Reducing application-stage latencies for real-time interactive systems

Latency is a pressing problem in Virtual Reality (VR) applications. Low latencies are required for VR to reduce perceptual artifacts and cyber sickness. Additionally, latency jitter denotes the variance in the pattern of latency changes which additionally may cause unwanted effects. This paper analyzes latency jitter caused by typical inter-thread communication (ITC) techniques commonly used in todays computer systems employed for VR, the influence of the operating system scheduler, and the effect of different garbage collection (GC) methods to understand their effect on latency spikes, here for different Java Virtual Machines (JVM). We measure the scalability and latencies for various ITC techniques with an increasing number of threads and actors performing prototypical concurrent tasks. Four different benchmark implementations on a vanilla Linux kernel as well as on a real-time (RT) Linux kernel assess if a RT variant of a multiuser multiprocess operating system can prevent latency spikes and how this behavior would apply to different programming languages and ITC techniques. We confirmed that scheduler and prioritization of the VR application both play an important role and identified the impact they have on the implementation strategies. Also, Linux RT can limit the latency jitter at the cost of throughput for certain implementations. As expected, the choice of a GC method also is critical and will change the latency patterns drastically. As a result, we suggest that coarse grained concurrency should be employed to avoid adding up of scheduler latencies and unwanted latency jitter for the native ITC case, while actor systems are found to support a higher degree of concurrency granularity and a higher level of abstraction.