Putting a Microscope on Your Application Monitoring via ebpf

The problem with most of the monitoring systems and tools today is that they identify symptoms of the problem like a high cpu utilization but are unable to tell the cause. This leads to visible proble …

The problem with most of the monitoring systems and tools today is that they identify symptoms of the problem like a high cpu utilization but are unable to tell the cause. This leads to visible problems in conducting a proper root cause analysis of the problem. In the session we talk about a low level Linux kernel tracing mechanism called ebpf and how it can be used to perform fine grained monitoring of the systems we run. We take an example of a production database and how with very low overhead, metrics like performance of the block layer like block latency, build up of process queues or network congestion/packet drops can be detected. We also talk of generating flame graphs on the fly which can show the hot code on and off cpu. The talk also will entail some other tools like perf and how they can be used to provide microscopic view of systems and applications.