Abstract

In the last year, cloud computing has become the subject of much hype. In particular, the perfect elasticity of supply means that it costs the same to lease ten virtual machines for one hour as it does to lease one virtual machine for ten hours. This bodes well for scaling large data-parallel jobs, but many researchers have found that the heterogeneity of cloud resources has made it challenging to extract the expected performance.

One solution to this problem has been to rewrite application logic so that it performs better in heterogeneous environments. However, this same heterogeneity makes it difficult to perform repeatable performance measurements, which are necessary to validate these modifications. In this talk, I will present some early ideas about how to capture the behaviour of cloud resources, and use this both to perform repeatable experiments, and to lower the cost of cloud-based experimentation.