Enhanced Userspace and In-Kernel Trace Filtering for Production Systems

Abstract

Trace tools like LTTng have a very low impact on the traced software as compared with traditional debuggers. However, for long runs, in resource constrained and high throughput environments, such as embedded network switching nodes and production servers, the collective tracing impact on the target software adds up considerably. The overhead is not just in terms of execution time but also in terms of the huge amount of data to be stored, processed and analyzed offline. This paper presents a novel way of dealing with such huge trace data generation by introducing a Just-In-Time (JIT) filter based tracing system, for sieving through the flood of high frequency events, and recording only those that are relevant, when a specific condition is met. With a tiny filtering cost, the user can filter out most events and focus only on the events of interest. We show that in certain scenarios, the JIT compiled filters prove to be three times more effective than similar interpreted filters. We also show that with the increasing number of filter predicates and context variables, the benefits of JIT compilation increase with some JIT compiled filters being even three times faster than their interpreted counterparts. We further present a new architecture, using our filtering system, which can enable co-operative tracing between kernel and process tracing VMs (virtual machines) that share data efficiently. We demonstrate its use through a tracing scenario where the user can dynamically specify syscall latency through the userspace tracing VM whose effect is reflected in tracing decisions made by the kernel tracing VM. We compare the data access performance on our shared memory system and show an almost 100 times improvement over traditional data sharing for co-operative tracing.