Abstract
E-Business applications are subject to significant variations in workload and this can
cause exceptionally long response times for users, the timing out of client requests
and/or the dropping of connections. One solution is to host these applications in virtualised
server pools, and to dynamically reassign compute servers between pools to
meet the demands on the hosted applications. Switching servers between pools is not
without cost, and this must therefore be weighed against possible system gain.
This work is concerned with dynamic resource allocation for multi-tiered, clusterbased
web hosting environments. Dynamic resource allocation is reactive, that is, when
overloading occurs in one resource pool, servers are moved from another (quieter) pool
to meet this demand. Switching servers comes with some overhead, so it is important
to weigh up the costs of the switch against possible system gains. In this thesis we
combine the reactive behaviour of two server switching policies – the Proportional
Switching Policy (PSP) and the Bottleneck Aware Switching Policy (BSP) – with the
proactive properties of several workload forecasting models.
We evaluate the behaviour of the two switching policies and compare them against
static resource allocation under a range of reallocation intervals (the time it takes to
switch a server from one resource pool to another) and observe that larger reallocation
intervals have a negative impact on revenue. We also construct model- and simulation-based environments in which the combination of workload prediction and dynamic
server switching can be explored. Several different (but common) predictors – Last
Observation (LO), Simple Average (SA), Sample Moving Average (SMA) and Exponential
Moving Average (EMA), Low Pass Filter (LPF), and an AutoRegressive Integrated
Moving Average (ARIMA) – have been applied alongside the switching policies.
As each of the forecasting schemes has its own bias, we also develop a number of
meta-forecasting algorithms – the Active Window Model (AWM), the Voting Model
(VM), the Selective Model (SM), the Dynamic Active Window Model (DAWM), and
a method based on Workload Pattern Analysis (WPA). The schemes are tested with
real-world workload traces from several sources to ensure consistent and improved results.
We also investigate the effectiveness of these schemes on workloads containing
extreme events (e.g. flash crowds). The results show that workload forecasting can be
very effective when applied alongside dynamic resource allocation strategies.