Biblio

The evolution of cloud-computing imposes many challenges on performance testing and requires not only a different approach and methodology of performance evaluation and analysis, but also specialized tools and frameworks to support such work. In traditional performance testing, typically a single workload was run against a static test configuration. The main metrics derived from such experiments included throughput, response times, and system utilization at steady-state. While this may have been sufficient in the past, where in many cases a single application was run on dedicated hardware, this approach is no longer suitable for cloud-based deployments. Whether private or public cloud, such environments typically host a variety of applications on distributed shared hardware resources, simultaneously accessed by a large number of tenants running heterogeneous workloads. The number of tenants as well as their activity and resource needs dynamically change over time, and the cloud infrastructure reacts to this by reallocating existing or provisioning new resources. Besides metrics such as the number of tenants and overall resource utilization, performance testing in the cloud must be able to answer many more questions: How is the quality of service of a tenant impacted by the constantly changing activity of other tenants? How long does it take the cloud infrastructure to react to changes in demand, and what is the effect on tenants while it does so? How well are service level agreements met? What is the resource consumption of individual tenants? How can global performance metrics on application- and system-level in a distributed system be correlated to an individual tenant's perceived performance? In this paper we present CloudPerf, a performance test framework specifically designed for distributed and dynamic multi-tenant environments, capable of answering all of the above questions, and more. CloudPerf consists of a distributed harness, a protocol-independent load generator and workload modeling framework, an extensible statistics framework with live-monitoring and post-analysis tools, interfaces for cloud deployment operations, and a rich set of both low-level as well as high-level workloads from different domains.