Mining Block I/O Traces for Cache Preloading with Sparse Temporal Non-parametric Mixture of Multivariate Poisson

Existing caching strategies, in the storage do-main, though well suited to exploit short range spatio-temporal patterns, are unable to leverage long-rangemotifs for improving hitrates. Motivated by this,we investigate novel Bayesian non-parametric model-ing(BNP) techniques for count vectors, to capture longrange correlations for cache preloading, by mining BlockI/O traces. Such traces comprise of a sequence ofmemory accesses that can be aggregated into high-dimensional sparse correlated count vector sequences.While there are several state of the art BNP algo-rithms for clustering and their temporal extensions forprediction, there has been no work on exploring thesefor correlated count vectors. Our first contribution ad-dresses this gap by proposing a DP based mixture modelof Multivariate Poisson (DP-MMVP) and its temporalextension(HMM-DP-MMVP) that captures the full co-variance structure of multivariate count data. However,modeling full covariance structure for count vectors iscomputationally expensive, particularly for high dimen-sional data. Hence, we exploit sparsity in our countvectors, and as our main contribution, introduce theSparse DP mixture of multivariate Poisson(Sparse-DP-MMVP), generalizing our DP-MMVP mixture model,also leading to more efficient inference. We then discussa temporal extension to our model for cache preloading.We take the first step towards mining historicaldata, to capture long range patterns in storage traces forcache preloading. Experimentally, we show a dramaticimprovement in hitrates on benchmark traces and laythe groundwork for further research in storage domainto reduce latencies using data mining techniques tocapture long range motifs.