Various parameter-efficient fine-tuning (PEFT) methods, including LoRA and its variants, have gained popularity for reducing computational costs. However, there is often an accuracy gap between PEFT approaches and full fine-tuning (FT), and this discrepancy has not yet been systematically explored. In this work, we introduce a method for selecting sparse sub-matrices that aims to minimize the performance gap between PEFT vs. full fine-tuning (FT) while also reducing both fine-tuning computational costs and memory costs. We explored both gradient-based and activation-based parameter selection methods to identify the most significant sub-matrices for downstream tasks, updating only these blocks during fine-tuning. In our experiments, we demonstrated that SMT consistently surpasses other PEFT baselines (e.g., LoRA and DoRA) in fine-tuning popular large language models such as LLaMA across a broad spectrum of tasks, while reducing the GPU memory footprint by 67% compared to FT. We also examine how the performance of LoRA and DoRA tends to plateau and decline as the number of trainable parameters increases, in contrast, our SMT method does not suffer from such issues.

Debugging and monitoring programs are integral to engineering and deploying software. Dynamic analyses monitor applications through source code or IR injection, machine code or bytecode rewriting, virtual machine APIs, or direct hardware support. While these techniques are viable within their respective domains, common tooling across techniques is rare, leading to fragmentation of skills, duplicated efforts, and inconsistent feature support. We address this problem in the WebAssembly ecosystem with Whamm, an instrumentation framework for Wasm that uses engine-level probing and has a bytecode rewriting fallback to promote portability. Whamm solves three problems: 1) tooling fragmentation, 2) prohibitive instrumentation overhead of general-purpose frameworks, and 3) tedium of tailoring low-level high-performance mechanisms. Whamm provides fully- programmable instrumentation with declarative match rules, static and dynamic predication, automatic state reporting, and user library support, achieving high performance through compiler and engine optimizations. The Whamm engine API allows instrumentation to be provided to a Wasm engine as Wasm code, reusing existing engine optimizations and unlocking new ones, most notably intrinsification, to minimize overhead. A key insight of our work is that explicitly requesting program state in match rules, rather than reflection, enables the engine to efficiently bundle arguments and even inline compiled probe logic. Whamm streamlines the tooling effort, as its bytecode-rewriting target can run instrumented programs everywhere, lowering fragmentation and advancing the state of the art for engine support. We evaluate Whamm with case studies of non-trivial monitors and show it is expressive, powerful, and efficient.

A key strength of managed runtimes over hardware is the ability to gain detailed insight into the dynamic execution of programs with instrumentation. Analyses such as code coverage, execution frequency, tracing, and debugging, are all made easier in a virtual setting. As a portable, low-level byte-code, WebAssembly offers inexpensive in-process sandboxing with high performance. Yet to date, Wasm engines have not offered much insight into executing programs, supporting at best bytecode-level stepping and basic source maps, but no instrumentation capabilities. In this paper, we show the first non-intrusive dynamic instrumentation system for WebAssembly in the open-source Wizard Research Engine. Our innovative design offers a flexible, complete hierarchy of instrumentation primitives that support building high-level, complex analyses in terms of low-level, programmable probes. In contrast to emulation or machine code instrumentation, injecting probes at the bytecode level increases expressiveness and vastly simplifies the implementation by reusing the engine's JIT compiler, interpreter, and deoptimization mechanism rather than building new ones. Wizard supports both dynamic instrumentation insertion and removal while providing consistency guarantees, which is key to composing multiple analyses without interference. We detail a fully-featured implementation in a high-performance multi-tier Wasm engine, show novel optimizations specifically designed to minimize instrumentation overhead, and evaluate performance characteristics under load from various analyses. This design is well-suited for production engine adoption as probes can be implemented to have no impact on production performance when not in use.

The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded “prompt templates”, i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, or imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric, by creating and collecting demonstrations. We conduct two case studies, showing that succinct DSPy programs can express and optimize pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, DSPy can automatically produce pipelines that outperform out-of-the-box few-shot prompting as well as expert-created demonstrations for GPT-3.5 and Llama2-13b-chat. On top of that, DSPy programs compiled for relatively small LMs like 770M parameter T5 and Llama2-13b-chat are competitive with many approaches that rely on large and proprietary LMs like GPT-3.5 and on expert-written prompt chains. DSPy is available at https://github.com/stanfordnlp/dspy

Circuit breakers are frequently deployed in microservice applications to improve their reliability. They achieve this by short circuiting RPC invocations issued to overloaded or failing services, thereby relieving pressure on those services and allowing them to recover. In this paper, we systematically examine the state of the art in industrial circuit breakers designs. We first present a taxonomy of existing, open-source circuit breaker designs and implementations based on a systematic mapping study. We then examine the relationship between these circuit breaker designs and application reliability. We make a clear case that incorrect application of circuit breakers to an application can hurt reliability in the process of trying to improve it. To address the deficiencies in the state of the art, we propose two new circuit breaker designs and provide guidance on how to properly structure microservice applications for the best circuit breaker use. Finally, we identify several open challenges in circuit breaker usage and design for future researchers.

Companies today increasingly rely on microservice architectures to deliver service for their large-scale mobile or web applications. However, not all developers working on these applications are distributed systems engineers and therefore do not anticipate partial failure: where one or more of the dependencies of their service might be unavailable once deployed into production. Therefore, it is paramount that these issues be raised early and often, ideally in a testing environment or before the code ships to production.

American Sign Language (ASL) classes are typically held face-to-face to increase interactivity and enhance the learning experience. However, the recent COVID-19 pandemic brought about many changes to course delivery methods, primarily resulting in a move to an online format, which had to occur in a short timeframe. The online format has presented students and teachers with many opportunities and challenges. In this experience report, we reflect on the student and teacher perspectives of learning ASL in an online setting. We use our experience to introduce new online ASL class guidelines, videoconferencing improvements, and suggest where future research is needed.

Automation tools like continuous integration services, code coverage reporters, style checkers, dependency managers, etc. are all known to provide significant improvements in developer productivity and software quality. Some of these tools are widespread, others are not. How do these automation "best practices" spread? And how might we facilitate the diffusion process for those that have seen slower adoption? In this paper, we rely on a recent innovation in transparency on code hosting platforms like GitHub---the use of repository badges---to track how automation tools spread in open-source ecosystems through different social and technical mechanisms over time. Using a large longitudinal data set, multivariate network science techniques, and survival analysis, we study which socio-technical factors can best explain the observed diffusion process of a number of popular automation tools. Our results show that factors such as social exposure, competition, and observability affect the adoption of tools significantly, and they provide a roadmap for software engineers and researchers seeking to propagate best practices and tools.

The Scala programming language offers two distinctive language features implicit parameters and implicit conversions, often referred together as implicits. Announced without fanfare in 2004, implicits have quickly grown to become a widely and pervasively used feature of the language. They provide a way to reduce the boilerplate code in Scala programs. They are also used to implement certain language features without having to modify the compiler. We report on a large-scale study of the use of implicits in the wild. For this, we analyzed 7,280 Scala projects hosted on GitHub, spanning over 8.1M call sites involving implicits and 370.7K implicit declarations across 18.7M lines of Scala code.

Abstract The most successful systems for “big data” processing have all adopted functional APIs. We present a new programming model, we call function passing , designed to provide a more principled substrate, or middleware, upon which to build data-centric distributed systems like Spark. A key idea is to build up a persistent functional data structure representing transformations on distributed immutable data by passing well-typed serializable functions over the wire and applying them to this distributed data. Thus, the function passing model can be thought of as a persistent functional data structure that is distributed , where transformations performed on distributed data are stored in its nodes rather than the distributed data itself. One advantage of this model is that failure recovery is simplified by design – data can be recovered by replaying function applications atop immutable data loaded from stable storage. Deferred evaluation is also central to our model; by incorporating deferred evaluation into our design only at the point of initiating network communication, the function passing model remains easy to reason about while remaining efficient in time and memory. Moreover, we provide a complete formalization of the programming model in order to study the foundations of lineage-based distributed computation. In particular, we develop a theory of safe, mobile lineages based on a subject reduction theorem for a typed core language. Furthermore, we formalize a progress theorem that guarantees the finite materialization of remote, lineage-based data. Thus, the formal model may serve as a basis for further developments of the theory of data-centric distributed programming, including aspects such as fault tolerance. We provide an open-source implementation of our model in and for the Scala programming language, along with a case study of several example frameworks and end-user programs written atop this model.

Understanding a program entails understanding its context; dependencies, configurations and even implementations are all forms of contexts. Modern programming languages and theorem provers offer an array of constructs to define contexts, implicitly. Scala offers implicit parameters which are used pervasively, but which cannot be abstracted over. This paper describes a generalization of implicit parameters to implicit function types , a powerful way to abstract over the context in which some piece of code is run. We provide a formalization based on bidirectional type-checking that closely follows the semantics implemented by the Scala compiler. To demonstrate their range of abstraction capabilities, we present several applications that make use of implicit function types. We show how to encode the builder pattern, tagless interpreters, reader and free monads and we assess the performance of the monadic structures presented.

As more applications migrate to the cloud, and as "big data" edges into even more production environments, the performance and simplicity of exchanging data between compute nodes/devices is increasing in importance. An issue central to distributed programming, yet often under-considered, is serialization or pickling, i.e., persisting runtime objects by converting them into a binary or text representation. Pickler combinators are a popular approach from functional programming; their composability alleviates some of the tedium of writing pickling code by hand, but they don't translate well to object-oriented programming due to qualities like open class hierarchies and subtyping polymorphism. Furthermore, both functional pickler combinators and popular, Java-based serialization frameworks tend to be tied to a specific pickle format, leaving programmers with no choice of how their data is persisted. In this paper, we present object-oriented pickler combinators and a framework for generating them at compile-time, called scala/pickling, designed to be the default serialization mechanism of the Scala programming language. The static generation of OO picklers enables significant performance improvements, outperforming Java and Kryo in most of our benchmarks. In addition to high performance and the need for little to no boilerplate, our framework is extensible: using the type class pattern, users can provide both (1) custom, easily interchangeable pickle formats and (2) custom picklers, to override the default behavior of the pickling framework. In benchmarks, we compare scala/pickling with other popular industrial frameworks, and present results on time, memory usage, and size when pickling/unpickling a number of data types used in real-world, large-scale distributed applications and frameworks.

Multimedia networking has been expanding its definition beyond communications of text, image, audio, and video as the Next Generation Internet evolves from social networks to cyber-physical networks. One application related to transportation infrastructure includes health monitoring of existing paved and un-surfaced roads. Embedding remote sensing and spatial information technology into roadways facilitates constant observation of road conditions and automatic detection of frost and thaw fronts during the spring thaw and recovery period. A system is being developed which provides critical quantitative data to eliminate or supplement components of current visual inspection procedures, and thus greatly assists transportation agencies in making spring load restriction (SLR) placement and removal decisions. This paper presents the data visualization module of a Decision Support System for Spring Load Restriction (DSS-SLR). After temperature data are collected by underground sensors and transferred by wireless/wired networks to a central database, a user can view the spatial and temporal temperature patterns via a Web browser. An embedded interpolation routine and a graphical user interface (GUI) enable the user to view a cross section showing frost and thaw depths over time. Our work pioneers this new frontier of multimedia networking, which will have significant applications with regard to health monitoring of existing paved roadway systems, as well as un-surfaced road evaluation and maintenance.