Efficient Data-Plane Memory Scheduling for In-Network Aggregation
iSING Lab, Hong Kong University of Science and Technology（香港科技大学 iSING 实验室）；Harvard University（哈佛大学）；Intel, Barefoot Switch Division（英特尔 Barefoot 交换机部门）；Peking University（北京大学）
Hao Wang, Yuxuan Qin, ChonLam Lao, Yanfang Le, Wenfei Wu, Kai Chen
As the scale of distributed training grows, communication becomes a bottleneck. To accelerate the communication, recent works introduce In-Network Aggregation (INA), which moves the gradients summation into network middle-boxes, e.g., programmable switches to reduce the traffic volume. However, switch memory is scarce compared to the volume of gradients transmitted in distributed training. Although literature applies methods like pool-based streaming or dynamic sharing to tackle the mismatch, switch memory is still a potential performance bottleneck. Furthermore, we observe the under-utilization of switch memory due to the synchronization requirement for aggregator deallocation in recent works. To improve the switch memory utilization, we propose ESA, an $\underline{E}$fficient Switch Memory $\underline{S}$cheduler for In-Network $\underline{A}$ggregation. At its cores, ESA enforces the preemptive aggregator allocation primitive and introduces priority scheduling at the data-plane, which improves the switch memory utilization and average job completion time (JCT). Experiments show that ESA can improve the average JCT by up to $1.35\times$.
Federated Learning (FL) is a newly emerged decentralized machine learning (ML) framework that combines on-device local training with server-based model synchronization to train a centralized ML model over distributed nodes. In this paper, we propose an asynchronous FL framework with periodic aggregation to eliminate the straggler issue in FL systems. For the proposed model, we investigate several device scheduling and update aggregation policies and compare their performances when the devices have heterogeneous computation capabilities and training data distributions. From the simulation resul…
Huazhong University of Science and Technology（华中科技大学）；Huawei Technologies Co., Ltd.（华为技术有限公司）
Computing workflows in heterogeneous multiprocessor systems are frequently modeled as directed acyclic graphs of tasks and data blocks, which represent computational modules and their dependencies in the form of data produced by a task and used by others. However, for some workflows, such as the task schedule in a digital signal processor may run out of memory by exposing too much parallelism. This paper focuses on the data allocation and task scheduling problem under memory constraints, and concentrates on shared memory platforms. We first propose an integer linear programming model to formul…
OptPipe: Memory- and Scheduling-Optimized Pipeline Parallelism for LLM Training
Shanghai University of Finance and Economics（上海金融学院）；National University of Singapore（新加坡国立大学）；Shanghai Jiao Tong University（上海交通大学）；Stanford University（斯坦福大学）
Loop scheduling techniques aim to achieve load-balanced executions of scientific applications. Dynamic loop self-scheduling (DLS) libraries for distributed-memory systems are typically MPI-based and employ a centralized chunk calculation approach (CCA) to assign variably-sized chunks of loop iterations. We present a distributed chunk calculation approach (DCA) that supports various types of DLS techniques. Using both CCA and DCA, twelve DLS techniques are implemented and evaluated in different CPU slowdown scenarios. The results show that the DLS techniques implemented using DCA outperform the…
Scientific applications often contain large computationally-intensive parallel loops. Loop scheduling techniques aim to achieve load balanced executions of such applications. For distributed-memory systems, existing dynamic loop scheduling (DLS) libraries are typically MPI-based, and employ a master-worker execution model to assign variably-sized chunks of loop iterations. The master-worker execution model may adversely impact performance due to the master-level contention. This work proposes a distributed chunk-calculation approach that does not require the master-worker execution scheme. Mor…
The problem of task scheduling with communication delays is strongly NP-hard. State-space search algorithms such as A* have been shown to be a promising approach to solving small to medium sized instances optimally. A recently proposed state-space model for task scheduling, known as Allocation-Ordering (AO), allows state-space search methods to be applied without the need for previously necessary duplicate avoidance mechanisms, and resulted in significantly improved A* performance. The property of a duplicate-free state space also holds particular promise for memory limited search algorithms, …
Recently, deep learning has been an area of intense research. However, as a kind of computing-intensive task, deep learning highly relies on the scale of GPU memory, which is usually prohibitive and scarce. Although some extensive works have been proposed for dynamic GPU memory management, they are hard to apply to systems with multiple dynamic workloads, such as in-database machine learning systems. In this paper, we demonstrated TENSILE, a method of managing GPU memory in tensor granularity to reduce the GPU memory peak, considering the multiple dynamic workloads. TENSILE tackled the cold-st…
Major chip manufacturers have all introduced multicore microprocessors. Multi-socket systems built from these processors are routinely used for running various server applications. Depending on the application that is run on the system, remote memory accesses can impact overall performance. This paper presents a new operating system (OS) scheduling optimization to reduce the impact of such remote memory accesses. By observing the pattern of local and remote DRAM accesses for every thread in each scheduling quantum and applying different algorithms, we come up with a new schedule of threads for…
Common resource management methods in supercomputing systems usually include hard divisions, capping, and quota allotment. Those methods, despite their 'advantages', have some known serious disadvantages including unoptimized utilization of an expensive facility, and occasionally there is still a need to dynamically reschedule and reallocate the resources. Consequently, those methods involve bad supply-and-demand management rather than a free market playground that will eventually increase system utilization and productivity. In this work, we propose the newly Optimized Memoryless Fair-Share H…
Multi-socket multi-core servers are used for solving some of the important problems in computing. Remote DRAM accesses can impact performance of certain applications running on such servers. This paper presents a new near linear operating system (OS) scheduling algorithm to reduce the impact of such remote DRAM accesses. By keeping track of the number of local and remote DRAM accesses, using performance counters, for every thread and applying this algorithm, I come up with a new schedule of threads for the next quantum. This new schedule reduces remote DRAM accesses and improves overall perfor…
Understanding the performance of data-parallel workloads when resource-constrained has significant practical importance but unfortunately has received only limited attention. This paper identifies, quantifies and demonstrates memory elasticity, an intrinsic property of data-parallel tasks. Memory elasticity allows tasks to run with significantly less memory that they would ideally want while only paying a moderate performance penalty. For example, we find that given as little as 10% of ideal memory, PageRank and NutchIndexing Hadoop reducers become only 1.2x/1.75x and 1.08x slower. We show tha…
GPUs in High-Performance Computing systems remain under-utilised due to the unavailability of schedulers that can safely schedule multiple applications to share the same GPU. The research reported in this paper is motivated to improve the utilisation of GPUs by proposing a framework, we refer to as schedGPU, to facilitate intra-node GPU co-scheduling such that a GPU can be safely shared among multiple applications by taking memory constraints into account. Two approaches, namely a client-server and a shared memory approach are explored. However, the shared memory approach is more suitable due …
The method of choice for parameter aggregation in Deep Neural Network (DNN) training, a network-intensive task, is shifting from the Parameter Server model to decentralized aggregation schemes (AllReduce) inspired by theoretical guarantees of better performance. However, current implementations of AllReduce overlook the interdependence of communication and computation, resulting in significant performance degradation. In this paper, we develop Caramel, a system that accelerates decentralized distributed deep learning through model-aware computation scheduling and communication optimizations fo…
Computationally-intensive loops are the primary source of parallelism in scientific applications. Such loops are often irregular and a balanced execution of their loop iterations is critical for achieving high performance. However, several factors may lead to an imbalanced load execution, such as problem characteristics, algorithmic, and systemic variations. Dynamic loop self-scheduling (DLS) techniques are devised to mitigate these factors, and consequently, improve application performance. On distributed-memory systems, DLS techniques can be implemented using a hierarchical master-worker exe…
In this paper we study a scheduling problem arising from executing numerical simulations on HPC architectures. With a constant number of parallel machines, the objective is to minimize the makespan under memory constraints for the machines. Those constraints come from a neighborhood graph G for the jobs. Motivated by a previous result on graphs G with bounded path-width, our focus is on the case when the neighborhood graph G has bounded tree-width. Our result is a bi-criteria fully polynomial time approximation algorithm based on a dynamic programming algorithm. It allows to find a solution wi…
The blockchain technology has been extensively studied to enable distributed and tamper-proof data processing in federated learning (FL). Most existing blockchain assisted FL (BFL) frameworks have employed a third-party blockchain network to decentralize the model aggregation process. However, decentralized model aggregation is vulnerable to pooling and collusion attacks from the third-party blockchain network. Driven by this issue, we propose a novel BFL framework that features the integration of training and mining at the client side. To optimize the learning performance of FL, we propose to…
This paper investigates the execution of tree-shaped task graphs using multiple processors. Each edge of such a tree represents a large IO file. A task can only be executed if all input and output files fit into memory, and a file can only be removed from memory after it has been consumed. Such trees arise, for instance, in the multifrontal method of sparse matrix factorization. The maximum amount of memory needed depends on the execution order of the tasks. With one processor the objective of the tree traversal is to minimize the required memory. This problem was well studied and optimal poly…
Optimizing LLM Inference: Fluid-Guided Online Scheduling with Memory Constraints
Institute for Data, Systems, and Society, Massachusetts Institute of Technology（数据、系统与社会研究所，麻省理工学院）；School of Mathematical Sciences, Peking University（北京大学数学科学学院）；Alibaba Group（阿里巴巴集团）
FedFiTS: Fitness-Selected, Slotted Client Scheduling for Trustworthy Federated Learning in Healthcare AI
Department of Computer Science University of the Western Cape（计算机科学系，西开普敦大学）；Department of Computer Science Missouri University of Science and Technology（计算机科学系，密苏里科学与技术大学）
This paper investigates the execution of tree-shaped task graphs using multiple processors. Each edge of such a tree represents some large data. A task can only be executed if all input and output data fit into memory, and a data can only be removed from memory after the completion of the task that uses it as an input data. Such trees arise, for instance, in the multifrontal method of sparse matrix factorization. The peak memory needed for the processing of the entire tree depends on the execution order of the tasks. With one processor the objective of the tree traversal is to minimize the req…
Many emerging cyber-physical systems, such as autonomous vehicles and robots, rely heavily on artificial intelligence and machine learning algorithms to perform important system operations. Since these highly parallel applications are computationally intensive, they need to be accelerated by graphics processing units (GPUs) to meet stringent timing constraints. However, despite the wide adoption of GPUs, efficiently scheduling multiple GPU applications while providing rigorous real-time guarantees remains a challenge. In this paper, we propose RTGPU, which can schedule the execution of multipl…
With the rapidly growing demand of graph processing in the real scene, they have to efficiently handle massive concurrent jobs. Although existing work enable to efficiently handle single graph processing job, there are plenty of memory access redundancy caused by ignoring the characteristic of data access correlations. Motivated such an observation, we proposed two-level scheduling strategy in this paper, which enables to enhance the efficiency of data access and to accelerate the convergence speed of concurrent jobs. Firstly, correlations-aware job scheduling allows concurrent jobs to process…
Pipeline parallelism has been demonstrated to be a remarkable approach to improve throughput for training deep neural networks with billions of parameters over heterogeneous clusters. The 1F1B scheduling plan is a widely adopted strategy for memory and performance optimization, which interchanges the forward and backward stage computations of different micro-batches. On the other hand, a common issue in using the 1F1B scheduling is that stage computation is delayed due to the data transfer when network resources are preempted by other tasks, even with the minimum communication between stages. …
Recurrent neural networks (RNNs) have shown state of the art results for speech recognition, natural language processing, image captioning and video summarizing applications. Many of these applications run on low-power platforms, so their energy efficiency is extremely important. We observed that cache-oblivious RNN scheduling during inference typically results in 30-50x more data transferred on and off the CPU than the application's working set size. This can potentially impact its energy efficiency. This paper presents a new metric called Data Reuse Efficiency to gauge the RNN scheduling eff…
Computational Grids, emerging as an infrastructure for next generation computing, enable the sharing, selection, and aggregation of geographically distributed resources for solving large-scale problems in science, engineering, and commerce. As the resources in the Grid are heterogeneous and geographically distributed with varying availability and a variety of usage and cost policies for diverse users at different times and, priorities as well as goals that vary with time. The management of resources and application scheduling in such a large and distributed environment is a complex task. This …
University of California, San Diego（加利福尼亚大学圣地亚哥分校）；Qualcomm AI Research（高通人工智能研究院）；Duke University（杜克大学）
Recent advances demonstrate that irregularly wired neural networks from Neural Architecture Search (NAS) and Random Wiring can not only automate the design of deep neural networks but also emit models that outperform previous manual designs. These designs are especially effective while designing neural architectures under hard resource constraints (memory, MACs, . . . ) which highlights the importance of this class of designing neural networks. However, such a move creates complication in the previously streamlined pattern of execution. In fact one of the main challenges is that the order of s…
Although many scalable event matching algorithms have been proposed to achieve scalability for large-scale content-based networks, content-based publish/subscribe networks (especially for large-scale real time systems) still suffer performance deterioration when subscription scale increases. While subscription aggregation techniques can be useful to reduce the amount of subscription dissemination traffic and the subscription table size by exploiting the similarity among subscriptions, efficient subscription aggregation is not a trivial task to accomplish. Previous research works have proved th…
Microsoft Research（微软研究院）；University of Texas at Austin（德克萨斯大学奥斯汀分校）；VMware Research（威睿研究院）
Training Deep Neural Networks (DNNs) is a widely popular workload in both enterprises and cloud data centers. Existing schedulers for DNN training consider GPU as the dominant resource, and allocate other resources such as CPU and memory proportional to the number of GPUs requested by the job. Unfortunately, these schedulers do not consider the impact of a job's sensitivity to allocation of CPU, memory, and storage resources. In this work, we propose Synergy, a resource-sensitive scheduler for shared GPU clusters. Synergy infers the sensitivity of DNNs to different resources using optimistic p…
We study computer systems with transactions executed on a set of shared objects. Transactions arrive continually subjects to constrains that are framed as an adversarial model and impose limits on the average rate of transaction generation and the number of objects that transactions use. We show that no deterministic distributed scheduler in the queue-free model of transaction autonomy can provide stability for any positive rate of transaction generation. Let a system consist of $m$ shared objects and an adversary be constrained such that each transaction may access at most $k$ shared objects.…
Efficient runtime task scheduling on complex memory hierarchy becomes increasingly important as modern and future High-Performance Computing (HPC) systems are progressively composed of multisocket and multi-chiplet nodes with nonuniform memory access latencies. Existing locality-aware scheduling schemes either require control of the data placement policy for memory-bound tasks or maximize locality for all classes of computations, resulting in a loss of potential performance. While such approaches are viable, an adaptive scheduling strategy is preferred to enhance locality and resource sharing …
With the development of federated learning (FL), mobile devices (MDs) are able to train their local models with private data and sends them to a central server for aggregation, thereby preventing sensitive raw data leakage. In this paper, we aim to improve the training performance of FL systems in the context of wireless channels and stochastic energy arrivals of MDs. To this purpose, we dynamically optimize MDs' transmission power and training task scheduling. We first model this dynamic programming problem as a constrained Markov decision process (CMDP). Due to high dimensions rooted from ou…
