NAP: Programming Data Planes with Approximate Data Structures

the University of Amsterdam, May 20-31, 1974.This Tract collects the lecture notes of five of the courses given.The sixth course, given by Dr. R. Kowalski on Predicate Logic as a Programming Language in Artificial Intelligence, is scheduled to appear elsewhere.We are very grateful to the Netherlands Organization for the Advancement of Pure Research (Z.W.O.)

To leverage modern hardware platforms to their fullest, more and more database systems embrace compilation of query plans to native code. In the research community, there is an ongoing debate about the best way to architect such query compilers. This is perceived to be a difficult task, requiring techniques fundamentally different from traditional interpreted query execution. In this dissertation, we contribute to this discussion by drawing attention to an old but underappreciated idea known as Futamura projections, which fundamentally link interpreters and compilers. Guided by this idea, we demonstrate that efficient query compilation can actually be very simple, using techniques that are no more difficult than writing a query interpreter in a high-level language. We first develop LB2: a high-level query compiler implemented in this style that is competitive with the best compiled query engines both in sequential and parallel execution on the standard TPC-H benchmark. Query engines process a variety of data types and structures including text, spatial, graphs, etc. Several spatial and graph engines are implemented as extensions to relational query engines to leverage optimized memory, storage, and evaluation. Still, the performance of these extensions is often stymied by the interpretive nature of the underlying data management, generic data structures, and the need to execute domain-specific external libraries. On that basis, compiling spatial and graph queries to native code is a desirable avenue to mitigate existing limitations and improve performance. To support compiling spatial queries, we extend the LB2 main-memory query compiler with spatial predicates, indexing structures, and spatial operators. To support compiling graph queries, we extend LB2 with graph data structures and operators. The spatial extension matches the performance of hand-written code and outperforms relational query engines and map-reduce extensions. Similarly, the graph extension matches, and sometimes outperforms, low-level graph engines.

Security types preserving compilation

Exact-match overlap graphs have been broadly used in the context of DNA assembly and the shortest super string problem where the number of strings n ranges from thousands to billions. The length ℓ of the strings is from 25 to 1000, depending on the DNA sequencing technologies. However, many DNA assemblers using overlap graphs suffer from the need for too much time and space in constructing the graphs. It is nearly impossible for these DNA assemblers to handle the huge amount of data produced by the next-generation sequencing technologies where the number n of strings could be several billions. If the overlap graph is explicitly stored, it would require Ω(n(2)) memory, which could be prohibitive in practice when n is greater than a hundred million. In this article, we propose a novel data structure using which the overlap graph can be compactly stored. This data structure requires only linear time to construct and and linear memory to store. For a given set of input strings (also called reads), we can informally define an exact-match overlap graph as follows. Each read is represented as a node in the graph and there is an edge between two nodes if the corresponding reads overlap sufficiently. A formal description follows. The maximal exact-match overlap of two strings x and y, denoted by ov(max)(x, y), is the longest string which is a suffix of x and a prefix of y. The exact-match overlap graph of n given strings of length ℓ is an edge-weighted graph in which each vertex is associated with a string and there is an edge (x, y) of weight ω=ℓ-|ov(max)(x, y)| if and only if ω ≤ λ, where |ov(max)(x, y)| is the length of ov(max)(x, y) and λ is a given threshold. In this article, we show that the exact-match overlap graphs can be represented by a compact data structure that can be stored using at most (2λ-1)(2⌈logn⌉+⌈logλ⌉)n bits with a guarantee that the basic operation of accessing an edge takes O(log λ) time. We also propose two algorithms for constructing the data structure for the exact-match overlap graph. The first algorithm runs in O(λℓnlogn) worse-case time and requires O(λ) extra memory. The second one runs in O(λℓn) time and requires O(n) extra memory. Our experimental results on a huge amount of simulated data from sequence assembly show that the data structure can be constructed efficiently in time and memory. Our DNA sequence assembler that incorporates the data structure is freely available on the web at http://www.engr.uconn.edu/~htd06001/assembler/leap.zip

Scalable and queryable compressed storage structure for raster data

The objective is to develop a data structure that is capable of handling large data volumes and offers support for querying, analysis and validation. Based on earlier results (i.e. the full decomposition of space, the use of a TEN structure and applying Poincare simplicial homology as mathematical foundation) a simplicial complex-based TEN structure is developed. Applying simplicial homology offers full control over orientation of simplexes and enables one to derive substantial parts of the TEN structure, instead of explicitly store the entire network. The described data structure is developed as a DBMS data structure and the usage of views, function based indexes and 3D R-trees result in a compact topological 3D data structure. Theoretical aspects of this approach are described earlier [12, 17,18]. This paper describes both theory and implementation of the approach.

Network management is challenging. To operate, maintain, and secure a communication network, network operators must grapple with low-level vendor-specific configuration to implement complex high-level network policies. Despite many previous proposals to make networks easier to manage, many solutions to network management problems amount to stop-gap solutions because of the difficulty of changing the underlying infrastructure. The rigidity of the underlying infrastructure presents few possibilities for innovation or improvement, since network devices have generally been closed, proprietary, and vertically integrated. A new paradigm in networking, software defined networking (SDN), advocates separating the data plane and the control plane, making network switches in the data plane simple packet forwarding devices and leaving a logically centralized software program to control the behavior of the entire network. SDN introduces new possibilities for network management and configuration methods. In this article, we identify problems with the current state-of-the-art network configuration and management mechanisms and introduce mechanisms to improve various aspects of network management. We focus on three problems in network management: enabling frequent changes to network conditions and state, providing support for network configuration in a highlevel language, and providing better visibility and control over tasks for performing network diagnosis and troubleshooting. The technologies we describe enable network operators to implement a wide range of network policies in a high-level policy language and easily determine sources of performance problems. In addition to the systems themselves, we describe various prototype deployments in campus and home networks that demonstrate how SDN can improve common network management tasks.

Expanded design procedures for learnable, usable interfaces (panel session)

The language in which evolving programs are expressed can have significant impacts on the problem-solving capabilities of a genetic programming system. These impacts stem both from the absolute computational power of the languages that are used, as elucidated by formal language theory, and from the ease with which various computational structures can be produced by random code generation and by the action of genetic operators. Highly expressive languages can facilitate the evolution of programs for any computable function using, when appropriate, multiple data types, evolved subroutines, evolved control structures, evolved data structures, and evolved modular program and data architectures. In some cases expressive languages can even support the evolution of programs that express methods for their own reproduction and variation (and hence for the evolution of their offspring). This tutorial will begin with a comparative survey of approaches to the evolution of programs in expressive programming languages ranging from machine code to graphical and grammatical representations. Within this context it will then provide a detailed introduction to the Push programming language, which was designed specifically for expressiveness and specifically for use in genetic programming systems. Push programs are syntactically unconstrained but can nonetheless make use of multiple data types and express arbitrary control structures, supporting the evolution of complex, modular programs in a particularly simple and flexible way. The Push language will be described and demonstrated, and ten years of Push-based research, including the production of human-competitive results, will be briefly surveyed. The tutorial will conclude with a discussion of recent enhancements to Push that are intended to support the evolution of complex and robust software systems.

The language in which evolving programs are expressed can have significant impacts on the problem-solving capabilities of a genetic programming system. These impacts stem both from the absolute computational power of the languages that are used, as elucidated by formal language theory, and from the ease with which various computational structures can be produced by random code generation and by the action of genetic operators. Highly expressive languages can facilitate the evolution of programs for any computable function using, when appropriate, multiple data types, evolved subroutines, evolved control structures, evolved data structures, and evolved modular program and data architectures. In some cases expressive languages can even support the evolution of programs that express methods for their own reproduction and variation (and hence for the evolution of their offspring). This tutorial will begin with a comparative survey of approaches to the evolution of programs in expressive programming languages ranging from machine code to graphical and grammatical representations. Within this context it will then provide a detailed introduction to the Push programming language, which was designed specifically for expressiveness and specifically for use in genetic programming systems. Push programs are syntactically unconstrained but can nonetheless make use of multiple data types and express arbitrary control structures, supporting the evolution of complex, modular programs in a particularly simple and flexible way. The Push language will be described and ten years of Push-based research, including the production of human-competitive results, will be briefly surveyed. The tutorial will conclude with a discussion of recent enhancements to Push that are intended to support the evolution of complex and robust software systems.

Software-Defined Networking (SDN) switch vendors are interested in extending switch data planes to support new and continuously evolving network protocols (e.g., NVGRE, VXLAN). Numerous commercial programmable data plane devices already enable a programmer to specify various aspects of the data plane including packet parsing, actions, and the layout of packet processing on the hardware device itself. Unlike OpenFlow-based devices, which only expose a series of fixed match-action table (MAT) abstraction, these specialized devices provide a more flexible abstraction for packet processing. Despite the increased programmability that these devices offer, however, the architecture of the target restricts the features that can be exposed to the programmer. Similarly, existing languages for programming the data planes in such devices (e.g., P4) assume a specific computational model, resembling the architecture of the device for which they are targeted for. Unfortunately, this model leads to similar limitations as in OpenFlow, where the high-level specification is coupled to the underlying device model.In this paper, we introduce NetASM, an intermediate representation for programmable data planes. NetASM is a device-independent language that is expressive enough to act as the target language for compilers for high-level languages, yet low-level enough to be efficiently assembled on various device architectures. It enables conventional compiler optimization techniques to significantly improve the performance and resource utilization of custom packet-processing pipelines on a variety of targets.

We look at program synthesis where the aim is to automatically synthesize a controller that operates on data structures and from which a concrete program can be easily derived. We do not aim at a fully-automatic process or tool that produces a program meeting a given specification of the program’s behaviour. Rather, we aim at the design of a clear and well-founded approach for supporting programmers at the design and implementation phases. Concretely, we first show that a program synthesis task can be modeled as a generalized planning problem. This is done at an abstraction level where the involved data structures are seen as black-boxes that can be interfaced with actions and observations, the first corresponding to the operations and the second to the queries provided by the data structure. The abstraction level is high enough to capture intuitive and common assumptions as well as general and simple strategies used by programmers, and yet it contains sufficient structure to support the automated generation of concrete solutions (in the form of controllers). From such controllers and the use of standard data structures, an actual program in a general language like C++ or Python can be easily obtained. Then, we discuss how the resulting generalized planning problem can be reduced to an LTL synthesis problem, thus making available any LTL synthesis engine for obtaining the controllers. We illustrate the effectiveness of the approach on a series of examples.

High level data structures are a cornerstone of modern programming and at the same time stand in the way of compiler optimizations. In order to reason about user- or library-defined data structures compilers need to be extensible. Common mechanisms to extend compilers fall into two categories. Frontend macros, staging or partial evaluation systems can be used to programmatically remove abstraction and specialize programs before they enter the compiler. Alternatively, some compilers allow extending the internal workings by adding new transformation passes at different points in the compile chain or adding new intermediate representation (IR) types. None of these mechanisms alone is sufficient to handle the challenges posed by high level data structures. This paper shows a novel way to combine them to yield benefits that are greater than the sum of the parts.Instead of using staging merely as a front end, we implement internal compiler passes using staging as well. These internal passes delegate back to program execution to construct the transformed IR. Staging is known to simplify program generation, and in the same way it can simplify program transformation. Defining a transformation as a staged IR interpreter is simpler than implementing a low-level IR to IR transformer. With custom IR nodes, many optimizations that are expressed as rewritings from IR nodes to staged program fragments can be combined into a single pass, mitigating phase ordering problems. Speculative rewriting can preserve optimistic assumptions around loops.We demonstrate several powerful program optimizations using this architecture that are particularly geared towards data structures: a novel loop fusion and deforestation algorithm, array of struct to struct of array conversion, object flattening and code generation for heterogeneous parallel devices. We validate our approach using several non trivial case studies that exhibit order of magnitude speedups in experiments.

Software-Defined Networking (SDN) promises the vision of more flexible and manageable networks, but requires certain level of programmability in the data plane. Such a flexible, programmable data plane implementation is OpenFlow (OF) which these days is seen as primary model of SDN data plane. In this paper we focus on the limitations of OF in packet switching performance. We share some measurement results we collected using an OF 1.3 prototype based on Intel's Data Plane Development Kit (DPDK) and we also describe some optimization ideas. While OF 1.0 can be implemented on high-speed Ethernet switch hardware it has certain disadvantages in the area of flexibility. On the other hand OF 1.3 offers good-enough flexibility, but the poor performance of OF 1.3 implementations seems to represent a roadblock to SDN adoption. In this paper we argue that contrast to the common view, the overhead of flexibility is relatively low. We also argue that the apparent difference between a programmable data plane and a state of the art layered data plane is not primarily due to flexibility itself, but because the lack of optimization in case of flexible implementations.

GPUs and other accelerators are popular devices for accelerating compute-intensive, parallelizable applications. However, programming these devices is a difficult task. Writing efficient device code is challenging, and is typically done in a low-level programming language. High-level languages are rarely supported, or do not integrate with the rest of the high-level language ecosystem. To overcome this, we propose compiler infrastructure to efficiently add support for new hardware or environments to an existing programming language. We evaluate our approach by adding support for NVIDIA GPUs to the Julia programming language. By integrating with the existing compiler, we significantly lower the cost to implement and maintain the new compiler, and facilitate reuse of existing application code. Moreover, use of the high-level Julia programming language enables new and dynamic approaches for GPU programming. This greatly improves programmer productivity, while maintaining application performance similar to that of the official NVIDIA CUDA toolkit.