Data Placement Mechanism Research Articles

SIMPLE: Efficient Temporal Graph Neural Network Training at Scale with Dynamic Data Placement

Dynamic graphs are essential in real-world scenarios like social media and e-commerce for tasks such as predicting links and classifying nodes. Temporal Graph Neural Networks (T-GNNs) stand out as a prime solution for managing dynamic graphs, employing temporal message passing to compute node embeddings at specific timestamps. Nonetheless, the high CPU-GPU data loading overhead has become the bottleneck for efficient training of T-GNNs over large-scale dynamic graphs. In this work, we present SIMPLE, a versatile system designed to address the major efficiency bottleneck in training existing T-GNNs on a large scale. It incorporates a dynamic data placement mechanism, which maintains a small buffer space in available GPU memory and dynamically manages its content during T-GNN training. SIMPLE is also empowered by systematic optimizations towards data processing flow. We compare SIMPLE to the state-of-the-art generic T-GNN training system TGL on four large-scale dynamic graphs with different underlying T-GNN models. Extensive experimental results show that SIMPLE effectively cuts down 80.5% ~ 96.8% data loading cost, and accelerates T-GNN training by 1.8× ~ 3.8× (2.6× on average) compared to TGL.

Modeling Retention Errors of 3D NAND Flash for Optimizing Data Placement

Considering 3D NAND flash has a new property of process variation (PV) , which causes different raw bit error rates (RBER) among different layers of the flash block. This paper builds a mathematical model for estimating the retention errors of flash cells, by considering the factor of layer-to-layer PV in 3D NAND flash memory, as well as the factors of program/erase (P/E) cycle and retention time of data. Then, it proposes classifying the layers of flash block in 3D NAND flash memory into profitable and unprofitable categories, according to the error correction overhead. After understanding the retention error variation of different layers in 3D NAND flash, we design a mechanism of data placement, which maps the write data onto a suitable layer of flash block, according to the data hotness and the error correction overhead of layers, to boost read performance of 3D NAND flash. The experimental results demonstrate that our proposed retention error estimation model can yield a R 2 value of 0.966 on average, verifying the accuracy of the model. Based on the estimated retention error rates of layers, the proposed data placement mechanism can noticeably reduce the read latency by 29.8 % on average, compared with state-of-the-art methods against retention errors for 3D NAND flash memory.

DownShift: Tuning Shift Reduction With Reliability for Racetrack Memories

Ultra-dense non-volatile <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">racetrack memories</i> (RTMs) have been investigated at various levels in the memory hierarchy for improved performance and reduced energy consumption. However, the innate <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">shift</i> operations in RTMs, required for data access, incur performance penalties and can induce position errors. These factors can hinder their applicability in replacing low-latency, reliable on-chip memories. Intelligent placement of memory objects in RTMs can significantly reduce the number of shifts per memory access with little to no hardware overhead. However, existing placement strategies may lead to sub-optimal performance when applied to different architectures. Additionally, the impact of these shift optimization techniques on RTM reliability has been insufficiently investigated. We propose DownShift, a generalized data placement mechanism that improves upon prior approaches by taking into account (1) the timing and liveliness information of memory objects and (2) the underlying memory architecture, including required shifting fault tolerance. Thus, we also propose a collaboratively designed new shift alignment reliability technique called GROGU. GROGU leverages the reduced shift window made possible through DownShift allowing improved reliability, area, and energy compared to the state-of-the-art reliability approaches. DownShift reduces the number of shifts, runtime, and energy consumption by 3.24×, 47.6%, and 70.8% compared to the state-of-the-art. GROGU consumes 2.2× less area and 1.3× less energy while providing 16.8× improvement in shift fault tolerance compared to the leading reliability approach for a latency degradation of only 3.2%.

A machine learning assisted data placement mechanism for hybrid storage systems

Emerging applications produce massive files that show different properties in file size, lifetime, and read/write frequency. Existing hybrid storage systems place these files onto different storage mediums assuming that the access patterns of files are fixed. However, we find that the access patterns of files are changeable during their lifetime. The key to improve the file access performance is to adaptively place the files on the hybrid storage system using the run-time status and the properties of both files and the storage systems. In this paper, we propose a machine learning assisted data placement mechanism that adaptively places files onto the proper storage medium by predicting access patterns of files. We design a PMFS based tracer to collect file access features for prediction and show how this approach is adaptive to the changeable access pattern. Based on data access prediction results, we present a linear data placement algorithm to optimize the data access performance on the hybrid storage mediums. Extensive experimental results show that the proposed learning algorithm can achieve over 90% accuracy for predicting file access patterns. Meanwhile, this paper can achieve over 17% improvement of system performance for file accesses compared with the state-of-the-art linear-time data placement methods.

SLA-Aware Multi-Criteria Data Placement in Cloud Storage Systems

This paper proposes a multi-criteria data placement mechanism for typical cloud storage systems. The primary goal of this study is to place the clients’ files in the storage cluster by taking into account the principles of service level agreements and the levels of user support services. For this purpose, a multi-criteria scoring model is defined based on a trade-off between three criteria that can be crucial from the clients’ viewpoint: data access latency, data privacy, and computational performance. The second goal of the proposed mechanism is to improve the fairness of the storage allocation mechanism for various kinds of files compared to the conventional mechanisms. In this case, the cloud service provider gives each data storage node a chance to become the host of an already received file using a probabilistic ranking model. The experimental results demonstrate that the proposed data placement mechanism increases the robustness of cloud storage architecture under different file, network, and storage configurations compared to the existing data placement mechanisms.

Data Placement Oriented Scheduling Algorithm for Scheduling Scientific Workflow in Cloud: A Budget-Aware Approach

Background: A large number of communities and enterprises deploy numerous scientific workflow applications on cloud service. Aims: The main aim of the cloud service provider is to execute the workflows with a minimal budget and makespan. Most of the existing techniques for budget and makespan are employed for the traditional platform of computing and are not applicable to cloud computing platforms with unique resource management methods and pricing strategies based on service. Methods: In this paper, we studied the joint optimization of cost and makespan of scheduling workflows in IaaS clouds, and proposed a novel workflow scheduling scheme. Also, data placement is included in the proposed algorithm. Results: In this scheme, DPO-HEFT (Data Placement Oriented HEFT) algorithm is developed which closely integrates the data placement mechanism with the list scheduling heuristic HEFT. Extensive experiments using the real-world and synthetic workflow demonstrate the efficacy of our scheme. Conclusion: Our scheme can achieve significantly better cost and makespan trade-off fronts with remarkably higher hypervolume and can run up to hundreds times faster than the state-of-the-art algorithms.

An adaptive replica placement approach for distributed key‐value stores

SummaryThe use of distributed key‐value stores (KVS) has experienced fast adoption by various applications in recent years due to key advantages such as hypertext transfer protocol‐based RESTful application programming interface, high availability and elasticity. Due to great scalability characteristics, KVS systems commonly use consistent hashing as data placement mechanism. Although KVS systems offer many advantages, they were not designed to dynamically adapt to changing workloads which often include data access skew. Furthermore, the underlying physical storage nodes may be heterogeneous and do not expose their performance capabilities to higher level data placement layers. In this paper, we address those issues and propose an essential step toward a dynamic autonomous solution by leveraging deep reinforcement learning. We design a self‐learning approach that incrementally changes the data placement, improving the load balancing. Our approach is dynamic in the sense that is capable of avoiding hot spots, that is, overloaded storage nodes when facing different workloads. Also, we design our solution to be pluggable. It assumes no previous knowledge of the storage nodes capabilities, thus different KVS deployments may make use of it. Our experiments show that our method performs well on changing workloads including data access skew aspects. We demonstrate the effectiveness of our approach through experiments in a distributed KVS deployment.

Lessons Learned from Optimizing the Sunway Storage System for Higher Application I/O Performance

It is hard for applications to make full utilization of the peak bandwidth of the storage system in highperformance computers because of I/O interferences, storage resource misallocations and complex long I/O paths. We performed several studies to bridge this gap in the Sunway storage system, which serves the supercomputer Sunway TaihuLight. To locate these issues and connections between them, an end-to-end performance monitoring and diagnosis tool was developed to understand I/O behaviors of applications and the system. With the help of the tool, we were about to find out the root causes of such performance barriers at the I/O forwarding layer and the parallel file system layer. An application-aware I/O forwarding allocation framework was used to address the I/O interferences and resource misallocations at the I/O forwarding layer. A performance-aware data placement mechanism was proposed to mitigate the impact of I/O interferences and performance variations of storage devices in the PFS. Together, applications obtained much better I/O performance. During the process, we also proposed a lightweight storage stack to shorten the I/O path of applications with -N I/O pattern. This paper summarizes these studies and presents the lessons learned from the process.

Data Replica Placement Mechanism for Open Heterogeneous Storage Systems

Abstract: Many big data-oriented network storage platforms are open and heterogeneous, which means that the data placement mechanism has a great impact on the system performance. The data replica placement mechanism needs to consider the data availability, the load balance and the quality of service for open heterogeneous storage systems. However, current data placement mechanisms are mainly for simple homogeneous storage systems, not entirely suitable for heterogeneous storage systems. We proposed a novel data replica placement mechanism with the adjustable replica deployment strategy (ARDS) for open heterogeneous data storage systems, which considers the data availability, the data access frequency and the storage capacity. The mechanism takes the data availability as premise of the initialization number of replica, then adjusts the number of replica according to the data access frequency, and places replicas based on the node spare capacity, which effectively avoids the condition of load unbalance. We implemented simulation experiments based on OMNet++. The experimental results show that the ARDS-based mechanism improves the system performance including the system load balance and the rate of successful data access.