Search for: All records

Asynchronous Many-task (AMT) runtime systems have gained increasing acceptance in the HPC community due to the performance improvements offered by fine-grained tasking runtime systems. At the same time, C++ standardization efforts are focused on creating higher-level interfaces able to replace OpenMP or OpenACC in modern C++ codes. These higher level functions have been adopted in standards conforming runtime systems such as HPX, giving users the ability to simply utilize fork-join parallelism in their own codes. Despite innovations in runtime systems and standardization efforts users face enormous challenges porting legacy applications. Not only must users port their own codes, but often users rely on highly optimized libraries such as BLAS and LAPACK which use OpenMP for parallization. Current efforts to create smooth migration paths have struggled with these challenges, especially as the threading systems of AMT libraries often compete with the treading system of OpenMP. To overcome these issues, our team has developed hpxMP, an implementation of the OpenMP standard, which utilizes the underlying AMT system to schedule and manage tasks. This approach leverages the C++ interfaces exposed by HPX and allows users to execute their applications on an AMT system without changing their code. In this work, we compare hpxMP with Clang's OpenMP library with four linear algebra benchmarks of the Blaze C++ library. While hpxMP is often not able to reach the same performance, we demonstrate viability for providing a smooth migration for applications but have to be extended to benefit from a more general task based programming model. 

As the era of high frequency, single core processors have come to a close, the new paradigm of many core processors has come to dominate. In response to these systems, asynchronous multitasking runtime systems have been developed as a promising solution to efficiently utilize these newly available hardware. Asynchronous multitasking runtime systems work by dividing a problem into a large number of fine grained tasks. However, as the number of tasks created increase, the overheads associated with task creation and management cannot be ignored. Task inlining, a method where the parent thread consumes a child thread, enables the runtime system to achieve the balance between parallelism and its overhead. As largely impacted by different processor architectures, the decision of task inlining is dynamic in nature. In this research, we present adaptive techniques for deciding, at runtime, whether a particular task should be inlined or not. We present two policies, a baseline policy that makes inlining decision based on a fixed threshold and an adaptive policy which decides the threshold dynamically at runtime. We also evaluate and justify the performance of these policies on different processor architectures. To the best of our knowledge, this is the first study of the impacts of adaptive policy at runtime for task inlining in an asynchronous multitasking runtime system on different processor architectures. From experimentation, we find that the baseline policy improves the execution time from 7.61% to 54.09%. Furthermore, the adaptive policy improves over the baseline policy by up to 74%. 

Experience shows that on today's high performance systems the utilization of different acceleration cards in conjunction with a high utilization of all other parts of the system is difficult. Future architectures, like exascale clusters, are expected to aggravate this issue as the number of cores are expected to increase and memory hierarchies are expected to become deeper. One big aspect for distributed applications is to guarantee high utilization of all available resources, including local or remote acceleration cards on a cluster while fully using all the available CPU resources and the integration of the GPU work into the overall programming model. For the integration of CUDA code we extended HPX, a general purpose C++ run time system for parallel and distributed applications of any scale, and enabled asynchronous data transfers from and to the GPU device and the asynchronous invocation of CUDA kernels on this data. Both operations are well integrated into the general programming model of HPX which allows to seamlessly overlap any GPU operation with work on the main cores. Any user defined CUDA kernel can be launched on any (local or remote) GPU device available to the distributed application. We present asynchronous implementations for the data transfers and kernel launches for CUDA code as part of a HPX asynchronous execution graph. Using this approach we can combine all remotely and locally available acceleration cards on a cluster to utilize its full performance capabilities. Overhead measurements show, that the integration of the asynchronous operations (data transfer + launches of the kernels) as part of the HPX execution graph imposes no additional computational overhead and significantly eases orchestrating coordinated and concurrent work on the main cores and the used GPU devices. 

Despite advancements in the areas of parallel and distributed computing, the complexity of programming on High Performance Computing (HPC) resources has deterred many domain experts, especially in the areas of machine learning and artificial intelligence (AI), from utilizing performance benefits of such systems. Researchers and scientists favor high-productivity languages to avoid the inconvenience of programming in low-level languages and costs of acquiring the necessary skills required for programming at this level. In recent years, Python, with the support of linear algebra libraries like NumPy, has gained popularity despite facing limitations which prevent this code from distributed runs. Here we present a solution which maintains both high level programming abstractions as well as parallel and distributed efficiency. Phylanx, is an asynchronous array processing toolkit which transforms Python and NumPy operations into code which can be executed in parallel on HPC resources by mapping Python and NumPy functions and variables into a dependency tree executed by HPX, a general purpose, parallel, task-based runtime system written in C++. Phylanx additionally provides introspection and visualization capabilities for debugging and performance analysis. We have tested the foundations of our approach by comparing our implementation of widely used machine learning algorithms to accepted NumPy standards. 

Overheads associated with fine grained communication in task based runtime systems are one of the major bottlenecks that limit the performance of distributed applications. In this research, we provide methodology and metrics for analyzing network overheads using the introspection capabilities of HPX, a task based runtime system. We demonstrate that our metrics show a strong correlation with the overall runtime of our test applications. Our aim is to eventually use these metrics to tune, at runtime, parameters relating to active message coalescing. This method improves on the postmortem analysis techniques that are currently employed to tune network settings in distributed applications.