Storage device performance prediction with CART models

Abstract

Acknowledgements: We thank the members and companies of the PDL Consortium (including EMC, Hewlett-Packard, Hitachi, Hitachi Global Storage Technologies, IBM, Intel, LSI Logic, Microsoft, Network Appliance, Oracle, Panasas, Seagate, Sun, and Veritas) for their interest, insights, feedback, and support. We thank IBM for partly funding this work through a CAS student fellowship and a faculty partnership award.

A hybrid approach based on regression tree and radial-based function network for dynamic storage device performance prediction

“…Caching effect is an important measure can make a significant difference on prediction accuracy, which was missing in the feature vector designed by Wang et al [9]. Considering that hitting in the buffer cache is basically determined by temporal locality of accessed blocks [12], we propose an approximate LRU stack to efficiently track recency of requested blocks and use it as a measure in the feature vector.…”

“…In contrast, the so-called black-box method treats the storage system as a black box without knowing the internal algorithms or components and can accommodate workloads of different characteristics. In this method, the training data sets, which contain quantified description of characteristics of input I/O requests and their corresponding response times from the system, are recorded in a table [8] and fed into a machine learning model [9], [10], or a statistic model [11].…”

“…Wang et al [9], [10] proposed regression tree (CART) [13] to model storage device performance because it is not only easy to fit for different workloads and has good interpretability, but also provides relative good approximations to highly nonlinear mappings. However, the regression tree model has its critical drawbacks-it is too coarse for approximating the highly nonlinear mappings [18] and cannot learn incrementally for on-line applications in storage systems.…”

Storage System Optimization in Data-Intensive Environment

“…In addition, some researchers using Petri nets model performance study. Wang [31] of Carnegie Mellon University present a CART models, it treats the storage system as a black box, and get the performance of system by the machine learning method. In the early years of the study and evaluating performance, people measure performance in view of specific components or technical provinces, such as disk array, server, SCSI bus, RAID, etc.…”

Storage Performance Optimization Based on ARIMA

“…Such as, SMART(Storage Management Analytics and Reasoning Technology), the storage management project of IBM Almaden Research Center; HP's AUTORAID [4], and so on. They mainly use rule-based, model-based or feedback-approach-based to realize decisionmaking and automation of storage systems [5], [6].…”