In this paper, we introduce a software-defined framework that enables the parallel utilization of all the programmable processing resources available in heterogeneous system-on-chip (SoC) including FPGA-based hardware accelerators and programmable CPUs. Two platforms with different architectures are considered, and a single C/C++ source code is used in both of them for the CPU and FPGA resources. Instead of simply using the hardware accelerator to offload a task from the CPU, we propose a scheduler that dynamically distributes the tasks among all the resources to fully exploit all computing devices while minimizing load unbalance. The multi-architecture study compares an ARMV7 and ARMV8 implementation with different number and type of CPU cores and also different FPGA micro-architecture and size. We measure that both platforms benefit from having the CPU cores assist FPGA execution at the same level of energy requirements.
Multicore architectures dominate today's processor market. Even though the number of cores and threads are pretty high and continues to grow, inherently serial algorithms do not benefit from the abundance of cores and threads. In this paper, we propose Runahead Computing, a technique which uses idle threads in a multi-threaded architecture for accelerating the execution time of serial algorithms. Through detailed evaluations targeting both CPU and GPU platforms and a specific serial algorithm, our approach reduces the execution latency up to 9x in our experiments.
Parallelizing Tool Execution and LLM Generation for Low-Latency Agent Serving
Shanghai Jiao Tong University（上海交通大学）；Microsoft Research（微软研究院）；Stevens Institute of Technology（Stevens 工程学院）；Google（谷歌）；Hong Kong University of Science and Technology（香港科学与技术大学）
Computational protein structure determination involves optimization in a problem space much too large to exhaustively search. Existing approaches include optimization algorithms such as gradient descent and simulated annealing, but these typically only find local minima. One novel approach implemented in REDcRAFT is to instead of folding a protein all at the same time, fold it residue by residue. This simulates a protein folding as each residue exits from the generating ribosome. While REDcRAFT exponentially reduces the problem space so it can be explored in polynomial time, it is still extrem…
Resource selection and task placement for distributed execution poses conceptual and implementation difficulties. Although resource selection and task placement are at the core of many tools and workflow systems, the methods are ad hoc rather than being based on models. Consequently, partial and non-interoperable implementations proliferate. We address both the conceptual and implementation difficulties by experimentally characterizing diverse modalities of resource selection and task placement. We compare the architectures and capabilities of two systems: the AIMES middleware and Swift workfl…
Infrastructure as a service clouds hide the complexity of maintaining the physical infrastructure with a slight disadvantage: they also hide their internal working details. Should users need knowledge about these details e.g., to increase the reliability or performance of their applications, they would need solutions to detect behavioural changes in the underlying system. Existing runtime solutions for such purposes offer limited capabilities as they are mostly restricted to revealing weekly or yearly behavioural periodicity in the infrastructure. This article proposes a technique for predicti…
Technische Universität Berlin（柏林工业大学）；Humboldt-Universität zu Berlin（柏林洪堡大学）
With the growing amount of data, data processing workloads and the management of their resource usage becomes increasingly important. Since managing a dedicated infrastructure is in many situations infeasible or uneconomical, users progressively execute their respective workloads in the cloud. As the configuration of workloads and resources is often challenging, various methods have been proposed that either quickly profile towards a good configuration or determine one based on data from previous runs. Still, performance data to train such methods is often lacking and must be costly collected.…
High performance computing systems have historically been designed to support applications comprised of mostly monolithic, single-job workloads. Pilot systems decouple workload specification, resource selection, and task execution via job placeholders and late-binding. Pilot systems help to satisfy the resource requirements of workloads comprised of multiple tasks. RADICAL-Pilot (RP) is a modular and extensible Python-based pilot system. In this paper we describe RP's design, architecture and implementation, and characterize its performance. RP is capable of spawning more than 100 tasks/second…
Performance optimization is the art of continuous seeking a harmonious mapping between the application domain and hardware. Recent years have witnessed a surge of deep learning (DL) applications in industry. Conventional wisdom for optimizing such workloads mainly focus on compute intensive ops (GEMM, Convolution, etc). Yet we show in this work, that the performance of memory intensive computations is vital to E2E performance in practical DL models. We propose \emph{FusionStitching}, a optimization framework capable of fusing memory intensive \emph{elementwise}, \emph{reduction} and fine grain…
The scalability and efficiency of graph applications are significantly constrained by conventional systems and their supporting programming models. Technology trends like multicore, manycore, and heterogeneous system architectures are introducing further challenges and possibilities for emerging application domains such as graph applications. This paper explores the space of effective parallel execution of ephemeral graphs that are dynamically generated using the Barnes-Hut algorithm to exemplify dynamic workloads. The workloads are expressed using the semantics of an Exascale computing execut…
This paper investigates co-scheduling algorithms for processing a set of parallel applications. Instead of executing each application one by one, using a maximum degree of parallelism for each of them, we aim at scheduling several applications concurrently. We partition the original application set into a series of packs, which are executed one by one. A pack comprises several applications, each of them with an assigned number of processors, with the constraint that the total number of processors assigned within a pack does not exceed the maximum number of available processors. The objective i…
Scientific computing sometimes involves computation on sensitive data. Depending on the data and the execution environment, the HPC (high-performance computing) user or data provider may require confidentiality and/or integrity guarantees. To study the applicability of hardware-based trusted execution environments (TEEs) to enable secure scientific computing, we deeply analyze the performance impact of AMD SEV and Intel SGX for diverse HPC benchmarks including traditional scientific computing, machine learning, graph analytics, and emerging scientific computing workloads. We observe three main…
The last decade has witnessed growth in the computational requirements for training deep neural networks. Current approaches (e.g., data/model parallelism, pipeline parallelism) parallelize training tasks onto multiple devices. However, these approaches always rely on specific deep learning frameworks and requires elaborate manual design, which make it difficult to maintain and share between different type of models. In this paper, we propose Auto-MAP, a framework for exploring distributed execution plans for DNN workloads, which can automatically discovering fast parallelization strategies th…
Current computational systems are heterogeneous by nature, featuring a combination of CPUs and GPUs. As the latter are becoming an established platform for high-performance computing, the focus is shifting towards the seamless programming of these hybrid systems as a whole. The distinct nature of the architectural and execution models in place raises several challenges, as the best hardware configuration is behaviour and workload dependent. In this paper, we address the execution of compound, multi-kernel, OpenCL computations in multi-CPU/multi-GPU environments. We address how these computatio…
Multi-accelerator servers are increasingly being deployed in shared multi-tenant environments (such as in cloud data centers) in order to meet the demands of large-scale compute-intensive workloads. In addition, these accelerators are increasingly being inter-connected in complex topologies and workloads are exhibiting a wider variety of inter-accelerator communication patterns. However, existing allocation policies are ill-suited for these emerging use-cases. Specifically, this work identifies that multi-accelerator workloads are commonly fragmented leading to reduced bandwidth and increased …
Many extreme scale scientific applications have workloads comprised of a large number of individual high-performance tasks. The Pilot abstraction decouples workload specification, resource management, and task execution via job placeholders and late-binding. As such, suitable implementations of the Pilot abstraction can support the collective execution of large number of tasks on supercomputers. We introduce RADICAL-Pilot (RP) as a portable, modular and extensible Python-based Pilot system. We describe RP's design, architecture and implementation. We characterize its performance and show its a…
Operating a distributed data stream processing workload efficiently at scale is hard. The operator of the workload must parallelize and lay out tasks of the workload with resources that match the requirement of target data rate. The challenge is that neither the operator nor the programmer is typically aware of the scaling behavior of the workload as a function of resources. An operator manually searches for a safe operating point that can handle predicted peak load and deploys with ample headroom for absorbing unpredictable spikes. Such empirical, static over-provisioning is wasteful of both …
Datacenter servers are increasingly heterogeneous: from x86 host CPUs, to ARM or RISC-V CPUs in NICs/SSDs, to FPGAs. Previous works have demonstrated that migrating application execution at run-time across heterogeneous-ISA CPUs can yield significant performance and energy gains, with relatively little programmer effort. However, FPGAs have often been overlooked in that context: hardware acceleration using FPGAs involves statically implementing select application functions, which prohibits dynamic and transparent migration. We present Xar-Trek, a new compiler and run-time software framework th…
We propose a dense tensor accelerator called VectorMesh, a scalable, memory-efficient architecture that can support a wide variety of DNN and computer vision workloads. Its building block is a tile execution unit~(TEU), which includes dozens of processing elements~(PEs) and SRAM buffers connected through a butterfly network. A mesh of FIFOs between the TEUs facilitates data exchange between tiles and promote local data to global visibility. Our design performs better according to the roofline model for CNN, GEMM, and spatial matching algorithms compared to state-of-the-art architectures. It ca…
Heterogeneous graph neural networks (HGNNs) deliver powerful capacity in heterogeneous graph representation learning. The execution of HGNNs is usually accelerated by GPUs. Therefore, characterizing and understanding the execution pattern of HGNNs on GPUs is important for both software and hardware optimizations. Unfortunately, there is no detailed characterization effort of HGNN workloads on GPUs. In this paper, we characterize HGNN workloads at inference phase and explore the execution of HGNNs on GPU, to disclose the execution semantic and execution pattern of HGNNs. Given the characterizat…
The advantages of distributing workloads and utilizing multiple distributed resources are now well established. The type and degree of heterogeneity of distributed resources is increasing, and thus determining how to distribute the workloads becomes increasingly difficult, in particular with respect to the selection of suitable resources. We formulate and investigate the resource selection problem in a way that it is agnostic of specific task and resource properties, and which is generalizable to range of metrics. Specifically, we developed a model to describe the requirements of tasks and to …
Many extreme scale scientific applications have workloads comprised of a large number of individual high-performance tasks. The Pilot abstraction decouples workload specification, resource management, and task execution via job placeholders and late-binding. As such, suitable implementations of the Pilot abstraction can support the collective execution of large number of tasks on supercomputers. We introduce RADICAL-Pilot (RP) as a portable, modular and extensible pilot-enabled runtime system. We describe RP's design, architecture and implementation. We characterize its performance and show it…
Motivated by the need to emulate workload execution characteristics on high-performance and distributed heterogeneous resources, we introduce Synapse. Synapse is used as a proxy application (or "representative application") for real workloads, with the advantage that it can be tuned in different ways and dimensions, and also at levels of granularity that are not possible with real applications. Synapse has a platform-independent application profiler, and has the ability to emulate profiled workloads on a variety of resources. Experiments show that the automated profiling performed using Synaps…
Binarized Neural Networks (BNNs) significantly reduce the computation and memory demands with binarized weights and activations compared to full-precision NNs. Executing a layer in a BNN on different devices of a heterogeneous multiprocessor platform consisting of CPU and GPU can affect the inference performance, i.e., accuracy and latency. Usually, a heterogeneous HW platform consisting of a CPU and a GPU is available to execute the BNN workloads. However, to use the heterogeneous HW effectively, it is necessary to find an efficient strategy for BNN workload mapping. In this work, we propose …
The paper aims to find an efficient way for processing large datasets having different types of workload queries with minimal replication. The work first identifies the complexity of queries best suited for the given data processing tool . The paper proposes Query Complexity Aware partitioning technique QCA with a lightweight query identification and partitioning algorithm. Different replication approaches have been studied to cover more use-cases for different application workloads. The technique is demonstrated using a scientific dataset known as Sloan Digital Sky Survey SDSS. The results sh…
Today's clusters often have to divide resources among a diverse set of jobs. These jobs are heterogeneous both in execution time and in their rate of arrival. Execution time heterogeneity has lead to the development of hybrid schedulers that can schedule both short and long jobs to ensure good task placement. However, arrival rate heterogeneity, or burstiness, remains a problem in existing schedulers. These hybrid schedulers manage resources on statically provisioned cluster, which can quickly be overwhelmed by bursts in the number of arriving jobs. In this paper we propose CloudCoaster, a hyb…
Network-on-Chip (NoC) plays a significant role in the performance of a DNN accelerator. The scalability and modular design property of the NoC help in improving the performance of a DNN execution by providing flexibility in running different kinds of workloads. Data movement in a DNN workload is still a challenging task for DNN accelerators and hence a novel approach is required. In this paper, we propose the In-Network Accumulation (INA) method to further accelerate a DNN workload execution on a many-core spatial DNN accelerator for the Weight Stationary (WS) dataflow model. The INA method ex…
Lowering costs by driving high utilization across deep learning workloads is a crucial lever for cloud providers. We present Singularity, Microsoft's globally distributed scheduling service for highly-efficient and reliable execution of deep learning training and inference workloads. At the heart of Singularity is a novel, workload-aware scheduler that can transparently preempt and elastically scale deep learning workloads to drive high utilization without impacting their correctness or performance, across a global fleet of AI accelerators (e.g., GPUs, FPGAs). All jobs in Singularity are preem…
The excessively increased volume of data in modern data management systems demands an improved system performance, frequently provided by data distribution, system scalability and performance optimization techniques. Optimized horizontal data partitioning has a significant influence of distributed data management systems. An optimally partitioned schema found in the early phase of logical database design without loading of real data in the system and its adaptation to changes of business environment are very important for a successful implementation, system scalability and performance improvem…
