Virtual packet loss (VPL) refers to a phenomenon affecting fully or partially virtualized applications. VPL can intermittently degrade application performance and often goes undetected by traditional network monitoring and application performance management (APM) tools.
VPL occurs when the hypervisor oversubscribes the number of physical CPUs with multiple virtual machines vying for the simultaneous execution of deadline-sensitive TCP code. TCP is very sensitive to variations in packet timing; if acknowledgements are not received within converged run-time bands, the sender may falsely intuit congestion and packet loss from such oversubscribed scheduling jitter, and may reduce its sending throughput to alleviate its perception of poor link quality.
It follows that although VPL creates symptoms suggesting network congestion, monitoring tools that gather data from switches and routers will find no evidence of packet loss. Moreover, VPL need not correlate with high CPU utilization; the primary cause is excessive concurrency demanded in multiple guest systems without enough physical CPUs to address the deadline request in a timely manner—a metric that can only be extracted from the hypervisor, as opposed to the guests. However, VPL can be detected by a combination of application awareness and advanced TCP analysis.
Read more about how to detect VPL by monitoring retransmission timeouts and high levels of jitter.