arXiv subjects

Matthias Springer

Publications and source records attributed to Matthias Springer.

6 recordsLinked to original sources

arXiv2022

Composable and Modular Code Generation in MLIR: A Structured and Retargetable Approach to Tensor Compiler Construction

Despite significant investment in software infrastructure, machine learning systems, runtimes and compilers do not compose properly. We propose a new design aiming at providing unprecedented degrees of modularity, composability and genericity. This paper discusses a structured approach to the construction of domain-specific code generators for tensor compilers, with the stated goal of improving the productivity of both compiler engineers and end-users. The approach leverages the natural structure of tensor algebra. It has been the main driver for the design of progressive lowering paths in \MLIR. The proposed abstractions and transformations span data structures and control flow with both functional (SSA form) and imperative (side-effecting) semantics. We discuss the implications of this infrastructure on compiler construction and present preliminary experimental results.

cs.PL↗

arXiv2019

Memory-Efficient Object-Oriented Programming on GPUs

Object-oriented programming is often regarded as too inefficient for high-performance computing (HPC), despite the fact that many important HPC problems have an inherent object structure. Our goal is to bring efficient, object-oriented programming to massively parallel SIMD architectures, especially GPUs. In this thesis, we develop various techniques for optimizing object-oriented GPU code. Most notably, we identify the object-oriented Single-Method Multiple-Objects (SMMO) programming model. We first develop an embedded C++ Structure of Arrays (SOA) data layout DSL for SMMO applications. We then design a lock-free, dynamic memory allocator that stores allocations in SOA layout. Finally, we show how to further optimize the memory access of SMMO applications with memory defragmentation.

cs.PL↗

arXiv2018

DynaSOAr: A Parallel Memory Allocator for Object-oriented Programming on GPUs with Efficient Memory Access

Object-oriented programming has long been regarded as too inefficient for SIMD high-performance computing, despite the fact that many important HPC applications have an inherent object structure. On SIMD accelerators, including GPUs, this is mainly due to performance problems with memory allocation and memory access: There are a few libraries that support parallel memory allocation directly on accelerator devices, but all of them suffer from uncoalesed memory accesses. We discovered a broad class of object-oriented programs with many important real-world applications that can be implemented efficiently on massively parallel SIMD accelerators. We call this class Single-Method Multiple-Objects (SMMO), because parallelism is expressed by running a method on all objects of a type. To make fast GPU programming available to average programmers, we developed DynaSOAr, a CUDA framework for SMMO applications. DynaSOAr consists of (1) a fully-parallel, lock-free, dynamic memory allocator, (2) a data layout DSL and (3) an efficient, parallel do-all operation. DynaSOAr achieves performance superior to state-of-the-art GPU memory allocators by controlling both memory allocation and memory access. DynaSOAr improves the usage of allocated memory with a Structure of Arrays data layout and achieves low memory fragmentation through efficient management of free and allocated memory blocks with lock-free, hierarchical bitmaps. Contrary to other allocators, our design is heavily based on atomic operations, trading raw (de)allocation performance for better overall application performance. In our benchmarks, DynaSOAr achieves a speedup of application code of up to 3x over state-of-the-art allocators. Moreover, DynaSOAr manages heap memory more efficiently than other allocators, allowing programmers to run up to 2x larger problem sizes with the same amount of memory.

cs.PL↗

arXiv2018

DynaSOAr: Accelerating Single-Method Multiple-Objects Applications on GPUs

Object-oriented programming (OOP) has long been regarded as too inefficient for SIMD high-performance computing, despite the fact that many important HPC applications have an inherent object structure. We discovered a broad subset of OOP that can be implemented efficiently on massively parallel SIMD accelerators. We call it Single-Method Multiple-Objects (SMMO), because parallelism is expressed by running a method on all objects of a type. To make fast GPU programming available to domain experts who are less experienced in GPU programming, we developed DynaSOAr, a CUDA framework for SMMO applications. DynaSOAr improves the usage of allocated memory with an SOA data layout and achieves low memory fragmentation through efficient management of free and allocated memory blocks with lock-free, hierarchical bitmaps.

cs.PL↗

arXiv2016

Inter-language Collaboration in an Object-oriented Virtual Machine

Multi-language virtual machines have a number of advantages. They allow software developers to use software libraries that were written for different programming languages. Furthermore, language implementors do not have to bother with low-level VM functionality and their implementation can benefit from optimizations in existing virtual machines. MagLev is an implementation of the Ruby programming language on top of the GemStone/S virtual machine for the Smalltalk programming language. In this work, we present how software components written in both languages can interact. We show how MagLev unifies the Smalltalk and the Ruby object model, taking into account Smalltalk meta classes, Ruby singleton classes, and Ruby modules. Besides, we show how we can call Ruby methods from Smalltalk and vice versa. We also present MagLev's concept of bridge methods for implementing Ruby method calling conventions. Finally, we compare our solution to other language implementations and virtual machines.

cs.PL↗

arXiv2016

Fast In-Memory SQL Analytics on Graphs

We study a class of graph analytics SQL queries, which we call relationship queries. Relationship queries are a wide superset of fixed-length graph reachability queries and of tree pattern queries. Intuitively, it discovers target entities that are reachable from source entities specified by the query. It usually also finds aggregated scores, which correspond to the target entities and are calculated by applying aggregation functions on measure attributes, which are found on the target entities, the source entities and the paths from the sources to the targets. We present real-world OLAP scenarios, where efficient relationship queries are needed. However, row stores, column stores and graph databases are unacceptably slow in such OLAP scenarios. We briefly comment on the straightforward extension of relationship queries that allows accessing arbitrary schemas. The GQ-Fast in-memory analytics engine utilizes a bottom-up fully pipelined query execution model running on a novel data organization that combines salient features of column-based organization, indexing and compression. Furthermore, GQ-Fast compiles its query plans into executable C++ source codes. Besides achieving runtime efficiency, GQ-Fast also reduces main memory requirements because, unlike column databases, GQ-Fast selectively allows more dense forms of compression including heavy-weighted compressions, which do not support random access. We used GQ-Fast to accelerate queries for two OLAP dashboards in the biomedical field. It outperforms Postgres by 2-4 orders of magnitude and outperforms MonetDB and Neo4j by 1-3 orders of magnitude when all of them are running on RAM. In addition, it generally saves space due to the appropriate use of compression methods.

cs.DB↗