Navaneeth Rameshan

Stay-Away , protecting sensitive applications from performance interference

While co-locating virtual machines improves utilization in resource shared environments, the resulting performance interference between VMs is difficult to model or predict. QoS sensitive applications can suffer from resource co-location with other less short-term resource sensitive or batch applications. The common practice of overprovisioning resources helps to avoid performance interference and guarantee QoS but leads to low machine utilization. Recent work that relies on static approaches suffer from practical limitations due to assumptions such as a priori knowledge of application behaviour and workload. To address these limitations, we present Stay-Away, a generic and adaptive mechanism to mitigate the detrimental effects of performance interference on sensitive applications when co-located with batch applications. Our mechanism complements the allocation decisions of resource schedulers by continuously learning the favourable and unfavourable states of co-execution and mapping them to a state-space representation. Trajectories in this representation are used to predict and prevent any transition towards interference of sensitive applications by proactively throttling the execution of batch applications. The representation also doubles as a template to prevent violations in the future execution of the repeatable sensitive application when co-located with other batch applications. Experimental results with realistic applications show that it is possible to guarantee a high level of QoS for latency sensitive applications while also improving machine utilization.

ProRenaTa: Proactive and Reactive Tuning to Scale a Distributed Storage System

Provisioning tasteful services in the Cloud that guarantees high quality of service with reduced hosting cost is challenging to achieve. There are two typical auto-scaling approaches: predictive and reactive. A prediction based controller leaves the system enough time to react to workload changes while a feedback based controller scales the system with better accuracy. In this paper, we show the limitations of using a proactive or reactive approach in isolation to scale a tasteful system and the overhead involved. To overcome the limitations, we implement an elasticity controller, ProRenaTa, which combines both reactive and proactive approaches to leverage on their respective advantages and also implements a data migration model to handle the scaling overhead. We show that the combination of reactive and proactive approaches outperforms the state of the art approaches. Our experiments with Wikipedia workload trace indicate that ProRenaTa guarantees a high level of SLA commitments while improving the overall resource utilization.

Energy Efficiency Dilemma: P2P-cloud vs. Datacenter

Energy consumption is increasing in the IT sector and a remarkable part of this energy is consumed in data centers. Numerous techniques have been proposed to solve the energy efficiency issue in cloud systems. Recently, there are some efforts to decentralize the cloud via distributing data centers in diverse geographical positions. In this paper, we elaborate on the energy consumption of different cloud architectures, from a mega-datacenter to a P2P-cloud that provides extreme decentralization in terms of datacenter size. P2P-cloud is defined as a set of commodity host machines, connected to each other to serve a community. Our evaluation results reveal the fact that the more decentralized the system is, the less energy may be consumed in the system. Studying the energy efficiency of P2P-cloud infrastructure shows that the additional system design complexity involved is warranted with improved energy-efficiency and better locality for some services. Our analysis indicates that such P2P-cloud outperforms the classic datacenter model as long as it meets the locality conditions, which are commonplace in communities. Moreover, we illustrate how much energy can be saved for MapReduce applications with a diverse range of specifications by switching to P2P-cloud.

Hubbub-Scale: Towards Reliable Elastic Scaling under Multi-tenancy

Elastic resource provisioning is used to guarantee service level objective (SLO) with reduced cost in a Cloud platform. However, performance interference in the hosting platform introduces uncertainty in the performance guarantees of provisioned services. Existing elasticity controllers are either unaware of this interference or over-provision resources to meet the SLO. In this paper, we show that assuming predictable performance of VMs to build an elasticity controller will fail if interference is not modelled. We identify and control the different sources of unpredictability and build Hubbub-Scale, an elasticity controller that is reliable in the presence of performance interference. Our evaluation with Redis and Memcached show that Hubbub-Scale efficiently conforms to the SLO requirements under scenarios where standard modelling approaches fail.

Augmenting Elasticity Controllers for Improved Accuracy

Elastic resource provisioning is used to guarantee service level objectives (SLO) at reduced cost in a Cloud platform. However, performance interference in the hosting platform introduces uncertainty in the performance guarantees of provisioned services. Existing elasticity controllers are either unaware of this interference or over-provision resources to meet the SLO. In this paper, we show that assuming predictable performance of VMs in a multi-tenant environment to scale, will result in long periods of SLO violations. We augment the elasticity controller to be aware of interference and improve the convergence time of scaling without over provisioning. We perform experiments with Memcached and compare our solution against a baseline elasticity controller that is unaware of performance interference. Our results show that augmentation can reduce SLO violations by 65% or more and also save provisioning costs compared to an interference oblivious controller.

Resource-Aware Scaling of Multi-threaded Java Applications in Multi-tenancy Scenarios

Cloud platforms are becoming more prevalent in every computational domain, particularly in e-Science. A typical scientific workload will have a long execution time or be data intensive. Providing an execution environment for these applications, which belong to different tenants, has to deal with the horizontal scaling of execution flows (i.e. threads) and an effective allocation of resources that takes into account the effective progress made by each tenant. While this is trivial for Bag-of-Tasks and embarrassingly parallel jobs, it is hard for HPC single-process multi-threaded applications because they cannot be scaled up automatically just by adding more virtual machines to execute the workload. In this paper we present MengTian, a distributed execution environment or platform capable of addressing the issues above. It encompasses several extensions to the Java execution environment, ranging from middleware to the virtual machine code and libraries. Our Java-based platform provides a Single System Image abstraction supported by a Partially Global Address Space to transparently spawn threads across a cluster of machines. It monitors progress with different levels-of-detail and accounts and restricts resource consumption. The overall goal is to redistribute resources among different JVM instances, increasing the unitary outcome of the progress vs. resource usage ratio over time.

Elastic Scaling in the Cloud: A Multi-tenant Perspective

Performance interference in the hosting platform introduces uncertainty in the performance guarantees of provisioned services. Existing elasticity controllers are either unaware of this interference or over-provision resources to meet the SLO. In this paper, we take a holistic view on elastic scaling from a multi-tenant perspective. We show that performance interference can significantly impact the accuracy of scaling and result in long periods of SLO violation. Using Memcached as a case-study, we show that making an elasticity controller interference aware can improve the accuracy of scaling decisions and significantly reduce the periods of SLO violation.

Profiling Memory Vulnerability of Big-Data Applications

Motivated by the increasing popularity of hosting in-memory big-data analytics in cloud, we present a profiling methodology that can understand how different memory subsystems, i.e., cache and memory bandwidth, are susceptible to the impact of interference from co-located applications. We first describe the design of the proposed tool and demonstrate a case study consisting of five Spark applications on real-life data set.

A monitoring system for community-lab

Community-Lab is an open and distributed infrastructure that provides a testbed for researchers to carry out experiments within wireless community networks. Community networks are an emergent model of infrastructures built with off-the-shelf communication equipment that aims to satisfy a communitys demand for Internet access and ICT services. Community-Lab consists of a set of nodes integrated into the existing community networks to give researchers access to the network and to allow them to perform experiments. The challenging environment of community networks needs a careful evaluation of experimental data to understand application behavior and spot any misbehavior or anomalies. This paper focuses on demonstrating a monitoring system tailored to meet the specific requirements of the testbed and proposes an architecture for self management to automate management. This demonstration aims to present the current status of the monitoring system, the data gathered and also invite others to experiment with the data generated by the monitoring system.