Resource allocation algorithms for virtualized service hosting platforms

Abstract

13/01/15 meb. Accepted version, Ok to add

Capacity Management as a Service for Enterprise Standard Software

“…Following our argumentation, we further investigated the listed publications with respect to the evaluation of QoS aspects and summarize the results in Table 1. [21] dynamic no spare capacity Rolia2005 [22] dynamic yes spare capacity Bichler2006 [23] dynamic no not explicitly considered Cherkasova2006 [41] dynamic yes spare capacity Stillwell2010 [19] static no not explicitly considered Xu2010 [66] static no not explicitly considered Speitkamp2010 [14] dynamic no quantiles of historical service demands Feller2011 [43] dynamic no not explicitly considered Gao2013 [67] static no not explicitly considered At first, Table 1 distinguishes between static and dynamic workload. If the optimization problem implicates a time dimension for service capacity demands, we consider the approach to be dynamic, following the classification by [43].…”

“…In their allocation problem, they maximize the minimum yield over all services to consider performance and fairness. However, actual response time estimations were not quantified [19]. In the same year, Speitkamp et al state that workloads which exceed the available capacity may lead to increased response times, but it remains unclear to which extent response times are affected [14].…”

“…However, solution feasibility is usually evaluated from a strong technical perspective, e.g., the fitness function of a genetic algorithm rates a solution candidate's feasibility with respect to capacity savings. As stated by Stillwell et al, "in practice, however, resource management objectives are expressed based on higher level metrics that are related to user concerns, such as service response time or throughput" [19]. Likewise, service level agreements (SLA) are typically expressed from a user perspective to ensure measurability.…”

“…Likewise, service level agreements (SLA) are typically expressed from a user perspective to ensure measurability. Server consolidation approaches consider those higher level metrics only indirectly by mapping them to resource capacity values in a reasonable way [19], e.g., a certain amount of CPU capacity allows for the execution of a certain number of business transactions per hour [20]. Yet this high level abstraction does not allow to qualify response times or other related user-level metrics and strongly limits solution evaluation with respect to given SLAs.…”

Vector repacking algorithms for power-aware computing

“…For practical vector packing in 1-dimensions, it is well known that FF and Best-fit (BF) [7] and their off-line counterparts, FFD and BFD, strike the best balance between their time complexity and performance in terms of number of bins. Stillwell et al [14] showed FFDSum to be a good choice for resource allocation algorithms for virtualized service platforms. Generalizing BF and BFD for vector packing can be done in many ways.…”