AI applications increasingly run on fast-evolving, heterogeneous hardware to maximize performance, but general-purpose libraries lag in supporting these features. Performance-minded programmers often build custom communication stacks that are fast but error-prone and non-portable. This paper introduces MSCCL++, a design methodology for developing high-performance, portable communication kernels. It provides (1) a low-level, performance-preserving primitive interface that exposes minimal hardware abstractions while hiding the complexities of synchronization and consistency, (2) a higher-level DSL for application developers to implement workload-specific communication algorithms, and (3) a library of efficient algorithms implementing the standard collective API, enabling adoption by users with minimal expertise. Compared to state-of-the-art baselines, MSCCL++ achieves geomean speedups of 1.7× (up to 5.4×) for collective communication and 1.2× (up to 1.38×) for AI inference workloads. MSCCL++ is in production of multiple AI services provided by Microsoft Azure, and has also been adopted by RCCL, the GPU collective communication library maintained by AMD. MSCCL++ is open source and available at https://github.com/microsoft/mscclpp. Our two years of experience with MSCCL++ suggests that its abstractions are robust, enabling support for new hardware features, such as multimem, within weeks of development.

Machine Learning (ML) models execute several parallel computations including Generalized Matrix Multiplication, Convolution, Dropout, etc. These computations are commonly executed on Graphics Processing Units (GPUs), by dividing the computation into independent processing blocks, known as tiles. Since the number of tiles are usually higher than the execution units of a GPU, tiles are executed on all execution units in one or more waves. However, the number of tiles is not always a multiple of the number of execution units. Thus, tiles executed in the final wave can under-utilize the GPU. To address this issue, we present cuSync, a framework for synchronizing dependent kernels using a user-defined fine-grained synchronization policy to improve the GPU utilization. cuSync synchronizes tiles instead of kernels, which allows executing independent tiles of dependent kernels concurrently. We also present a compiler to generate diverse fine-grained synchronization policies based on dependencies between kernels. Our experiments found that synchronizing CUDA kernels using cuSync reduces the inference times of four popular ML models: MegatronLM GPT-3 by up to 15%, LLaMA by up to 14%, ResNet-38 by up to 22%, and VGG-19 by up to 16% over several batch sizes.

Increasingly complex and diverse deep neural network (DNN) models necessitate distributing the execution across multiple devices for training and inference tasks, and also require carefully planned schedules for performance. However, existing practices often rely on predefined schedules that may not fully exploit the benefits of emerging diverse model-aware operator placement strategies. Handcrafting high-efficiency schedules can be challenging due to the large and varying schedule space. This paper presents Tessel, an automated system that searches for efficient schedules for distributed DNN training and inference for diverse operator placement strategies. To reduce search costs, Tessel leverages the insight that the most efficient schedules often exhibit repetitive pattern (repetend) across different data inputs. This leads to a two-phase approach: repetend construction and schedule completion. By exploring schedules for various operator placement strategies, Tessel significantly improves both training and inference performance. Experiments with representative DNN models demonstrate that Tessel achieves up to 5.5 x training performance speedup and up to 38 % inference latency reduction.

Increasingly complex and diverse deep neural network (DNN) models necessitate distributing the execution across multiple devices for training and inference tasks, and also require carefully planned schedules for performance. However, existing practices often rely on predefined schedules that may not fully exploit the benefits of emerging diverse model-aware operator placement strategies. Handcrafting high-efficiency schedules can be challenging due to the large and varying schedule space. This paper presents Tessel, an automated system that searches for efficient schedules for distributed DNN training and inference for diverse operator placement strategies. To reduce search costs, Tessel leverages the insight that the most efficient schedules often exhibit repetitive pattern (repetend) across different data inputs. This leads to a two-phase approach: repetend construction and schedule completion. By exploring schedules for various operator placement strategies, Tessel significantly improves both training and inference performance. Experiments with representative DNN models demonstrate that Tessel achieves up to 5.5× training performance speedup and up to 38% inference latency reduction.

Machine learning models with millions or billions of parameters are increasingly trained and served on large multi-GPU systems. As models grow in size and execute on more GPUs, collective communication becomes a bottleneck. Custom collective algorithms optimized for both particular network topologies and application-specific communication patterns can alleviate this bottleneck and help these applications scale. However, implementing correct and efficient custom algorithms is challenging.

We develop a new derivative based theory and algorithm for nonbacktracking regex matching that supports anchors and counting, preserves backtracking semantics, and can be extended with lookarounds. The algorithm has been implemented as a new regex backend in .NET and was extensively tested as part of the formal release process of .NET7. We present a formal proof of the correctness of the algorithm, which we believe to be the first of its kind concerning industrial implementations of regex matchers. The paper describes the complete foundation, the matching algorithm, and key aspects of the implementation involving a regex rewrite system, as well as a comprehensive evaluation over industrial case studies and other regex engines.

Recent trends towards large machine learning models require both training and inference tasks to be distributed. Considering the huge cost of training these models, it is imperative to unlock optimizations in computation and communication to obtain best performance. However, the current logical separation between computation and communication kernels in machine learning frameworks misses optimization opportunities across this barrier. Breaking this abstraction can provide many optimizations to improve the performance of distributed workloads. However, manually applying these optimizations requires modifying the underlying computation and communication libraries for each scenario, which is both time consuming and error-prone.

Many applications today, such as natural language processing, network and code analysis, rely on semantically embedding objects into low-dimensional fixed-length vectors. Such embeddings naturally provide a way to perform useful downstream tasks, such as identifying relations among objects and predicting objects for a given context. Unfortunately, training accurate embeddings is usually computationally intensive and requires processing large amounts of data. This paper presents a distributed training framework for a class of applications that use Skip-gram-like models to generate embeddings. We call this class Any2Vec and it includes Word2Vec (Gensim), and Vertex2Vec (DeepWalk and Node2Vec) among others. We first formulate Any2Vec training algorithm as a graph application. We then adapt the state-of-the-art distributed graph analytics framework, D-Galois, to support dynamic graph generation and re-partitioning, and incorporate novel communication optimizations. We show that on a cluster of 3248-core hosts our framework GraphAny2Vec matches the accuracy of the state-of-the-art shared-memory implementations of Word2Vec and Vertex2Vec, and gives geo-mean speedups of 12 x and 5 x respectively. Furthermore, GraphAny2Vec is on average 2 x faster than DMTK, the state-of-the-art distributed Word2Vec implementation, on 32 hosts while yielding much better accuracy.

Collective communication algorithms are an important component of distributed computation. Indeed, in the case of deep-learning, collective communication is the Amdahl's bottleneck of data-parallel training.

Fully-Homomorphic Encryption (FHE) offers powerful capabilities by enabling secure offloading of both storage and computation, and recent innovations in schemes and implementations have made it all the more attractive. At the same time, FHE is notoriously hard to use with a very constrained programming model, a very unusual performance profile, and many cryptographic constraints. Existing compilers for FHE either target simpler but less efficient FHE schemes or only support specific domains where they can rely on expert-provided high-level runtimes to hide complications.

We propose a solution to the problem of efficient matching regular expressions (regexes) with bounded repetition, such as (ab){1,100}, using deterministic automata. For this, we introduce novel counting-set automata (CsAs) , automata with registers that can hold sets of bounded integers and can be manipulated by a limited portfolio of constant-time operations. We present an algorithm that compiles a large sub-class of regexes to deterministic CsAs. This includes (1) a novel Antimirov-style translation of regexes with counting to counting automata (CAs) , nondeterministic automata with bounded counters, and (2) our main technical contribution, a determinization of CAs that outputs CsAs. The main advantage of this workflow is that the size of the produced CsAs does not depend on the repetition bounds used in the regex (while the size of the DFA is exponential to them). Our experimental results confirm that deterministic CsAs produced from practical regexes with repetition are indeed vastly smaller than the corresponding DFAs. More importantly, our prototype matcher based on CsA simulation handles practical regexes with repetition regardless of sizes of counter bounds. It easily copes with regexes with repetition where state-of-the-art matchers struggle.

Fully Homomorphic Encryption (FHE) refers to a set of encryption schemes that allow computations on encrypted data without requiring a secret key. Recent cryptographic advances have pushed FHE into the realm of practical applications. However, programming these applications remains a huge challenge, as it requires cryptographic domain expertise to ensure correctness, security, and performance.

Symbolic regex matcher is a new open source .NET regular expression matching tool and match generator in the Microsoft Automata framework. It is based on the .NET regex parser in combination with a set based representation of character classes. The main feature of the tool is that the core matching algorithms are based on symbolic derivatives that support extended regular expression operations such as intersection and complement and also support a large set of commonly used features such as bounded loop quantifiers. The particularly useful features of the tool are that it supports full UTF16 encoded strings, the match generation is backtracking free, thread safe, and parallelizes with low overhead in multithreaded applications. We discuss the main design decisions behind the tool, explain the core algorithmic ideas and how the tool works, discuss some practical usage scenarios, and compare it to existing state of the art.

List comprehensions provide a powerful abstraction mechanism for expressing computations over ordered collections of data declaratively without having to use explicit iteration constructs. This paper puts forth effectful comprehensions as an elegant way to describe list comprehensions that incorporate loop-carried state. This is motivated by operations such as compression/decompression and serialization/deserialization that are common in log/data processing pipelines and require loop-carried state when processing an input stream of data.

This article focuses on computing minimal test suites for multithreaded programs. Based on previous work on test case generation for multithreaded programs using unfoldings, this article shows how this unfolding can be used to generate minimal test suites covering all local states of the program. Generating such minimal test suites is shown to be NP-complete in the size of the unfolding. We propose an SMT encoding for this problem and two methods based on heuristics which only approximate the solution, but scale better in practice. Finally, we apply our methods to compute the minimal test suites for several benchmarks.

In multithreaded programs both environment input data and the nondeterministic interleavings of concurrent events can affect the behavior of the program. One approach to systematically explore the nondeterminism caused by input data is dynamic symbolic execution. For testing multithreaded programs we present a new approach that combines dynamic symbolic execution with unfoldings, a method originally developed for Petri nets but also applied to many other models of concurrency. We provide an experimental comparison of our new approach with existing algorithms combining dynamic symbolic execution and partial-order reductions and show that the new algorithm can explore the reachable control states of each thread with a significantly smaller number of test runs. In some cases the reduction to the number of test runs can be even exponential allowing programs with long test executions or hard-to-solve constrains generated by symbolic execution to be tested more efficiently.