No abstract

Cloud computing enables on-demand access to remote computing resources. It provides dynamic scalability and elasticity with a low upfront cost. As the adoption of this computing model is rapidly growing, this increases the system complexity, since virtual machines (VMs) running on multiple virtualization layers become very difficult to monitor without interfering with their performance. In this paper, we present hypertracing, a novel method for tracing VMs by using various paravirtualization techniques, enabling efficient monitoring across virtualization boundaries. Hypertracing is a monitoring infrastructure that facilitates seamless trace sharing among host and guests. Our toolchain can detect latencies and their root causes within VMs, even for boot-up and shutdown sequences, whereas existing tools fail to handle these cases. We propose a new hypervisor optimization, for handling efficient nested paravirtualization, which allows hypertracing to be enabled in any nested environment without triggering VM exit multiplication. This is a significant improvement over current monitoring tools, with their large I/O overhead associated with activating monitoring within each virtualization layer.

Performance has become an important and difficult issue for software development and maintenance on increasingly parallel systems. To address this concern, teams of developers use tracing tools to improve the performance, or track performance related bugs. In this work, we developed an automated technique to find the root cause of performance issues, which does not require deep knowledge of the system. This approach is capable of highlighting the performance cause, using a comparative methodology on slow and fast execution runs. We applied the solution on some use cases and were able to find the specific cause of issues. Furthermore, we implemented the solution in a framework to help developers working with similar problems.