Re-evaluating Network Onload vs. Offload for the Many-Core Era.
Abstract not provided.
SEARCH · Search NASA
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Abstract not provided.
The OpenMP language features have been evolving to meet the rapid development in hardware platforms. DOE applications tend to push the bleeding edge of features ratified in the OpenMP specification and tend to expose the rough edges of the features' implementations. The software harness on DOE supercomputers such as Titan and (upcoming) Summit include Cray, Clang, Flang, XL and GCC compilers. It is critical, especially for Summit, that the compilers support OpenMP offloading features. This paper focuses on evaluating support for OpenMP 4.5 target offload directives across compiler implementations on Titan and Summitdev, an early access system, which is one generation removed from Summit's architecture enabling application teams to test the systems' architecture. Our tests not only evaluate the OpenMP implementations but also expose ambiguities in the OpenMP 4.5 specification. We also evaluate compiler implementations using kernels extracted from production DOE applications. This helps in assessing the interaction of different OpenMP directives independent of other application artifacts. We are aware that the implementations are constantly evolving and are advertised as having only partial OpenMP 4.x support. We see this as a synergistic effort to help identify and correct features that are required by DOE applications and prevent deployment delays later on. Going forward, we also plan to interact with standard benchmarking bodies like SPEC/HPG to donate our tests and mini-apps/kernels for potential inclusion in the next release versions of SPEC OMP and SPEC ACCEL benchmark suites.
In the last few years, many scientific applications have been developed for powerful graphics processing units (GPUs) and have achieved remarkable speedups. This success can be partially attributed to high performance host callable GPU library routines that are offloaded to GPUs at runtime. These library routines are based on C/C++-like programming toolkits such as CUDA from NVIDIA and have the same calling signatures as their CPU counterparts. Recently, with the sufficient support of C++ templates from CUDA, the emergence of template libraries have enabled further advancement in code reusability and rapid software development for GPUs. However, Expression Templates (ET), which have been very popular for implementing data parallel scientific software for host CPUs because of their intuitive and mathematics-like syntax, have been underutilized by GPU development libraries. The lack of ET usage is caused by the difficulty of offloading expression templates from hosts to GPUs due to the inability to pass instantiated expressions to GPU kernels as well as the absence of the exact form of the expressions for the templates at the time of coding. This paper presents a general approach that enables automatic offloading of C++ expression templates to CUDA enabled GPUs by using the C++ metaprogramming technique and Just-In-Time (JIT) compilation methodology to generate and compile CUDA kernels for corresponding expression templates followed by executing the kernels with appropriate arguments. This approach allows developers to port applications to run on GPUs with virtually no code modifications. More specifically, this paper uses a large ET based data parallel physics library called QDP++ as an example to illustrate many aspects of the approach to offload expression templates automatically and to demonstrate very good speedups for typical QDP++ applications running on GPUs against running on CPUs using this method of offloading. In addition, this approach of automatic offlo- ding expression templates could be applied to other many-core accelerators that provide C++ programming toolkits with the support of C++ template.
The HPC industry is inexorably moving towards an era of extremely heterogeneous architectures, with more devices configured on any given HPC platform and potentially more kinds of devices, some of them highly specialized. Writing a separate code suitable for each target system for a given HPC application is not practical. The better solution is to use directive-based parallel programming models such as OpenMP. OpenMP provides a number of options for offloading a piece of code to devices like GPUs. To select the best option from such options during compilation, most modern compilers use analytical models to estimate the cost of executing the original code and the different offloading code variants. Building such an analytical model for compilers is a difficult task that necessitates a lot of effort on the part of a compiler engineer. Recently, machine learning techniques have been successfully applied to build cost models for a variety of compiler optimization problems. In this paper, we present COMPOFF, a cost model which uses the multi-layer perceptrons to statically estimates the Cost of OpenMP OFFloading. We used six different transformations on a parallel code of Wilson Dslash Operator to support GPU offloading, and we predicted their cost of execution on different GPUs using COMPOFF during compile time. Our results show that this model can predict offloading costs with a root mean squared error in prediction of less than 0.5 seconds. Our preliminary findings indicate that this work will make it much easier and faster for scientists and compiler developers to port legacy HPC applications that use OpenMP to new heterogeneous computing environment.
Efficient collective operations are a major component of application scalability. Offload of collective operations onto the network interface reduces many of the latencies that are inherent in network communications and, consequently, reduces the time to perform the collective operation. To support offload, it is desirable to expose semantic building blocks that are simple to offload and yet powerful enough to implement a variety of collective algorithms. This paper presents the implementation of barrier and broadcast leveraging triggered operations - a semantic building block for collective offload. Triggered operations are shown to be both semantically powerful and capable of improving performance.
Abstract We present Jay , a software framework for offloading applications in hybrid edge clouds. Jay provides an API, services, and tools that enable mobile application developers to implement, instrument, and evaluate offloading applications using configurable cloud topologies, offloading strategies, and job types. We start by presenting Jay 's job model and the concrete architecture of the framework. We then present the programming API with several examples of customization. Then, we turn to the description of the internal implementation of Jay instances and their components. Finally, we describe the Jay Workbench, a tool that allows the setup, execution, and reproduction of experiments with networks of hosts with different resource capabilities organized with specific topologies. The complete source code for the framework and workbench is provided in a GitHub repository.
Recent reliability issues with one of the fastest supercomputers in the world, Titan at Oak Ridge National Laboratory (ORNL), demonstrated the need for resilience in large-scale heterogeneous computing. OpenMP currently does not address error and failure behavior. This paper takes a first step toward resilience for heterogeneous systems by providing the concepts for resilient OpenMP offload to devices. Using real-world error and failure observations, the paper describes the concepts and terminology for resilient OpenMP target offload, including error and failure classes and resilience strategies. It details the experienced general-purpose computing graphics processing unit (GPGPU) errors and failures in Titan. It further proposes improvements in OpenMP, including a preliminary prototype design, to support resilient offload to devices for efficient handling of errors and failures in heterogeneous high-performance computing (HPC) systems.
This work evaluated the use of OpenMP tasking with target GPU offloading as a potential solution for programming productivity and performance on heterogeneous systems. Also, it is proposed a new OpenMP specification to make the implementation of heterogeneous codes simpler by using OpenMP target task, which integrates both OpenMP tasking and target GPU offloading in a single OpenMP pragma. As a test case, the authors used one of the most popular and widely used Basic Linear Algebra Subprogram Level-3 routines: triangular solver (TRSM). To benefit from the heterogeneity of the current high-performance computing systems, the authors propose a different parallelization of the algorithm by using a nonuniform decomposition of the problem. This work used target GPU offloading inside OpenMP tasks to address the heterogeneity found in the hardware. This new approach can outperform the state-of-the-art algorithms, which use a uniform decomposition of the data, on both the CPU-only and hybrid CPU-GPU systems, reaching speedups of up to one order of magnitude. The performance that this approach achieves is faster than the IBM ESSL math library on CPU and competitive relative to a highly optimized heterogeneous CUDA version. One node of Oak Ridge National Laboratory’s supercomputer, Summit, was used for performance analysis.
Modern High-Performance Computing (HPC) centers are facing a data deluge from emerging scientific applications. Supporting large data entails a significant commitment of the highthroughput center storage system, scratch space. However, the scratch space is typically managed using simple purge policies, without sophisticated end-user data services to balance resource consumption and user serviceability. End-user data services such as offloading are performed using point-to-point transfers that are unable to reconcile center s purge and users delivery deadlines, unable to adapt to changing dynamics in the end-toend data path and are not fault-tolerant. Such inefficiencies can be prohibitive to sustaining high performance. In this paper, we address the above issues by designing a framework for the timely, decentralized offload of application result data. Our framework uses an overlay of user-specified intermediate and landmark sites to orchestrate a decentralized fault-tolerant delivery. We have implemented our techniques within a production job scheduler (PBS) and data transfer tool (BitTorrent). Our evaluation using both a real implementation and supercomputer job log-driven simulations show that: the offloading times can be significantly reduced (90.4% for a 5 GB data transfer); the exposure window can be minimized while also meeting center-user Service Level Agreements.
Summary Computational chemistry is at the forefront of solving urgent societal problems, such as polymer upcycling and carbon capture. The complexity of modeling these processes at appropriate length and time scales is mainly manifested in the number and types of chemical species involved in the reactions and may require models of several thousand atoms and large basis sets to accurately capture the chemical complexity and heterogeneity in the physical and chemical processes. The quantum chemistry package General Atomic and Molecular Electronic Structure System (GAMESS) has a wide array of methods that can efficiently and accurately treat complex chemical systems. In this work, we have used the GAMESS Effective Fragment Molecule Orbital (EFMO) method for electronic structure calculation of a challenging mesoporous silica nanoparticle (MSN) model surrounded by about 4700 water molecules to investigate the strong scaling and GPU offloading on hybrid CPU‐GPU nodes. Experiments were performed on the Perlmutter platform at the National Energy Research Scientific Computing Center. Good strong scaling and load balancing have been observed on up to 88 hybrid nodes for different settings of the execution parameters for the calculation considered here. When GPUs are oversubscribed by offloading work from multiple CPU processes, using the NVIDIA multi‐process service (MPS) has consistently reduced time to solution and energy consumed. Additionally, for some configuration parameter settings, oversubscription with MPS improved performance by up to 5.8% over the case without oversubscription.
Here, the latest OpenMP standard offers automatic device offloading capabilities which facilitate GPU programming. Despite this, there remain many challenges. One of these is the unified memory feature introduced in recent GPUs. GPUs in current and future HPC systems have enhanced support for unified memory space. In such systems, CPU and GPU can access each other's memory transparently, that is, the data movement is managed automatically by the underlying system software and hardware. Memory over subscription is also possible in these systems. However, there is a significant lack of knowledge about how this mechanism will perform, and how programmers should use it. We have modified several benchmarks codes, in the Rodinia benchmark suite, to study the behavior of OpenMP accelerator extensions and have used them to explore the impact of unified memory in an OpenMP context. We moreover modified the open source LLVM compiler to allow OpenMP programs to exploit unified memory. The results of our evaluation reveal that, while the performance of unified memory is comparable with that of normal GPU offloading for benchmarks with little data reuse, it suffers from significant overhead when GPU memory is over subcribed for benchmarks with large amount of data reuse. Based on these results, we provide several guidelines for programmers to achieve better performance with unified memory.
Following the mass adoption of external accelerators for high performance computing, the overall performance of many applications has become increasingly dependent on relatively small accelerated kernels. As static analysis is fundamentally limited by dynamic values and external definitions, standard ahead-of-time compilation is not always sufficient to achieve the best performance. Furthermore, many users looking to port an existing application to run on an external accelerator will not want to fundamentally restructure their programs. These and other problems can be addressed through both link-time optimization (LTO) and just-in-time (JIT) compilation, but until now had sparse and inconsistent support from the compiler. In this work, we present a new compilation method that enables device-side LTO as well as a transparent JIT compilation tool-chain for OpenMP target offloading. Our contributions include an entirely new device linking and embedding scheme to enable LTO as well as a novel JIT engine to efficiently optimize OpenMP offloading regions at run-time. We also introduce a persistent caching system to improve end-to-end runtime using the JIT engine and minimize kernel launching overheads. We measure the performance of our LTO and JIT implementation via several real-world scientific applications. With our optimizations we observe significant improvements through LTO on large applications as well as significant end-to-end execution time improvement using JIT.
The OpenMP specification recently introduced support for unified shared memory, allowing implementation to leverage underlying system software to provide a simpler GPU offloading model where explicit mapping of variables is optional. Support for this feature is becoming more available in different OpenMP implementations on several hardware platforms. A deeper understanding of the different implementation’s execution profile and performance is crucial for applications as they consider the performance portability implications of adopting a unified memory offloading programming style. This work introduces a benchmark tool to characterize unified memory support in several OepnMP compilers and runtimes, with emphasis on identifying discrepancies between different OpenMP implementations as to how they various memory allocation strategies interact with unified shared memory. The benchmark tool is used to characterize OpenMP compilers on three leading High Performance Computing platforms supporting different CPU and device architectures. The benchmark tool is used to assess the impact of enabling unified shared memory on the performance of memory-bound code, highlighting implementation differences that should be accounted for when applications consider performance portability across platforms and compilers.
Ceph is a widely used distributed object store, but its messenger layer imposes substantial CPU overhead on the host. To address this limitation, we propose DoCeph, a DPU-offloaded storage architecture for Ceph that disaggregates the system by offloading the communication-intensive messaging component to the DPU while retaining the storage backend on the host. The DPU efficiently manages communication, using lightweight RPC for metadata operations and DMA for data transfer. Moreover, DoCeph introduces a pipelining technique that overlaps data transmission with buffer preparation, mitigating hardware-imposed transfer size limitations. We implemented DoCeph on a Ceph cluster with NVIDIA BlueField-3 DPUs. Evaluation results indicate that DoCeph cuts host CPU usage by up to 92% while sustaining stable throughput and providing larger performance benefits for object writes over 1 MB.
Performing a variety of numerical computations efficiently and, at the same time, in a portable fashion requires both an overarching design followed by a number of implementation strategies. All of these are exemplified below as we present transitioning the PLASMA numerical library from relying on dependence-driven large tasks to achieving utilization of fine grain tasking and offload to hardware accelerators while keeping its core dependence sets: OpenMP source code pragmas and runtime for most system-level functionality and basic low-level numerical kernels provided directly by hardware vendors or open source projects with vendor contributions. We also present new algorithmic methods and their efficient parallel implementations including fine grained tasking for eigen-spectrum slicing and offload for mixed-precision eigenvalue refinement. We provide performance, scaling, and numerical results showing sizable gains over the available solutions from either the open source and vendor-provided packages.
This paper introduces the newly developed Infini-Band (IB) Management Queue capability, used by the Host Channel Adapter (HCA) to manage network task data flow dependancies, and progress the communications associated with such flows. These tasks include sends, receives, and the newly supported wait task, and are scheduled by the HCA based on a data dependency description provided by the user. This functionality is supported by the ConnectX-2 HCA, and provides the means for delegating collective communication management and progress to the HCA, also known as collective communication offload. This provides a means for overlapping collective communications managed by the HCA and computation on the Central Processing Unit (CPU), thus making it possible to reduce the impact of system noise on parallel applications using collective operations. This paper further describes how this new capability can be used to implement scalable Message Passing Interface (MPI) collective operations, describing the high level details of how this new capability is used to implement the MPI Barrier collective operation, focusing on the latency sensitive performance aspects of this new capability. This paper concludes with small scale benchmark experiments comparing implementations of the barrier collective operation, using the new network offload capabilities, with established point-to-point based implementations of these same algorithms, which manage the data flow using the central processing unit. These early results demonstrate the promise this new capability provides to improve the scalability of high performance applications using collective communications. The latency of the HCA based implementation of the barrier is similar to that of the best performing point-to-point based implementation managed by the central processing unit, starting to outperform these as the number of processes involved in the collective operation increases.
Fourier pseudo-spectral methods for nonlinear partial differential equations are of wide interest in many areas of advanced computational science, including direct numerical simulation of three-dimensional (3-D) turbulence governed by the Navier-Stokes equations in fluid dynamics. This paper presents a new capability for simulating turbulence at a new record resolution up to 35 trillion grid points, on the world's first exascale computer, Frontier, comprising AMD MI250x GPUs with HPE's Slingshot interconnect and operated by the US Department of Energy's Oak Ridge Leadership Computing Facility (OLCF). Key programming strategies designed to take maximum advantage of the machine architecture involve performing almost all computations on the GPU which has the same memory capacity as the CPU, performing all-to-all communication among sets of parallel processes directly on the GPU, and targeting GPUs efficiently using OpenMP offloading for intensive number-crunching including 1-D Fast Fourier Transforms (FFT) performed using AMD ROCm library calls. With 99% of computing power on Frontier being on the GPU, leaving the CPU idle leads to a net performance gain via avoiding the overhead of data movement between host and device except when needed for some I/O purposes. Memory footprint including the size of communication buffers for MPI_ALLTOALL is managed carefully to maximize the largest problem size possible for a given node count. Detailed performance data including separate contributions from different categories of operations to the elapsed wall time per step are reported for five grid resolutions, from 2048 3 on a single node to 32768 3 on 4096 or 8192 nodes out of 9408 on the system. Both 1D and 2D domain decompositions which divide a 3D periodic domain into slabs and pencils respectively are implemented. The present code suite (labeled by the acronym GESTS, GPUs for Extreme Scale Turbulence Simulations) achieves a figure of merit (in grid points per second) exceeding goals set in the Center for Accelerated Application Readiness (CAAR) program for Frontier. The performance attained is highly favorable in both weak scaling and strong scaling, with notable departures only for 2048 3 where communication is entirely intra-node, and for 32768 3 , where a challenge due to small message sizes does arise. Communication performance is addressed further using a lightweight test code that performs all-to-all communication in a manner matching the full turbulence simulation code. Performance at large problem sizes is affected by both small message size due to high node counts as well as dragonfly network topology features on the machine, but is consistent with official expectations of sustained performance on Frontier. Overall, although not perfect, the scalability achieved at the extreme problem size of 32768 3 (and up to 8192 nodes — which corresponds to hardware rated at just under 1 exaflop/sec of theoretical peak computational performance) is arguably better than the scalability observed using prior state-of-the-art algorithms on Frontier's predecessor machine (Summit) at OLCF. New science results for the study of intermittency in turbulence enabled by this code and its extensions are to be reported separately in the near future.
Honey bee colonies pollinate about one-third of the world’s food crops, and their rapid decline directly threatens agricultural productivity and ecosystem stability. Understanding how colony-level social defenses influence pesticide fate and the circumstances under which they fail is therefore a crucial question in pollinator biology. We used biological accelerator mass spectrometry (BioAMS), a sensitive radiotracer technique, to track the movement of a model pesticide through a small honey bee colony under laboratory conditions. We tested the hypothesis that social buffering protects honey bees from toxic accumulation and that this protection can be overcome, leading to maternal offloading of the pesticide to developing eggs. Consistent with this hypothesis, our results identified three key mechanisms governing chemical movement within a social insect colony: (1) worker bees initially decrease dietary pesticide levels by 95% through diet filtering and deposition in honeycombs, though this declines to 86% by day 10; (2) queen bees maintain markedly lower pesticide levels than workers but, over time, they accumulate the pesticide in their ovaries and transfer it into developing eggs, revealing a previously undocumented protective mechanism in reproductive individuals; and (3) the presence of a queen bee shifts colony-wide chemical distribution by concentrating worker exposure and increasing pesticide deposition in wax. Our findings show that honey bee colonies function as integrated detoxification networks, in which chemical fate depends on complex social behaviors and caste-specific physiology. When social buffering is overwhelmed, reproductive queens may survive by transferring their chemical burden to their offspring.