Jianzhe Tai

ArA: Adaptive resource allocation for cloud computing environments under bursty workloads

Cloud computing nowadays becomes quite popular among a community of cloud users by offering a variety of resources. However, burstiness in user demands often dramatically degrades the application performance. In order to satisfy peak user demands and meet Service Level Agreement (SLA), efficient resource allocation schemes are highly demanded in the cloud. However, we find that conventional load balancers unfortunately neglect cases of bursty arrivals and thus experience significant performance degradation. Motivated by this problem, we propose new burstiness-aware algorithms to balance bursty workloads across all computing sites, and thus to improve overall system performance. We present a smart load balancer, which leverages the knowledge of burstiness to predict the changes in user demands and on-the-fly shifts between the schemes that are “greedy” (i.e., always select the best site) and “random” (i.e., randomly select one) based on the predicted information. Both simulation and real experimental results show that this new load balancer can adapt quickly to the changes in user demands and thus improve performance by making a smart site selection for cloud users under both bursty and non-bursty workloads.

GReM: Dynamic SSD resource allocation in virtualized storage systems with heterogeneous IO workloads

In a shared virtualized storage system that runs VMs with heterogeneous IO demands, it becomes a problem for the hypervisor to cost-effectively partition and allocate SSD resources among multiple VMs. There are two straightforward approaches to solving this problem: equally assigning SSDs to each VM or managing SSD resources in a fair competition mode. Unfortunately, neither of these approaches can fully utilize the benefits of SSD resources, particularly when the workloads frequently change and bursty IOs occur from time to time. In this paper, we design a Global SSD Resource Management solution — GReM, which aims to fully utilize SSD resources as a second-level cache under the consideration of performance isolation. In particular, GReM takes dynamic IO demands of all VMs into consideration to split the entire SSD space into a long-term zone and a short-term zone, and cost-effectively updates the content of SSDs in these two zones. GReM is able to adaptively adjust the reservation for each VM inside the long-term zone based on their IO changes. GReM can further dynamically partition SSDs between the long- and short-term zones during runtime by leveraging the feedbacks from both cache performance and bursty workloads. Experimental results show that GReM can capture the cross-VM IO changes to make correct decisions on resource allocation, and thus obtain high IO hit ratio and low IO management costs, compared with both traditional and state-of-the-art caching algorithms.

Improving Flash Resource Utilization at Minimal Management Cost in Virtualized Flash-Based Storage Systems

Effectively leveraging Flash resources has emerged as a highly important problem in enterprise storage systems. One of the popular techniques today is to use Flash as a secondary-level host-side cache in the virtual machine environment. Although this approach delivers IO acceleration for VMs’ IO workloads, it might not be able to fully exploit the outstanding performance of Flash and justify the high cost-per-GB of Flash resources. In this paper, we design new VMware Flash Resource Managers (vFRM and glb-vFRM) under the consideration of both performance and the incurred cost for managing Flash resources. Specifically, vFRM and glb-vFRM aim to maximize the utilization of Flash resources with minimal CPU, memory and IO cost in managing and operating Flash for a dedicated enterprise workload and multiple heterogeneous enterprise workloads, respectively. Our new Flash resource managers adopt the ideas of thermodynamic heating and cooling to identify data blocks that can benefit the most from being put on Flash and migrate data blocks between Flash and magnetic disks in a lazy and asynchronous mode. Experimental evaluation of the prototype shows that both vFRM and glb-vFRM achieve better cost-effectiveness than traditional caching solutions, i.e., obtaining IO hit ratios even slightly better than some of the conventional algorithms as Flash size increases yet costing orders of magnitude less IO bandwidth.

LsPS: A Job Size-Based Scheduler for Efficient Task Assignments in Hadoop

The MapReduce paradigm and its open source implementation Hadoop are emerging as an important standard for large-scale data-intensive processing in both industry and academia. A MapReduce cluster is typically shared among multiple users with different types of workloads. When a flock of jobs are concurrently submitted to a MapReduce cluster, they compete for the shared resources and the overall system performance in terms of job response times, might be seriously degraded. Therefore, one challenging issue is the ability of efficient scheduling in such a shared MapReduce environment. However, we find that conventional scheduling algorithms supported by Hadoop cannot always guarantee good average response times under different workloads. To address this issue, we propose a new Hadoop scheduler, which leverages the knowledge of workload patterns to reduce average job response times by dynamically tuning the resource shares among users and the scheduling algorithms for each user. Both simulation and real experimental results from Amazon EC2 cluster show that our scheduler reduces the average MapReduce job response time under a variety of system workloads compared to the existing FIFO and Fair schedulers.

VFRM: Flash Resource Manager in VMware ESX Server

One popular approach of leveraging Flash technology in the virtual machine environment today is using it as a secondary-level host-side cache. Although this approach delivers I/O acceleration for a single VM workload, it might not be able to fully exploit the outstanding performance of Flash and justify the high cost-per-GB of Flash resources. In this paper, we present the design for VMware Flash Resource Manager (VFRM), which aims to maximize the utilization of Flash resources with minimal CPU, memory and I/O cost for managing and operating Flash. It borrows the ideas of heating and cooling from thermodynamics to identify the data blocks that benefit most from being put on Flash, and lazily and asynchronously migrates the data blocks between Flash and spinning disks. Experimental evaluation of the prototype shows that VFRM achieves better cost-effectiveness than traditional caching solutions, and costs orders of magnitude less memory and I/O bandwidth.

Live Data Migration for Reducing SLA Violations in Multi-tiered Storage Systems

Today, the volume of data in the world has been tremendously increased. Large-scaled and diverse data sets are raising new big challenges of storage, process, and query. Tiered storage architectures combining solid-state drives (SSDs) with hard disk drives (HDDs), become attractive in enterprise data centers for achieving high performance and large capacity simultaneously. However, how to best use these storage resources and efficiently manage massive data for providing high quality of service (QoS) is still a core and difficult problem. In this paper, we present a new approach for automated data movement in multi-tiered storage systems, which lively migrates the data across different tiers, aiming to support multiple service level agreements (SLAs) for applications with dynamic workloads at the minimal cost. Trace-driven simulations show that compared to the no migration policy, LMsT significantly improves average I/O response times, I/O violation ratios and I/O violation times, with only slight degradation (e.g., up to 6% increase in SLA violation ratio) on the performance of high priority applications.

Load balancing for cluster systems under heavy-tailed and temporal dependent workloads

Abstract Large-scaled cluster systems have been employed in various areas by offering pools of fundamental resources. Efficient allocation of the shared resources in a cluster system is a critical but challenging issue, which has been extensively studied in the past few years. Despite the fact that existing load balancing policies, such as Random, Join Shortest Queue and size-based polices, are widely implemented in actual systems due to their simplicity and efficiency, the performance benefits of these policies diminish when workloads are highly variable and temporally correlated. In this paper, we propose a new load balancing policy, named AD u S, which attempts to partition jobs according to their present sizes and further rank the servers based on their loads. By dispatching jobs of similar sizes to the corresponding ranked servers, AD u S can adaptively balance user traffic and system load in a cluster and thus achieve significant performance benefits. Extensive trace-driven simulations using both synthetic and real traces show the effectiveness and robustness of AD u S under many different environments.

Decentralized Scheduling of Bursty Workload on Computing Grids

Bursty workloads are often observed in a variety of systems such as grid services, multi-tier architectures, and large storage systems. Studies have shown that such burstiness can dramatically degrade system performance because of overloading, increased response time, and unavailable service. Computing grids, which often use distributed, autonomous resource management, are particularly susceptible to load imbalances caused by bursty workloads. In this paper, we use a simulation environment to investigate the performance of decentralized schedulers under various intensity levels of burstiness. We first demonstrate a significant performance degradation in the presence of strong and moderate bursty workloads. Then, we describe two new hybrid schedulers, based on duplication-invalidation, and assess the effectiveness of these schedulers under different intensities of burstiness. Our simulation results show that compared to the conventional decentralized methods, the proposed schedulers achieve a 40% performance improvement under the bursty condition while obtaining similar performance in non-bursty conditions.

SLA-aware data migration in a shared hybrid storage cluster

Data volume in today’s world has been tremendously increased. Large-scaled and diverse data sets are raising new big challenges of storage, process, and query. Particularly, real-time data analysis becomes more and more frequently. Multi-tiered, hybrid storage architectures, which provide a solid way to combine solid-state drives with hard disk drives (HDDs), therefore become attractive in enterprise data centers for achieving high performance and large capacity simultaneously. However, from service provider’s perspective, how to efficiently manage all the data hosted in data center in order to provide high quality of service (QoS) is still a core and difficult problem. The modern enterprise data centers often provide the shared storage resources to a large variety of applications which might demand for different service level agreements (SLAs). Furthermore, any user query from a data-intensive application could easily trigger a scan of a gigantic data set and inject a burst of disk I/Os to the back-end storage system, which will eventually cause disastrous performance degradation. Therefore, in the paper, we present a new approach for automated data movement in multi-tiered, hybrid storage clusters, which lively migrates the data among different storage media devices, aiming to support multiple SLAs for applications with dynamic workloads at the minimal cost. Detailed trace-driven simulations show that this new approach significantly improves the overall performance, providing higher QoS for applications and reducing the occurrence of SLA violations. Sensitivity analysis under different system environments further validates the effectiveness and robustness of the approach.

ADuS: Adaptive resource allocation in cluster systems under heavy-tailed and bursty workloads

A large-scaled cluster system has been employed in various areas by offering pools of fundamental resources. How to effectively allocate the shared resources in a cluster system is a critical but challenging issue, which has been extensively studied in the past few years. Despite the fact that classic load balancing policies, such as Random, Join Shortest Queue and size-based polices, are widely implemented in actual systems due to their simplicity and efficiency, the performance benefits of these policies diminish when workloads are highly variable and heavily dependent. In this paper, we propose a new load balancing policy named ADuS, which attempts to partition jobs according to their sizes and to further rank the servers based on their loads. By dispatching jobs of similar size to the servers with the same ranking, ADuS can adaptively balance user traffic and system load in the system and thus achieve significant performance benefits. Extensive simulations show the effectiveness and the robustness of ADuS under many different environments.