power consumption, c) no need transceivers, d) negligible forwarding latency, e) for frequent upgrade....

Emerging distributed applications, such as microservices, machine learning, big data analysis, consist of both compute and network tasks. DAG-based abstraction primarily targets tasks has no explicit network-level scheduling. In contrast, Coflow collectively schedules flows among but lacks the end-to-end view application DAG. Because dependencies interactions between these two types tasks, it is sub-optimal to only consider one them. We argue that co-scheduling can help applications towards...

We develop link rate control policies to minimize the queueing delay of packets in overloaded networks. show that increasing rates does not guarantee reduction during overload. consider a fluid model facilitates explicit characterization packets, and establish conditions on can average maximum both single-hop multi-stage (switching) These min-delay require maintaining an identical ratio between ingress egress different nodes at same layer network. term follow these rate-proportional...

There has been increasing interest in Vehicular Ad Hoc networks (VANET). The sharing of resources among vehicles is one the most crucial topics VANET when two transfer files to each other for some demands. However, due high speed mobility vehicles, link between frequently disconnected, which makes transmission time unpredictable. This paper analyzes probability that packet can be received, labile instability based on lower-layer VANET. And we use nonhomogeneous Markov model research...

Recent data center applications rely on lossless networks to achieve high network performance. Lossless networks, however, can suffer from in-network deadlocks induced by hop-by-hop flow control protocols like PFC. Once occur, large parts of the could be blocked. Existing solutions mainly a deadlock avoidance strategy; unfortunately, they are not foolproof. Thus, detection is necessary last resort. In this paper, we propose DCFIT, new mechanism performed entirely in plane detect and solve...

We consider the problem of overload balancing in single-hop networks with bounded buffers. show that backpressure policy, which is known to achieve most balanced for unbounded buffers, does not balance formulate buffers by leveraging ordinary differential equations (ODE) model queue dynamics. prove choosing service rates on each transmission link minimizes quadratic sum leads overload. Based this result, we propose a queue-based policy combining maxweight scheduling backpressure, can...

We study the pattern of queueing delay minimization in overloaded networks using link rate control. show that serving packets with maximum rates does not minimize when are on overload. To understand this counter-intuitive observation and identify optimal solution, we characterize packet explicitly through a fluid queue model, which enables us to sufficient necessary condition for policy be delay-optimal single-hop networks. setting proportional arrival minimizes both average among all...

Large-scale distributed training is increasingly becoming communication bound. Many gradient compression algorithms have been proposed to reduce the overhead and improve scalability. However, it has observed that in some cases may even harm performance of training. In this paper, we propose MergeComp, a scheduler optimize scalability communication-efficient It automatically schedules operations without knowledge model architectures or system parameters. We applied MergeComp nine popular...

Distributed applications, such as database queries and distributed training, consist of both compute network tasks. DAG-based abstraction primarily targets tasks has no explicit network-level scheduling. In contrast, Coflow collectively schedules flows among but lacks the end-to-end view application DAG. Because dependencies interactions between these two types tasks, it is sub-optimal to only consider one them. We argue that co-scheduling can help applications towards globally optimal...

Lossless networks are increasingly popular for high-performance applications in data centers and cloud environments. To realize a lossless network Ethernet, the Priority-based Flow Control (PFC) protocol is adopted to guarantee zero packet loss. PFC, however, can induce in-network deadlocks severe cases cause entire be blocked. Existing solutions have focused on deadlock avoidance strategies; unfortunately, they not foolproof. Therefore, detection necessity. In this paper, we propose ITSY,...

An optical circuit-switched network core has the potential to overcome inherent challenges of a conventional electrical packet-switched today's compute clusters. As circuit switches (OCS) directly handle photon beams without any optical-electrical-optical (O/E/O) conversion and packet processing, OCS-based cores have following desirable properties: a) agnostic data-rate, b) negligible/zero power consumption, c) no need transceivers, d) negligible forwarding latency, e) for frequent upgrade....