Workload Burstiness Research Articles

A Model-Based Approach Towards Real-Time Analytics in NFV Infrastructures

Network Functions Virtualization (NFV) has recently gained much popularity in the research scene for the flexibility and programmability that it will bring with the software implementation of network functions on commercial off-the-shelf (COTS) hardware. To substantiate its roll out, a number of issues (e.g., COTS’ inherent performance and energy efficiency, virtualization overhead, etc.) must be addressed, in a scalable and sustainable manner. Numerous works in the scientific literature manifest the strong correlation of network key performance indicators (KPIs) with the burstiness of the traffic. This paper proposes a novel model-based analytics approach for profiling virtualized network function (VNF) workloads, towards real-time estimation of network KPIs (specifically, power and latency), based on an MX/G/1/SET queueing model that captures both the workload burstiness and system setup times (caused by interrupt coalescing and power management actions). Experimental results show good estimation accuracies for both VNF workload profiling and network KPI estimation, with respect to the input traffic and actual measurements, respectively. This demonstrates that the proposed approach can be a powerful tool for scalable and sustainable network/service management and orchestration.

On evaluating the resource usage effectiveness of multi-tenant cloud storage

Abstract As multi-tenant cloud storage services are becoming increasingly popular in commercial clouds, there is a growing need for evaluating their resource usage effectiveness. Although tail latency and resource utilization are the most important and popular performance metrics, it is biased to use only one of them on evaluating the resource usage effectiveness. In this paper, we present a unified framework, Stochastic Model-based Effectiveness Analyzer (SMEA), to evaluate the resource usage effectiveness of multi-tenant cloud storage systems. Instead of evaluating tail latency and resource utilization separately, we propose a new metric called resource-productivity to combine both performance. Since evaluating tail latency or resource utilization solely is challenging due to contention, queueing and burstiness of workloads, a Markov-modulated Poisson process (MMPP) is adopted in SMEA to properly characterize the behavior of workloads. Moreover, the Stochastic Network Calculus (SNC)-based network latency analysis model is extended to evaluate the end-to-end tail latency over all system components, with the principle “treating the computer as a network”. After obtaining the analyzed tail latency and resource utilization, we derive the resource-productivity mathematically. We have implemented SMEA on a small-scale testbed. Extensive trace-driven experiments demonstrate that SMEA is efficient in evaluating the resource usage effectiveness of the multi-tenant cloud storage system with relative error less than 13%.

Admission control in shared memory switches

Cloud applications bring new challenges to the design of network elements, in particular the burstiness of traffic workloads. A shared memory switch is a good candidate architecture to exploit buffer capacity; in this work, we analyze the performance of this architecture. Our goal is to explore the impact of additional traffic characteristics such as varying processing requirements and packet values on objective functions. The outcome of this work is a better understanding of the relevant parameters for buffer management to achieve better performance in dynamic environments of data centers. We consider a model that captures more of the properties of the target architecture than previous work and consider several scheduling and buffer management algorithms that are specifically designed to optimize its performance. In particular, we provide analytic guarantees for the throughput performance of our algorithms that are independent from specific distributions of packet arrivals. We furthermore report on a comprehensive simulation study which validates our analytic results.

Performability evaluation of server virtualized systems under bursty workload

This paper proposes a performability analysis based on stochastic reward nets (SRNs) of an energy-efficient server virtualized system (SVS) model handling data-intensive applications and implementing a workload-aware power management (PM) mechanism. The SVS model accounts for workload-aware dependent software aging, unplanned failures and Migrate-VM rejuvenation. We show through numerical analysis of perfomability metrics how workload burstiness and software aging impact SVS behavior in terms of availability and performance.

Burstiness-aware service level planning for enterprise application clouds

Enterprise applications are being increasingly deployed on cloud infrastructures. Often, a cloud service provider (SP) enters into a Service Level Agreement (SLA) with a cloud subscriber, which specifies performance requirements for the subscriber’s applications. An SP needs systematic Service Level Planning (SLP) tools that can help estimate the resources needed and hence the cost incurred to satisfy their customers’ SLAs. Enterprise applications typically experience bursty workloads and the impact of such bursts needs to be considered during SLP exercises. Unfortunately, most existing approaches do not consider workload burstiness. We propose a Resource Allocation Planning (RAP) technique, which allows an SP to identify a time varying allocation plan of resources to applications that satisfies bursts. Extensive simulation results show that the proposed RAP variants can identify resource allocation plans that satisfy SLAs without exhaustively generating all possible plans. Furthermore, the results show that RAP can permit SPs to more accurately determine the capacity required for meeting specified SLAs compared to other competing techniques especially for bursty workloads.

DiskAccel

Disk traces are typically used to analyze real-life workloads and for replay-based evaluations. This approach benefits from capturing important details such as varying behavior patterns, bursty activity, and diurnal patterns of system activity, which are often missing from the behavior of workload synthesis tools. However, accurate capture of such details requires recording traces containing long durations of system activity, which are difficult to use for replay-based evaluation. One way of solving the problem of long storage trace duration is the use of disk simulators. While publicly available disk simulators can greatly accelerate experiments, they have not kept up with technological innovations in the field. The variety, complexity, and opaque nature of storage hardware make it very difficult to implement accurate simulators. The alternative, replaying the whole traces on real hardware, suffers from either long run-time or required manual reduction of experimental time, potentially at the cost of reduced accuracy. On the other hand, burstiness, auto-correlation, and complex spatio-temporal properties of storage workloads make the known methods of sampling workload traces less effective. In this paper, we present a methodology called DiskAccel to efficiently select key intervals of a trace as representatives and to replay them to estimate the response time of the whole workload. Our methodology extracts a variety of spatial and temporal features from each interval and uses efficient data mining techniques to select the representative intervals. To verify the proposed methodology, we have implemented a tool capable of running whole traces or selective intervals on real hardware, warming up hardware state in an accelerated manner, and emulating request causality while minimizing request inter-arrival time error. Based on our experiments, DiskAccel manages to speed up disk replay by more than two orders of magnitude, while keeping average estimation error at 7.6%.

Automated and Agile Server ParameterTuning by Coordinated Learning and Control

Automated server parameter tuning is crucial to performance and availability of Internet applications hosted in cloud environments. It is challenging due to high dynamics and burstiness of workloads, multi-tier service architecture, and virtualized server infrastructure. In this paper, we investigate automated and agile server parameter tuning for maximizing effective throughput of multi-tier Internet applications. A recent study proposed a reinforcement learning based server parameter tuning approach for minimizing average response time of multi-tier applications. Reinforcement learning is a decision making process determining the parameter tuning direction based on trial-and-error, instead of quantitative values for agile parameter tuning. It relies on a predefined adjustment value for each tuning action. However it is nontrivial or even infeasible to find an optimal value under highly dynamic and bursty workloads. We design a neural fuzzy control based approach that combines the strengths of fast online learning and self-adaptiveness of neural networks and fuzzy control. Due to the model independence, it is robust to highly dynamic and bursty workloads. It is agile in server parameter tuning due to its quantitative control outputs. We implemented the new approach on a testbed of virtualized data center hosting RUBiS and WikiBench benchmark applications. Experimental results demonstrate that the new approach significantly outperforms the reinforcement learning based approach for both improving effective system throughput and minimizing average response time.

Sizing multi-tier systems with temporal dependence: benchmarks and analytic models

Abstract Temporal dependence, as a synonym for burstiness, is often found in workloads (i.e., arrival flows and/or service times) in enterprise systems that use the multi-tier paradigm. Despite the fact that burstiness has deleterious effects on performance, existing modeling and benchmarking techniques do not provide an effective capacity planning for multi-tier systems with temporal dependence. In this paper, we first present strong evidence that existing models cannot capture bursty conditions and accurately predict performance. Therefore, we propose a simple and effective sizing methodology to integrate workload burstiness into models and benchmarking tools used in system sizing. This modeling methodology is based on the index of dispersion which jointly captures variability and burstiness of the service process in a single number. We report experimentation on a real testbed that validates the accuracy of our modeling technique by showing that experimental and model prediction results are in excellent agreement under both bursty and non-bursty workloads. To further support the capacity planning process under burstiness, we propose an enhanced benchmarking technique that can emulate workload burstiness in systems. We find that most existing benchmarks, like the standard TPC-W benchmark, are designed to assess system performance only under non-bursty conditions. In this work, we rectify this deficiency by introducing a new module into existing benchmarks, which allows to inject burstiness into the arrival stream in a controllable and reproducible manner by using the index of dispersion as a single turnable knob. This approach enables a better understanding of system performance degradation due to burstiness and makes a strong case for the usefulness of the proposed benchmark enhancement for capacity planning of enterprise systems.

Bound analysis of closed queueing networks with workload burstiness

Burstiness and temporal dependence in service processes are often found in multi-tier architectures and storage devices and must be captured accurately in capacity planning models as these features are responsible of significant performance degradations. However, existing models and approximations for networks of first-come first-served (FCFS) queues with general independent (GI) service are unable to predict performance of systems with temporal dependence in workloads. To overcome this difficulty, we define and study a class of closed queueing networks where service times are represented by Markovian Arrival Processes (MAPs), a class of point processes that can model general distributions, but also temporal dependent features such as burstiness in service times. We call these models MAP queueing networks. We introduce provable upper and lower bounds for arbitrary performance indexes (e.g., throughput, response time, utilization) that we call Linear Reduction (LR) bounds. Numerical experiments indicate that LR bounds achieve a mean accuracy error of 2 percent. The result promotes LR bounds as a versatile and reliable bounding methodology of the performance of modern computer systems.