Dataframes have become universally popular as a means to represent data in various stages of structure, and manipulate it using a rich set of operators---thereby becoming an essential tool in the data scientists' toolbox. However, dataframe systems, such as pandas, scale poorly---and are non-interactive on moderate to large datasets. We discuss our experiences developing Modin, our first cut at a parallel dataframe system, which already has users across several industries and over 1M downloads. Modin translates pandas functions into a core set of operators that are individually parallelized via columnar, row-wise, or cell-wise decomposition rules that we formalize in this paper. We also introduce metadata independence to allow metadata---such as order and type---to be decoupled from the physical representation and maintained lazily. Using rule-based decomposition and metadata independence, along with careful engineering, Modin is able to support pandas operations across both rows and columns on very large dataframes---unlike Koalas and Dask DataFrames that either break down or are unable to support such operations, while also being much faster than pandas.

The amount of data being generated in genomics is predicted to be between 2 and 40 exabytes per year for the next decade, making genomic analysis the new frontier and the new challenge for precision medicine. This paper explores targeted deployment of hardware accelerators in the cloud to improve the runtime and throughput of immensescale genomic data analyses. In particular, INDEL (INsertion/DELetion) realignment is a critical operation that enables diagnostic testings of cancer through error correction prior to variant calling. It is the slowest part of the somatic (cancer) genomic analysis pipeline, the alignment refinement pipeline, and represents roughly one-third of the execution time of timesensitive diagnostics for acute cancer patients. To accelerate genomic analysis, this paper describes a hardware accelerator for INDEL realignment (IR), and a hardware-software framework leveraging FPGAs-as-a-service in the cloud. We chose to implement genomics analytics on FPGAs because genomic algorithms are still rapidly evolving (e.g. the de facto standard “GATK Best Practices” has had five releases since January of this year). We chose to deploy genomics accelerators in the cloud to reduce capital expenditure and to provide a more quantitative performance and cost analysis. We built and deployed a sea of IR accelerators using our hardware-software accelerator development framework on AWS EC2 F1 instances. We show that our IR accelerator system performed 81× better than multi-threaded genomic analysis software while being 32× more cost efficient.

Classifier evasion consists in finding for a given instance x the “nearest” instance x’ such that the classifier predictions of x and x’ are different. We present two novel algorithms for systematically computing evasions for tree ensembles such as boosted trees and random forests. Our first algorithm uses a Mixed Integer Linear Program solver and finds the optimal evading instance under an expressive set of constraints. Our second algorithm trades off optimality for speed by using symbolic prediction, a novel algorithm for fast finite differences on tree ensembles. On a digit recognition task, we demonstrate that both gradient boosted trees and random forests are extremely susceptible to evasions. Finally, we harden a boosted tree model without loss of predictive accuracy by augmenting the training set of each boosting round with evading instances, a technique we call adversarial boosting.

"Next generation" data acquisition technologies are allowing scientists to collect exponentially more data at a lower cost. These trends are broadly impacting many scientific fields, including genomics, astronomy, and neuroscience. We can attack the problem caused by exponential data growth by applying horizontally scalable techniques from current analytics systems to accelerate scientific processing pipelines.

GPUs have become part of most commodity systems. Nonetheless, they are often underutilized when not executing graphics-intensive or special-purpose numerical computations, which are rare in consumer workloads. Emerging architectures, such as integrated CPU/GPU combinations, may create an opportunity to utilize these otherwise unused cycles for offloading traditional systems tasks. Garbage collection appears to be a particularly promising candidate for offloading, due to the popularity of managed languages on consumer devices.

Machine learning’s ability to rapidly evolve to changing and complex situations has helped it become a fundamental tool for computer security. That adaptability is also a vulnerability: attackers can exploit machine learning systems. We present a taxonomy identifying and analyzing attacks against machine learning systems. We show how these classes influence the costs for the attacker and defender, and we give a formal structure defining their interaction. We use our framework to survey and analyze the literature of attacks against machine learning systems. We also illustrate our taxonomy by showing how it can guide attacks against SpamBayes, a popular statistical spam filter. Finally, we discuss how our taxonomy suggests new lines of defenses.

Statistical machine learning techniques have recently garnered increased popularity as a means to improve network design and security. For intrusion detection, such methods build a model for normal behavior from training data and detect attacks as deviations from that model. This process invites adversaries to manipulate the training data so that the learned model fails to detect subsequent attacks.

In recent work, we proposed D-Trigger, a framework for tracking a global condition over a large network that allows us to detect anomalies while only collecting a very limited amount of data from distributed monitors. In this paper, we expand our previous work by designing a new class of queries (conditions) that can be tracked for anomaly violations. We show how security violations can be detected over a time window of any size. This is important because security operators do not know in advance the window of time in which measurements should be made to detect anomalies. We also present an algorithm that determines how each machine should filter its time series measurements before back-hauling them to a central operations center. Our filters are computed analytically such that upper bounds on false positive and missed detection rates are guaranteed. In our evaluation, we show that botnet detection can be carried out successfully over a distributed set of machines, while simultaneously filtering out 80 to 90% of the measurement data.

There has been growing interest in building large-scale distributed monitoring systems for sensor, enterprise, and ISP networks. Recent work has proposed using principal component analysis (PCA) over global traffic matrix statistics to effectively isolate network-wide anomalies. To allow such a PCA-based anomaly detection scheme to scale, we propose a novel approximation scheme that dramatically reduces the burden on the production network. Our scheme avoids the expensive step of centralizing all the data by performing intelligent filtering at the distributed monitors. This filtering reduces monitoring bandwidth overheads, but can result in the anomaly detector making incorrect decisions based on a perturbed view of the global data set. We employ stochastic matrix perturbation theory to bound such errors. Our algorithm selects the filtering parameters at local monitors such that the errors made by the detector are guaranteed to lie below a user-specified upper bound. Our algorithm thus allows network operators to explicitly balance the tradeoff between detection accuracy and the amount of data communicated over the network. In addition, our approach enables real-time detection because we exploit continuous monitoring at the distributed monitors. Experiments with traffic data from Abilene backbone network demonstrate that our methods yield significant communication benefits while simultaneously achieving high detection accuracy.

Machine learning systems offer unparalled flexibility in dealing with evolving input in a variety of applications, such as intrusion detection systems and spam e-mail filtering. However, machine learning algorithms themselves can be a target of attack by a malicious adversary. This paper provides a framework for answering the question, "Can machine learning be secure?" Novel contributions of this paper include a taxonomy of different types of attacks on machine learning techniques and systems, a variety of defenses against those attacks, a discussion of ideas that are important to security for machine learning, an analytical model giving a lower bound on attacker's work function, and a list of open problems.

We consider the problem of network anomaly detection in large distributed systems. In this setting, Principal Component Analysis (PCA) has been proposed as a method for discover- ing anomalies by continuously tracking the projection of the data onto a residual subspace. This method was shown to work well empirically in highly aggregated networks, that is, those with a limited number of large nodes and at coarse time scales. This approach, how- ever, has scalability limitations. To overcome these limitations, we develop a PCA-based anomaly detector in which adaptive local data (cid:2)lters send to a coordinator just enough data to enable accurate global detection. Our method is based on a stochastic matrix perturba- tion analysis that characterizes the tradeoff between the accuracy of anomaly detection and the amount of data communicated over the network.

We present Tapestry, a peer-to-peer overlay routing infrastructure offering efficient, scalable, location-independent routing of messages directly to nearby copies of an object or service using only localized resources. Tapestry supports a generic decentralized object location and routing applications programming interface using a self-repairing, soft-state-based routing layer. The paper presents the Tapestry architecture, algorithms, and implementation. It explores the behavior of a Tapestry deployment on PlanetLab, a global testbed of approximately 100 machines. Experimental results show that Tapestry exhibits stable behavior and performance as an overlay, despite the instability of the underlying network layers. Several widely distributed applications have been implemented on Tapestry, illustrating its utility as a deployment infrastructure.

Structured peer-to-peer overlays provide a natural infrastructure for resilient routing via efficient fault detection and precomputation of backup paths. These overlays can respond to faults in a few hundred milliseconds by rapidly shifting between alternate routes. In this paper, we present two adaptive mechanisms for structured overlays and illustrate their operation in the context of Tapestry, a fault-resilient overlay from Berkeley. We also describe a transparent, protocol-independent traffic redirection mechanism that tunnels legacy application traffic through overlays. Our measurements of a Tapestry prototype show it to be a highly responsive routing service, effective at circumventing a range of failures while incurring reasonable cost in maintenance bandwidth and additional routing latency.

We present a novel domain analysis methodology that enables network researchers to quickly select the most accurate modeling and analysis method or methods for a given wired or wireless network path and network characteristic of interest e.g., delay, loss, or error process). Our approach includes two classical models, and two data preconditioning models developed in the Tapas project.

Charging a higher price during periods of congestion should more efficiently allocate scarce resources by encouraging users to conserve. We study its benefits by conducting several pricing experiments over two semesters with students in the dormitories using a computer-telephony-service. Users can use the service to make and receive phone calls from their computers or telephones. While we do not charge users real money, we limit each user to a certain number of tokens a week. With this experimental setup, we conducted a different pricing experiment each week to better understand how prices can be used to entice users to talk less, talk at another time, or use a lower quality connection. With our token scheme as a budget constraint, we can use static pricing policies to influence users' behaviors, but cannot use a simple congestion pricing scheme to encourage users to talk less. For example, we can use time-of-day pricing to encourage users to shift 30% of their usages from the peak to the off-peak hours. We can also use call-duration pricing, a higher rate as a call lasts longer, to encourage 3 times as many calls (18% instead of 6%) to terminate after a price increase. However, when using a simple congestion pricing scheme that charges a rate depending on the number of people calling, we find that we cannot get users to terminate their calls earlier. We believe that users do not change their behaviors because they do not know how long the price increases or decreases will last. Thus to make a congestion pricing scheme more effective, we believe that the price changes need to be more permanent to entice users to change their behaviors'.

Techniques for modeling and simulating channel conditions play an essential role in understanding network protocol and application behavior. In [11], we demonstrated that inaccurate modeling using a traditional analytical model yielded significant errors in error control protocol parameters choices. In this paper, we demonstrate that time-varying effects on wireless channels result in wireless traces which exhibit non-stationary behavior over small window sizes. We then present an algorithm that divides traces into stationary components in order to provide analytical channel models that, relative to traditional approaches, more accurately represent characteristics such as burstiness, statistical distribution of errors, and packet loss processes. Our algorithm also generates artificial traces with the same statistical characteristics as actual collected network traces. For validation, we develop a channel model for the circuit-switched data service in GSM and show that it: (1) more closely approximates GSM channel characteristics than a traditional Gilbert model and (2) generates artificial traces that closely match collected traces' statistics. Using these traces in a simulator environment enables future protocol and application testing under different controlled and repeatable conditions.

This paper describes the REAP (Remote Evaluation in Application-level Protocols) toolkit, a reusable solution for enabling evaluation in such protocols. By using this toolkit, server developers can reduce the number of bytes sent/received and round-trip times used by their protocols to fulfill a user task, making them more efficient in high-latency, low-bandwidth networks. This is accomplished by allowing the clients to upload simple mobile procedures to the server to be executed locally. This paper provides an overview of the REAP toolkit design and architecture, and gives experimental results showing significant reduction in the latency of two popularly-used application-level protocols.

The current generation of wireless links pose a significant challenge for sending video streams, as these links have low bit rate and high error rate compared to wired links. Sending high bit rate delay-sensitive traffic over wireless links requires appropriate error resilient video compression algorithms and transport/link layer protocols. We have built a wireless video system using an off-the-shelf error resilient low bit rate video codec with our implementations of UDP Lite and PPP Lite, which are modifications to transport and link layer protocols respectively. The flexible checksumming schemes in these protocols allow error-resilient application policies to be reflected in the networking stack. We conducted simulations and experimental evaluations of this wireless video system using quantitative performance metrics (i.e., throughput, jitter, packet loss, and end-to-end latency).

The demand for streaming multimedia applications is growing at an incr edible rate. In this paper, we propose Bayeux, an efficient application-level multicast system that scales to arbitrarily large receiver groups while tolerating failures in routers and network links. Bayeux also includes specific mechanisms for load-balancing across replicate root nodes and more efficient bandwidth consumption. Our simulation results indicate that Bayeux maintains these properties while keeping transmission overhead low. To achieve these properties, Bayeux leverages the architecture of Tapestry, a fault-tolerant, wide-area overlay routing and location network.

Because of the emergence of heterogeneous access devices and diverse wired and wireless networks, and a substantial lack of support for integrating these networks, we are building a communication network and a service infrastructure that provides integrated telephony and data services across these networks. We generalize the basic call service to support communication between two or more call parties using any number of devices through any media. Call setup has already been addressed by protocols such as the session initiation protocol (SIP), so we focus on a missing component: scalable and fault-tolerant call session maintenance and control after call setup. In this paper, we present a new signaling protocol that offers scalability, high availability, robustness, and flexibility in creating new call processing services. The protocol is compatible with existing call setup protocols, but provides lightweight call session management using a completely decentralized, soft-state group control protocol. We show that this approach simplifies the implementation of the basic call service, including multi-device calling and service handoff between diverse access networks. The design of our protocol follows the principle of separation of control (signaling information) from data. The signaling protocol has been implemented in ICEBERG, an IP-based core network testbed with access to heterogeneous networks.

The widespread deployment of inexpensive communications technology, computational resources in the networking infrastructure, and network-enabled end devices poses an interesting problem for end users: how to locate a particular network service or device out of hundreds of thousands of accessible services and devices.This paper presents the architecture and implementation of a secure Service Discovery Service (SDS).Service providers use the SDS to advertise complex descriptions of available or already running services, while clients use the SDS to compose complex queries for locating these services.Service descriptions and queries use the extensible Markup Language (XML) to encode such factors as cost, performance, location, and device-or service-specific capabilities.The SDS provides a highlyavailable, fault-tolerant, incrementally scalable service for locating services in the wide-area.Security is a core component of the SDS and, where necessary, communications are both encrypted and authenticated.Furthermore, the SDS uses an hybrid access control list and capability system to control access to service information.

We present the results of a performance evaluation of link layer error recovery over wireless links.Our analysis is based upon a case study of the circuit-switched data service implemented in GSM digital cellular networks.We collected a large set of block erasure traces in different radio environments and used a measurement-based approach to derive real-world models of the wireless link.We show that the throughput of the GSM circuit-switched data channel can be improved by up to 25 percent by increasing the (fixed) frame size of the reliable link layer protocol.Our results also suggest that adaptive frame length control could further increase the channel throughput.In general, our case study shows that pure end-to-end error recovery fails to optimize throughput when wireless links form parts of the end-to-end path.In many cases, it leads to decreased end-to-end throughput, an unfair load on a best-effort network, such as the Internet, and a waste of valuable radio resources (e.g., spectrum and transmission power).In fact, we show that link layer error recovery over wireless links is essential for reliable flows to avoid these problems.

It is well-known that TCP performance may degrade over paths that include wireless links, where packet losses are often not related to congestion. We examine this problem in the context of the GSM digital cellular network, where the wireless link is protected by a reliable link layer protocol. We propose the use of multi-layer tracing as a powerful methodology to analyze the complex protocol interactions between the layers. Our measurements show that TCP throughput over GSM is mostly ideal and that spurious timeouts are extremely rare. The multi-layer tracing tool we developed allowed us to identify the primary causes of degraded performance: (1) inefficient interactions with TCP/IP header compression, and (2) excessive queuing caused by overbuffered links. We conclude that link layer solutions alone can solve the problem of “TCP over wireless links”. We further argue that it is imperative to deploy active queue management and explicit congestion notification mechanisms in wide-area wireless networks; which we expect will be the bottleneck in a future Internet.

Rover is a software toolkit that supports the construction of both mobile-transparent and mobile-aware applications. The mobile-transparent approach aims to enable existing applications to run in a mobile environment without alteration. This transparency is achieved by developing proxies for system services that hide the mobile characteristics of the environment from applications. However, to excel, applications operating in the harsh conditions of a mobile environment must often be aware of and actively adapt to those conditions. Using the programming and communication abstractions present in the Rover toolkit, applications obtain increased availability, concurrency, resource allocation efficiency, fault tolerance, consistency, and adaptation. Experimental evaluation of a suite of mobile applications demonstrates that use of the toolkit requires relatively little programming overhead, allows correct operation, substantially increases interactive performance, and dramatically reduces network utilization.

This paper discusses extensions to the Rover toolkit for constructing reliable mobile-aware applications. The extensions improve upon the existing failure model, which addressed client or communication failures and guaranteed reliable message delivery from clients to server, but did not address server failures (e.g., the loss of an incoming message due to server failure) [16]. Due to the unpredictable, intermittent communication connectivity typically found in mobile client environments, it is inappropriate to make clients responsible for guaranteeing request completion at servers. The extensions discussed in this paper provide both system- and language-level support for reliable operation in the form of stable logging of each message received by a server, per-application stable variables, programmer-supplied failure recovery procedures, server process failure detection, and automatic server process restart. The design and implementation of fault-tolerance support is optimized for high...

The Rover toolkit combines relocatable dynamic objects and queued remote procedure calls to provide unique services for "roving" mobile applications.A relocatable dynamic object is an object with a well-defined interface that can be dynamically loaded into a client computer from a server computer (or vice versa) to reduce clientserver communication requirements.Queued remote procedure call is a communication system that permits applications to continue to make non-blocking remote procedure call requests even when a host is disconnected, with requests and responses being exchanged upon network reconnection.The challenges of mobile environments include intermittent connectivity, limited bandwidth, and channeluse optimization.Experimental results from a Rover-based mail reader, calendar program, and two non-blocking versions of World-Wide Web browsers show that Rover's services are a good match to these challenges.The Rover toolkit also offers advantages for workstation applications by providing a uniform distributed object architecture for code shipping, object caching, and asynchronous object invocation. 1