The emergence of network function virtualization has enabled network function chaining as a flexible approach for building complex network services. However, the high degree of flexibility envisioned for orchestrating network function chains introduces several challenges to support dynamism in workloads and the environment necessary for their realization. Existing works mostly consider supporting dynamism by re-adjusting provisioning of network function instances, incurring reaction times that are prohibitively high in practice. Existing solutions to dynamic packet scheduling rely on centralized schedulers and a priori knowledge of traffic characteristics, and cannot handle changes in the environment like link failures.We fill this gap by presenting FUMES, a reinforcement learning based distributed agent design for the runtime scheduling problem of assigning packets undergoing treatment by network function chains to network function instances. Our design consists of multiple distributed agents that cooperatively work on the scheduling problem. A key design choice enables agents, once trained, to be applicable for unknown chains and traffic patterns including branching, and different environments including link failures. The paper presents the system design and shows its suitability for realistic deployments. We empirically compare FUMES with state-of-the-art runtime scheduling solutions showing improved scheduling decisions at lower server capacity.

The recent trend towards more programmable switching hardware in data centers opens up new possibilities for distributed applications to leverage in-network computing (INC). Literature so far has largely focused on individual application scenarios of INC, leaving aside the problem of coordinating usage of potentially scarce and heterogeneous switch resources among multiple INC scenarios, applications, and users. Alas, the traditional model of resource pools of isolated compute containers does not fit an INC-enabled data center. This paper describes HIRE, a holistic INC-aware resource manager which allows for server-local and INC resources to be coordinated in unison. HIRE introduces a novel flexible resource (meta-)model to address heterogeneity and resource interchangeability, and includes two approaches for INC scheduling: (a) retrofitting existing schedulers; (b) designing a new one. For (a), HIRE presents a retrofitting API and demonstrates it with four state-of-the-art schedulers. For (b), HIRE proposes a flow-based scheduler, cast as a min-cost max-flow problem, where a unified cost model is used to integrate the different costs. Experiments with a workload trace of a 4000 machine cluster show that HIRE makes better use of INC resources by serving 8–30% more INC requests, while simultaneously reducing network detours by 20% and reducing tail placement latency by 50%.

The recent trend towards more programmable switching hardware in data centers opens up new possibilities for distributed applications to leverage in-network computing (INC). Literature so far has largely focused on individual application scenarios of INC, leaving aside the problem of coordinating usage of potentially scarce and heterogeneous switch resources among multiple INC scenarios, applications, and users. The traditional model of resource pools of isolated compute containers does not fit an INC-enabled data center.

Aggregation is common in data analytics and crucial to distilling information from large datasets, but current data analytics frameworks do not fully exploit the potential for optimization in such phases. The lack of optimization is particularly notable in current “online” approaches that store data in main memory across nodes, shifting the bottleneck away from disk I/O toward network and compute resources, thus increasing the relative performance impact of distributed aggregation phases. We present ROME, an aggregation system for use within data analytics frameworks or in isolation. ROME uses a set of novel heuristics based primarily on basic knowledge of aggregation functions combined with deployment constraints to efficiently aggregate results from computations performed on individual data subsets across nodes (e.g., merging sorted lists resulting from top- k ). The user can either provide minimal information that allows our heuristics to be applied directly, or ROME can autodetect the relevant information at little cost. We integrated ROME as a subsystem into the Spark and Flink data analytics frameworks. We use real-world data to experimentally demonstrate speedups up to 3× over single-level aggregation overlays, up to 21% over other multi-level overlays, and 50% for iterative algorithms like gradient descent at 100 iterations.

Network function virtualization has introduced a high degree of flexibility for orchestrating service functions. The provisioning of chains of service functions requires making decisions on both (1) placement of service functions and (2) scheduling of traffic through them. The placement problem (1) can be tackled during the planning phase, by exploiting coarse-grained traffic information, and has been studied extensively. However, runtime traffic scheduling (2) for optimizing system utilization and service quality, as required for future edge cloud and mobile carrier scenarios, has not been addressed so far.We fill this gap by presenting a queuing-based system model to characterize the runtime traffic scheduling problem for service function chaining. We propose a throughput-optimal scheduling policy, called integer allocation maximum pressure policy (IA-MPP). To ensure practicality in large distributed settings, we propose multi-site cooperative IA-MPP (STEAM), fulfilling runtime requirements while achieving near-optimal performance. We examine our policies in various settings representing real-world scenarios. STEAM closely matches IA-MPP in terms of throughput, and significantly outperforms (possible adaptations of) existing static or coarse-grained dynamic solutions, requiring 30%-60% less server capacity for similar service quality. Our STEAM prototype shows feasibility running on a standard server.

Data center networks lie at the core of distributed data analytics frameworks running in large scale environments. Recent research seek to improve the system performance by optimizing the end-host network usage, e.g., optimally use RDMA [2] or zero copy I/O frameworks [5] for distributed data analytics frameworks. Such approaches allow these systems to leverage the high network-bandwidth at end-hosts, however, keep the network itself untouched which does not solve contention and scalability issues.

In times of cloud services and IoT, network communication takes heavily part of distributed applications.The underlying network protocol processing causes substantial CPU utilization, although recent kernel-bypass frameworks or specialized server network cards reduce the burden for the CPU. Since hardware acceleration is expensive and still not very common, the leverage of available compute power is a logical step, that leads to GPGPU utilization. The new heterogeneous uniform memory access architecture unites CPU and GPU on one die and therefore features a shared memory excluding the overhead of PCI-E communication. In this paper we present SocketCL, a user-space network protocol stack acceleration. In order to utilize the GPU we ported CPU based Ethernet, IPv4 and UDP network protocols to OpenCL 2.0 kernels leveraging the new hUMA architecture. Resulting from minimal copy processes and optimal utilization of a system's compute resources, SocketCL achieves substantial acceleration of individual network protocols.

The world is driven by the Internet and there is no doubt about its importance in our daily life. However, the Internet has rarely been upgraded since its advent, although the ISO OSI model has already provided the required flexibility. With innovations being blocked, the Internet is suffering from a high-degree of ossification (e.g., the slow progress of IPv6 update), leading to suboptimal efficiency for emerging applications as well as enlarged maintenance cost.In this paper, we present VirtualStack, which aims at bringing back the interchangeability of network layers. VirtualStack is based on the idea of protocol virtualization, where the most suitable protocol stack can be dynamically composed and applied on the fly according to the characteristics of both the application and the physical link. Through a comprehensive study, we show that many existing but not widely deployed protocols outperform the omnipresent TCP under various link technologies and network conditions. This provides the necessary insight for dynamic composition of the network protocol stack. We further evaluate VirtualStack under a typical Internet setting with multiple hops under different conditions. The experimental results confirm the benefits as well as the potential of VirtualStack.