On the Impact of Reliable Protocols on Run-Time Software System Performance Predictability

This paper presents the results of OMNet++ based simulations that explore the statistical effects reliable protocols have on software system run-time performance predictability. Real-world relevance is addressed by focusing the behaviors generated via the simulation of a real-world industry-held in-production Large-scale Distributed Software System (LDSS). The impacts of reliable and non-reliable protocols contrasted through simulating the same LDSS under statistically identical service workloads and deployment regimes. In this manner, it is shown that as event recovery actions of reliable protocols are triggered, there is a substantial decrease in the LDSS's runtime performance predictability. This somewhat counter-intuitive result arises due to reliable protocols inherently generating additional workloads to be serviced into LDSS systems already experiencing the onsets of overload conditions. The presented work qualifies these impacts while highlighting easily measurable production indicators of when performance predictability has been lost.