An Architecture for Load Balance in Computer Cluster Applications

Amid a data revolution that is transforming industries around the globe, computing systems have undergone a paradigm shift where many applications are scaled out to run on multiple computers in a computing cluster. As the storage and processing capabilities of a single machine are unable to keep pace with the amount of data, companies turn to distributed solutions to organize, persist, and analyze this Big Data. These new software solutions divide datasets into partitions that are processed in parallel on separate machines. A common problem in current cluster computing frameworks is load imbalance and limited parallelism due to skewed data distributions, processing times, and machine speeds. Load imbalance occurs when a few machines have more work and take longer to finish while the others remain idle, resulting in reduced overall performance and low resource utilization. The underlying cause for these issues is that data locality, where machines process the data stored locally, leads to tight coupling between a partition and the machine on which it is placed. This dissertation proposes a novel scatter architecture for computer cluster applications that effectively addresses the load imbalance problem and improves resource utilization. Scatter systems abandon data locality in favor of load balance. Existing systems first target locality, bringing computation to the data, and only then attempt to minimize load imbalance. Instead, we disregard any locality concerns and focus solely on achieving load balance by scattering all data across machines and bringing data to the computation as needed. The scatter architecture disaggregates compute and storage resources and pools all resources together across machines. We ensure storage load balance by spreading the data for each partition in small blocks across all storage devices and allow machines to retrieve data using an efficient, decentralized scheme. Scatter systems achieve compute load balance at runtime by dynamically adjusting the parallelism within a partition, either through work-sharing or by offloading background tasks to other machines. We design and implement three cluster applications inspired by the scatter architecture: a graph processing system that scales out using secondary storage, a general-purpose analytics framework, and a filesystem substrate that mitigates load imbalance for existing distributed databases. We demonstrate that each application can gracefully handle significant load imbalance with minimal performance degradation and that these applications can perform up to an order of magnitude faster than existing systems.

Optimal application-oriented resource brokering in a high performance computing grid

Vincent Keller

In the world of High Performance Computing (newly renamed "High Productivity Computing"), where the race for performance is on, where the hunt for the last Flop rages and where the developers swear by the Cult of Power, the different applications have different needs. Some of them are single threaded, others are parallel, some are memory bandwidth, others processor speed dominated. Some are embarrassingly parallel, point-to-point or multicast dominated, all depending on the algorithms and numerical methods they implement. Computing resources are also different, some are low-cost clusters of PlayStations 3, clusters with more or less efficient networks, others are expensive vector supercomputers. To express that diversity, a parameterization of the applications and the resources is proposed. Based on this parameterization, it is possible to quantify the needs of the different types of applications in terms of memory, processor and network requirements. On the other hand, the parameterization of the resources quantifies what a resource "offers" in term of memory, processor or network performance. Comparing the needs of the application and the offers of the resources, we present a model that predicts the computation and the communication times, and consequently the execution times of an application on different resources. We define a cost model that finds the best matches between an application and the available resources in an HPC Grid. It takes into account the cost of execution time of the application (based on the prediction previously mentioned), the waiting time cost (based on the information given by the local resource management systems), the cost of the licenses of the different software the application may use, the cost of the data transfer, and the ecological cost. The choice is made on a QOS (Quality of Service) given by the user, such as minimum turnaround time or minimum cost. Computing these different costs requires a fine knowledge of the behavior of the application on the resource. Thus we present a detailed application-oriented monitoring that consists in mapping information from the hardware monitoring and the different middleware the application (or the resource) uses, such as MPI detailed information, the hardware counters or the information of the execution on a local resource management system. The monitored data is stored in an archive system. The archived monitoring data concerning the behavior of the applications on the different resources, can then be used to help deciders purchasing new best suited resources for the current (and future) applications. In the second part of this work, we go further in matching the applications and resources: if it is possible to predict the behavior of an application on a given resource, it is also possible to fit the resource to the needs of the application. Knowing the processor performance needs of the application during its execution, it is possible to adjust its frequency through the frequency stepping mechanism that authorizes an operating system kernel or a program to modify on the fly the frequency (and the voltage) of the present days processors. Hence the resource exactly fits the needs of the application and the energy consumption of the processor is reduced. Finally, we propose a methodology to determine poorly implemented applications and thus giving users and developers hints and advice to detect flaws in an automatic manner. We present a real-life example, a spectral element code in CFD, for which the complexity changed when more than 1024 processors were used.

EPFL2008

Linear Scalability of Distributed Applications

Nicolas Bonvin

The explosion of social applications such as Facebook, LinkedIn and Twitter, of electronic commerce with companies like Amazon.com and Ebay.com, and of Internet search has created the need for new technologies and appropriate systems to manage effectively a considerable amount of data and users. These applications must run continuously every day of the year and must be capable of surviving sudden and abrupt load increases as well as all kinds of software, hardware, human and organizational failures. Increasing (or decreasing) the allocated resources of a distributed application in an elastic and scalable manner, while satisfying requirements on availability and performance in a cost-effective way, is essential for the commercial viability but it poses great challenges in today's infrastructures. Indeed, Cloud Computing can provide resources on demand: it now becomes easy to start dozens of servers in parallel (computational resources) or to store a huge amount of data (storage resources), even for a very limited period, paying only for the resources consumed. However, these complex infrastructures consisting of heterogeneous and low-cost resources are failure-prone. Also, although cloud resources are deemed to be virtually unlimited, only adequate resource management and demand multiplexing can meet customer requirements and avoid performance deteriorations. In this thesis, we deal with adaptive management of cloud resources under specific application requirements. First, in the intra-cloud environment, we address the problem of cloud storage resource management with availability guarantees and find the optimal resource allocation in a decentralized way by means of a virtual economy. Data replicas migrate, replicate or delete themselves according to their economic fitness. Our approach responds effectively to sudden load increases or failures and makes best use of the geographical distance between nodes to improve application-specific data availability. We then propose a decentralized approach for adaptive management of computational resources for applications requiring high availability and performance guarantees under load spikes, sudden failures or cloud resource updates. Our approach involves a virtual economy among service components (similar to the one among data replicas) and an innovative cascading scheme for setting up the performance goals of individual components so as to meet the overall application requirements. Our approach manages to meet application requirements with the minimum resources, by allocating new ones or releasing redundant ones. Finally, as cloud storage vendors offer online services at different rates, which can vary widely due to second-degree price discrimination, we present an inter-cloud storage resource allocation method to aggregate resources from different storage vendors and provide to the user a system which guarantees the best rate to host and serve its data, while satisfying the user requirements on availability, durability, latency, etc. Our system continuously optimizes the placement of data according to its type and usage pattern, and minimizes migration costs from one provider to another, thereby avoiding vendor lock-in.

EPFL2012

Architecture Exploration and Optimization of Heterogeneous Many-Core Compute and Memory Architectures with Architectural Extensions

Yasir Mahmood Qureshi

The expeditious proliferation of Internet connectivity and the growing adoption of digital products have transformed various spheres of our everyday lives. This increased digitization of society has led to the emergence of new applications, which are deployed all the way from High Performance Computing (HPC) systems and cloud servers to mobile devices. These new emerging applications like video analytics, autonomous driving, natural language processing, content recommendation, bioinformatics, and genome sequencing, have different performance/energy requirements, and are deployed on a variety of different platforms. To meet the performance and Quality-of-Service (QoS)constraints, cloud servers, and HPC platforms are comprised of many-core processors. Even the processors in today's mobile and edge devices are multi-core. To optimize energy efficiency and performance, a system-level simulator is required. This simulator must be capable of simultaneously executing multi-threaded applications in a full-system environment with a complete operating system (OS), on a heterogeneous many-core system.In this thesis, I present gem5-X, a system-level simulation platform to optimize many-core heterogeneous compute and memory architectures. Gem5-X exploration and optimization methodology is also proposed to optimize both the compute and memory sub-system for multi-thread applications. Gem5-X extends gem5 with architectural extensions for the compute and memory sub-systems, including in-caching computing accelerator, clustered heterogeneous cores, in-memory computing engine, and 3D stacked High Bandwidth Memory v2 (HBM2). It also adds support enhancements to gem5, including, Virtual Machine (VM) support, profiling support in the simulator using gperf profiler, enhanced checkpointing, and file sharing between the simulated and the host system. Finally, all the extensions and enhancements are fully supported in Full System (FS)mode of gem5-X, enabling booting and running a full Linux stack on top of the simulator, allowing almost any application running in a Linux system to be executed using gem5-X.To demonstrate the capabilities of gem5-X, various compute-dominated and memory-dominated applications are used as case studies. These state-of-the-art applications include video encoding, video analytics, VMs in the cloud, Binary Neural Networks (BNNs) and Recurrent Neural Networks (RNNs) as compute-intensive workloads. In addition to compute-dominated workloads, memory-dominated applications, including genome sequence alignment based on Next Generation Sequencing (NGS) techniques, and Convolutional Neural Networks (CNNs) are presented as case studies to demonstrate the memory sub-system extensions of gem5-X. Gem5-X exploration methodology is used to optimize architectures for these applications, achieving both performance and energy benefits. All the applications and workloads used in this thesis are case studies to showcase the capabilities of gem5-X. Gem5-X is generic and can be used to optimize and explore architectures for any multi-threaded (or single-threaded) application that can run on a Linux-based OS.

EPFL2021