Positioning of this Thesis

A solution for resource contention in an interconnected distributed system quires strategies to avoid contentious situation in a way that the number of re-source contentions decreases to the minimum possible. As there are always cir-cumstances that resource contention takes place, the solution requires remedial strategies to handle the contentious situation. As discussed in this chapter, these strategies can be implemented in different components of resource management systems.

This dissertation proposes a solution for origin-initiated contention in Inter-Grid, as a platform that enables resource sharing between computational Grids, and meets the aforementioned requirements. InterGrid employs VM-based leases as the resource provisioning model.

The idea of InterGrid is inspired by the manner Internet Service Providers (ISPs) establish peering arrangements in the Internet [2]. Similar arrangements are established between InterGrid Gateways (IGGs) in InterGrid that facilitate resource sharing between Grids. IGGs are considered global schedulers (meta-scheduler) for each Grid. They handle resource sharing with other peer Grids based on peering agreements. Additionally, provisioning rights over resource providers (RPs) within each Grid are delegated to the IGG which enables it to schedule arriving external requests on the RPs. External requests scheduled in an RP contend with local requests of the RP to gain access to the computational resources. Therefore, RPs in InterGrid are prone to origin-initiated resource con-tention between local and external requests.

Similar to VioCluster [9] and Haizea [13], we consider VM preemption mech-anism to resolve the origin-initiated resource contention. This mechmech-anism stops the running request and frees the resources for the higher priority request. Later, the preempted request can resume its execution from the preempted point. Specif-ically, preemption is an appropriate mechanism, due to using VM technology in InterGrid and the VMs’ability to suspend, resume, and migrate without affecting the computation inside them.

We introduce the preemption of external requests in favour of local requests in InterGrid when there are not sufficient resources to serve local requests. We consider the side-effects of preempting VM-based leases in terms of the imposed overhead time, number of resource contention, and response time for external requests. Different proposed strategies in this dissertation operate within various components of the InterGrid platform and try to handle these side-effects.

With regard to the imposed overhead time for preempting VMs, we

investi-gated and modelled the overhead time based on various possible operations that can be performed on VMs. Our proposed model extends the existing model of the Haizea [17] scheduler and considers communication between the VMs of a lease.

Another extension in comparison to the model proposed by Sotomayor et al. [17]

is that we model the overhead of migrating VMs in addition to suspending and resuming.

We propose a preemption policy in the local scheduler of the RPs that is aware of resource contention between requests. The policy proactively selects a set of leases for preemption in a way that the resource contention decreases.

An idea similar to the preemption policy was investigated by Snell et al. [10], however, they have not taken into account the overhead side-effect in their deci-sion making. Indeed, they terminate the job for preemption, which reduces the imposed overhead to zero.

We propose a contention-aware scheduling in IGG, that aims at avoiding re-source contention via scheduling of arriving external requests on different rere-source providers within its domain. It is an on-line scheduling policy that proactively schedules the external requests with regard to the local workload characteristics, such as arrival rate, in each RP. The scheduling policy also considers the situ-ation that some external requests are more valuable and reduces the likelihood of contention for them. The most similar research work to this scheduling policy is the one proposed by Li [5]. Nonetheless, Li’s scheduling goal is to minimise the response time of external requests whereas our goal is to minimise the overall number of resource contentions. The other significant difference is that Li has solved the problem for the local scheduler of a Cluster whereas our scheduling policy works in the global scheduler of a Grid.

As the occurrence of origin-initiated contention is inevitable, we propose remedial strategies that react to the contentions and try to handle their side-effects, such as long response time for low priority requests. More specifically, we propose a contention-aware policy in the admission control unit of the resource providers in InterGrid. The policy reduces the impact of preemption mechanism on the response time of external requests and prevents long response time for them. The admission control policy works based on limiting the queue length for external requests in a way that their response time would be limited. According to the policy, external requests are accepted until their predicted response time is less than a threshold value.

The proposed policy is different from the one proposed by Sandholm et al. [52]

in the sense that it performs the feasibility test for each arriving request. Con-versely, our proposed policy finds out the ideal queue length and does not impose

overhead to the system for each arriving request. Although our scenario share similarities with the research undertaken by Gong et al. [7], the main difference is that we consider several parallel external requests whereas their policy handles one sequential external request that is allocated to a node. In other words, they do not consider queuing external requests, whereas our policy concentrates on finding the number of external requests that can be accepted while their response time is limited.

Recently, many research works were conducted on energy-aware manage-ment of resources in a provider. Energy managemanage-ment policies commonly utilise resource consolidation mechanism and switch off lightly loaded resources in or-der to preserve energy. However, reducing the number of active resources in a contention-prone system raises the likelihood of contentions within RPs. More importantly, an aggressive energy saving policy that switches on few resources can lead to long response time for external (low priority) requests.

We propose a contention-aware energy management policy for resource providers in InterGrid that employs consolidation to save energy while it considers the contention side-effect for external requests. Specifically, the policy tries to min-imise the energy consumption of resource providers while external requests can be served within a limited response time. Based on the classifications provided in this chapter, the proposed contention-aware energy management policy reacts to the contentious situation to handle its side-effect in terms of response time.

Additionally, this policy is applied in the resource provider level along with the local scheduler.

2.8 Summary

Focusing on interconnected distributed systems, in this chapter we introduced different possible types of resource contentions along with common approaches to resolve them. Then, we investigated the potential role of different components in a resource management system to resolve various types of contention. Particularly, we recognised the role of resource provisioning model, local scheduling, global scheduling, and admission control unit in a resource management system.

We realised that the emergence of VM-based resource provisioning model has posed the preemption mechanism as a predominant solution for different types of resource contention. Therefore, in this chapter, we also dealt with the chal-lenges and opportunities of preempting VMs. We reviewed existing approaches in Clusters, Grids, and Clouds from the contention management perspective and categorised them based on their operational model, the type of contention they

deal with, the component of resource management system involved in resolving the contention, and the provisioning model where the contention is considered.

We also presented the requirements for handling resource contention between dis-tributed systems and positioned this dissertation with regards to existing works.

In the next chapters, we describe the contention management strategies, aim-ing to provide resources to external requests without impactaim-ing the performance of providers local users.

Preemption-based Contention