Lukas Burkhalter, Anwar Hithnawi, Alexander Viand, Hossein Shafagh, Sylvia Ratnasamy
A growing number of devices and services collect detailed time series data that is stored in the cloud. Protecting the confidentiality of this vast and continuously generated data is an acute need for many applications in this space. At the same time, we must preserve the utility of this data by enabling authorized services to securely and selectively access and run analytics. This paper presents TimeCrypt, a system that provides scalable and real-time analytics over large volumes of encrypted time series data. TimeCrypt allows users to define expressive data access and privacy policies and enforces it cryptographically via encryption. In TimeCrypt, data is encrypted end-to-end, and authorized parties can only decrypt and verify queries within their authorized access scope. Our evaluation of TimeCrypt shows that its memory overhead and performance are competitive and close to operating on data in the clear.
Silvery Fu, Radhika Mittal, Lei Zhang, Sylvia Ratnasamy
Container is becoming the canonical way of deploying compute tasks at the edge. Unfortunately, container startup latency and overhead remain high, limiting responsiveness and of edge deployment. This latency comes mostly from fetching container dependencies including system libraries, tools, configuration files, and data files. To address this, we propose that schedulers in container orchestrators take into account a task's dependencies. Hence, in dependency scheduling, the scheduler tries to place a task at a node that has the maximum number of the task's dependencies stored locally. We implement dependency scheduling within Kubernetes and evaluate it through extensive experiments and measurement-driven simulations. We show that dependency scheduling improves task startup latency by 1.4-2.3x relative to current dependency-agnostic scheduling for typical scenarios.
Network Evolution for DNNs
Michael Alan Chang, Aurojit Panda, Scott Shenker
Deep Neural Networks increasingly power applications like image search, voice recognition, autonomous vehicles, spam detection, datacenter power management, etc. Many of these applications require DNNs to be periodically retrained, thereby improving prediction quality. As a result improving DNN training time has a significant impact on application performance. As a result DNN training is increasingly distributed across machines, and executed on GPUs, ASICs, or other specialized hardware. In this paper we analyze how the network fabric impacts DNN training time in order to determine how the network fabric should change to better accommodate these jobs. We rely on analytical models and trace driven simulation for our analysis and find that changing the network fabric can significantly impact DNN training performance, but unlike traditional data parallel systems the biggest improvements come from improving data distribution mechanisms rather than aggregation mechanisms.
Datacenter Congestion Control: Identifying what is essential and making it practical
Aisha Mushtaq, Radhika Mittal, James Murphy McCauley, Mohammad Alizadeh, Sylvia Ratnasamy, Scott Shenker
Recent years have seen a slew of papers on datacenter congestion control mechanisms. In this work, we ask whether the bulk of this research is needed for the common case where congestion control involves hosts responding to simple congestion signals from the network and the performance goal is reducing some average measure of flow completion time. We raise this question because we find that, out of all the possible variations one could make in congestion control algorithms, the most essential feature is the switch scheduling algorithm. More specifically, we find that congestion control mechanisms that use Shortest-Remaining-Processing-Time (SRPT) achieve superior performance as long as the rate-setting algorithm at the host is reasonable. We further find that while SRPT’s performance is quite robust to host behaviors, the performance of schemes that use scheduling algorithms like FIFO or Fair Queuing depend far more crucially on the rate-setting algorithm, and their performance is typically worse than what can be achieved with SRPT. Given these findings, we then ask whether it is practical to realize SRPT in switches without requiring custom hardware. We observe that approximate and deployable SRPT (ADS) designs exist, which leverage the small number of priority queues supported in almost all commodity switches, and require only software changes in the host and the switches. Our evaluations with one very simple ADS design shows that it can achieve performance close to true SRPT and is significantly better than FIFO. Thus, the answer to our basic question - whether the bulk of recent research on datacenter congestion control algorithms is needed for the common case - is no.
Sangjin Han, Keon Jang, Dongsu Han, Sylvia Ratnasamy
Modern NICs implement various features in hardware, such as protocol offloading, multicore supports, traffic control, and self virtualization. This approach exposes several issues: protocol dependence, limited hardware resources, and incomplete/buggy/non-compliant implementation. Even worse, the slow evolution of hardware NICs due to increasingly overwhelming design complexity cannot keep up in time with the new protocols and rapidly changing network architectures. We introduce the SoftNIC architecture to fill the gap between hardware capabilities and user demands. Our current SoftNIC prototype implements sophisticated NIC features on a few dedicated processor cores, while assuming only streamlined functionalities in hardware. The preliminary evaluation results show that most NIC features can be implemented in software with minimum performance cost, while the flexibility of software provides further potential benefits.
Shoumik Palkar, Chang Lan, Sangjin Han, Keon Jang, Aurojit Panda, Melvin Walls, Christian Maciocco, Sylvia Ratnasamy, Joshua Reich, Luigi Rizzo, Scott Shenker
By moving network appliance functionality from proprietary hardware to software, Network Function Virtualization promises to bring the advantages of cloud computing to network packet processing. However, the evolution of cloud computing (particularly for data analytics) has greatly benefited from application-independent methods for scaling and placement that achieve high efficiency while relieving programmers of these burdens. NFV has no such general management solutions. To this end, we present E2 -- a scalable and application-agnostic scheduling framework for packet processing.
Aurojit Panda, Sangjin Han, Keon Jang, Melvin Walls, Sylvia Ratnasamy, Scott Shenker
The move from hardware middleboxes to software net-work functions, as advocated by NFV, has proven morechallenging than expected. Developing new NFs remains a tedious process, with developers frequently having to re-discover and reapply the same set of optimizations, while current techniques for safely running multiple NFs (using VMs or containers) incur high performance overheads. In this paper we describe NetBricks, a new NFV framework that aims to improve both the building and running of NFs. For building NFs we take inspiration from databases and modern data analytics frameworks (e.g.,Spark andMap Reduce) and build a framework with a small set of customizable network processing elements. To improve execution performance, NetBricks builds on safe languages and runtimes to provide isolation in software, rather than relying on hardware isolation. NetBricks provides memory isolation comparable to VMs, without the associated performance penalties. To provide efficient I/O, we introducea novel technique called zero-copy software isolation.
Ethan J. Jackson, et. al
Quilt aims to be the easiest way to deploy and network containers. Traditional container orchestrators have a procedural API focused narrowly on compute. The network, usually an afterthought, must be managed by a separate system with its own independent API. This leaves operators with a complex task: write a deployment script that configures everything necessary to get their application up and running. Quilt takes a different approach. It relies on a new domain specific language, Stitch, to specify distributed applications, independent of the specific infrastructure they run on. Given a stitch, Quilt can automatically deploy in a variety of environments: Amazon EC2, Microsoft Azure, and Google Compute Engine, with more coming soon. Furthermore it can do this with no setup -- just point Quilt at a stitch and it will take care of the rest: booting virtual machines, starting containers on those VMs, and ensuring they can communicate. Quilt is currently in alpha and under heavy development. Please try it out! We are eager for feedback!
Recursively Cautious Congestion Control
Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, Scott Shenker
Any congestion control mechanism has two primary goals - to fill the pipe and to do no harm to other flows in the network. These two goals conflict with eachother – the former requires aggressiveness, whereas the latter requires caution. Traditional approaches use the same mechanism (the sending rate) to achieve these two conflicting goals. For example, TCP cautiously probes for bandwidth using slow-start, starting with a small initial window and then ramping up, in order to fill the pipe. As a result, it often takes flows several round-trip times to fully utilize the available bandwidth. RC3 simply decouples these two goals by sending additional packets from the flow using several layers of low priority service, to fill the pipe, while TCP runs as usual at higher priority. It can therefore, quickly take advantage of available capacity from the very first RTT to achieve near-optimal throughputs and smaller flow completion times while preserving TCP-friendliness and fairness. In common wide-area scenarios, RC3 results in a 40% reduction in average flow completion times, with strongest improvements – more than 70% reduction in flow completion time – seen in medium to large sized (100KB - 3MB) flows.
CANDID - Classifying Assets in Networks by Determining Importance and Dependencies
Scott Marshall, Sylvia Ratnasamy, and Vern Paxson
CANDID is a passive NetFlow-based network traffic analysis platform targeted at inferring relationships and dependencies among services running on hosts in enterprise networks. These networks present challenges of great scale, complexity, and nonstop dynamism, which hinder the ability for network administrators to maintain insight into the complex relationships that exist in these networks. Consequently, administrators do not always know how best to proceed if a network failure occurs. CANDID strives to empower administrators by illuminating these relationships, such that they will be prepared to remedy complex service failures. The solutions we present take the first steps towards understanding these complex in-network relationships, with a special focus on inferring one class of dependencies and detecting load balanced services. The current focal point of our work is two radically different, yet complementary, strategies for inferring the presence of load balancing for pairs of systems. We leverage a case study using real NetFlow data from the network located at Lawrence Berkeley National Lab to validate our strategies. Promising results indicate this problem space is rich with unanswered research questions and is worthy of further exploration.
Colin Scott, Andreas Wundsam, Barath Raghavan, Aurojit Panda, Zhi Liu, Sam Whitlock, Ahmed El-Hassany, Andrew Or, Jefferson Lai, Eugene Huang, Kyriakos Zarifis, and Scott Shenker
Software bugs are inevitable in software-defined networking control software, and troubleshooting is a tedious, time-consuming task. In this paper we discuss how to improve control software troubleshooting by presenting a technique for automatically identifying a minimal sequence of inputs responsible for triggering a given bug, without making assumptions about the language or instrumentation of the software under test. We apply our technique to five open source SDN control platforms—Floodlight, NOX, POX, Pyretic, ONOS—and illustrate how the minimal causal sequences our system found aided the troubleshooting process.
Kay Outerhout, Patrick Wendall, Matei Zaharia, Ion Stoica
Sparrow is a high throughput, low latency distributed cluster scheduler. Sparrow is designed for applications that require frequent research allocations due to launching very short jobs (e.g., jobs composed of 100ms tasks). To ensure that scheduling does not become a bottleneck, Sparrow distributes scheduling over several loosely coordianted machines. Each scheduler uses a constant-time scheduling algorithm based on on-demand feedback acquired by probing slave machines. Sparrow can perform task scheduling in milliseconds, two orders of magnitude faster than existing approaches.
Sangjin Han, Scott Marshall, Byung-Gon Chun, and Sylvia Ratnasamy
BSD Sockets has been a de facto standard API for network programming. While it provides a simple and portable way to perform network I/O, it shows suboptimal performance for “message-oriented” network workloads, where connections are short or messages are small. This problem is exacerbated by its poor scalability on multi-core processors. In this work, we explore the benefits of a clean-slate design of network APIs aimed at achieving both high performance and ease of programming. We present MegaPipe, a new API for efficient, scalable network I/O, and evaluate its efficiency and effectiveness with a proof-of-concept implementation.
Junda Liu, Aurojit Panda, Ankit Singla, P. Brighten Godfrey, Michael Schapira, Scott Shenker
Ensuring basic connectivity in the network is generally handled by the control plane. However control plane convergence times are several orders of magnitude larger than the rate at which switches forward packets, which means that after a failure the network might be disrupted for a while, even though the network is not partitioned. Traditionally, networks have handled this problem by precomputing a set of backup paths, over which traffic can be redirected in the case of link failures. The most widely deployed example of such a mechanism is MPLS FRR, however MPLS fast-reroute can only handle a limited number of link failures. A natural question that arises is whether one could design a static mechanism, capable of dealing with any arbitrary set of link failures? We have shown that a static mechanism cannot handle an arbitrary set of failures. Given this result, one must rely on a dynamic mechanism to guarantee ideal connectivity (i.e. packets are delivered as long as a network is connected, and barring congestion related drops). Previous work in this area, has resulted in algorithms that both require unbounded space in packet headers, and NP-complete computations to provide this guarantee. We propose a new algorithm, Data Driven Connectivity, which guarantees ideal connectivity, uses a single bit in the packet header, and can be carried out at line rate.
Justine Sherry, Shaddi Hasan, Colin Scott, Arvind Krishnamurthy, Sylvia Ratnasamy, Vyas Sekar
Middleboxes – such as firewalls, proxies, and WAN optimizers – have become almost ubiquitous in modern enterprises. In a survey of 57 enterprise network administrators, we found that these devices, while popular, are costly, error prone, and hard to manage. To ease these challenges, we developed APLOMB: a service for outsourcing middleboxes to the cloud entirely. With APLOMB, enterprise clients tunnel all of their Internet traffic to and from a cloud provider; the traffic undergoes middlebox processing at the cloud before being forwarded out to the Internet at large. Our implementation is built from open source components including Vyatta and OpenVPN; we hope to have a publicly-available service soon.
Justine Sherry, Shaddi Hasan, Colin Scott, Arvind Krishnamurthy, Sylvia Ratnasamy, Vyas Sekar
Modern networks deploy middleboxes to support numerous advanced processing capabilities such as firewalling, traffic compression, and caching. Despite the widespread deployment of these features, they are nevertheless invisible to the end hosts using the network. We designed ‘network calls’ (netcalls) to allow end hosts to make function calls to the network processing their traffic, allowing the end hosts to invoke and configure the numerous advanced features provided by the networks their traffic traverses. For example, we built a web server that, upon detecting it is under attack using application-layer knowledge, adds additional filters to the firewalls in its network. A key challenge to netcalls is that we want to allow advanced configuration to end hosts not only in their local network, but in any network their traffic traverses. Thus, the netcalls architecture lies primarily in two components: an intra-domain protocol for end hosts to invoke function calls with their network provider, and an inter-domain protocol, by which providers invoke features in each others networks on behalf of their clients. Source code forthcoming.