Extensões ao simulador cloudsim baseado no modelo de sobrecarga por acoplamento parasítico de primeira ordem para estudar o escalonamento de CPU em sistemas cloud

Abstract

Parallel CPU performance scaling is dependent of many factors such applications behavior, cache available, efficiency of cache use. Computer simulators are used to predict or benchmark the performance of multi-core processors; they are categorized as cycle accurate or high-level models. The cycle accurate models are based on queue theory to step-by-step conduct the simulation to perform the operations needed to consume a model that represents the application size. The high-level models are used to collect, usually via additional hardware, not accessible information that represents the behavior of the running system. In those two cases, they are complex because a very long sequence of queues (cycle accurate) controls the process, making the simulation very time consuming. And in the case of high-level, sometimes it needs a complete different hardware from the actual system to read the needed states. In addition, sometimes programmable logic systems are used to implement the cache controls, queues and communication protocols what enlarges the running time, to accommodate the clock restrictions of the additional hardware. Based on a simple model proposed by Kandalintsev and Lo Cigno, the Behavioral First Order Performance Model (BFO), all the complexity of the internals of the architecture is studied as parasitic interference of one CPU/core fighting for shared resources. They introduced a coupling factor, which incorporates the essence of the observed performance behavior in a system. In this work, we used the BFO model to implement a fast cycle-accurate simulator capable of handling even large systems. To do it the CloudSim simulator was extended to behave as a cycle-accurate simulator. Basically we implemented CPU and application models that can handle the complexity of the different types of scenarios (distinct CPU behavior when running different kinds of applications). Using off line measurements we trained the model to extract the coupling factor needed to feed the simulator. We tested the pairwise behavior to make sure that the implementation could reproduce the measured tests. Our simulations consider up to 14 cores (in case of Xeon E5-2880v4) and to improve the results quality we introduced a correction factor to minimize the (seemly exponential) error observed, with good results.