【摘要】：當(dāng)前,大規(guī)模存儲(chǔ)系統(tǒng)普遍采用副本、糾刪碼等技術(shù)來(lái)提供高可靠性。隨著存儲(chǔ)系統(tǒng)規(guī)模和復(fù)雜性的不斷增長(zhǎng),傳統(tǒng)的冗余機(jī)制難以提供足夠的可靠性,構(gòu)建高可靠性的存儲(chǔ)系統(tǒng)成為了巨大的挑戰(zhàn)。目前,幾乎所有的硬盤都支持SMART故障預(yù)測(cè)技術(shù)。SMART全稱為" Self-Monitoring, Analysis and Reporting Technology",它會(huì)監(jiān)控硬盤內(nèi)部的一些與健康狀況相關(guān)的屬性,如果某個(gè)屬性值超過(guò)了閾值,它就會(huì)發(fā)出故障預(yù)警信息。然而,SMART技術(shù)的預(yù)測(cè)準(zhǔn)確率非常有限,僅能在0.1%的誤報(bào)率的情況下預(yù)測(cè)出3-10%的硬盤故障。一些研究者研究了基于硬盤的SMART信息,采用統(tǒng)計(jì)學(xué)和機(jī)器學(xué)習(xí)的方法建立硬盤故障預(yù)測(cè)模型。然而,這些模型僅能夠在保持較低誤報(bào)率的情況下,預(yù)測(cè)出60%左右的硬盤故障。 本文嘗試了改進(jìn)和優(yōu)化基于支持向量機(jī)(Support Vector Machine, S VM)的硬盤故障預(yù)測(cè)模型。本文還提出了基于反向傳播(Backpropagation, BP)算法的人工神經(jīng)網(wǎng)絡(luò)(Artificial Neural Network, ANN)來(lái)建立預(yù)測(cè)模型,并且對(duì)該BP-ANN預(yù)測(cè)模型用AdaBoost算法進(jìn)行了優(yōu)化。實(shí)驗(yàn)數(shù)據(jù)集來(lái)自實(shí)際的數(shù)據(jù)中心,包含了多達(dá)23,395塊硬盤的SMART記錄。本文采用了新的樣本處理、選取方法以及特征構(gòu)造的方法,提高了模型的預(yù)測(cè)準(zhǔn)確率。本文還提出了一種基于投票的故障檢測(cè)算法,該方法能夠有效降低模型的故障誤報(bào)率。為了更加準(zhǔn)確地描述硬盤的健康程度(即故障概率),本文也嘗試了使用BP算法建立硬盤的健康度模型。實(shí)驗(yàn)結(jié)果顯示,本文提出的模型取得了非常好的故障預(yù)測(cè)性能。優(yōu)化的SVM模型能夠取得最低的誤報(bào)率(0.03%),而B(niǎo)P-ANN模型則能在保持較低誤報(bào)率時(shí)達(dá)到95%以上的故障檢測(cè)率。本文利用馬爾可夫模型計(jì)算了不同存儲(chǔ)系統(tǒng)的可靠性,結(jié)果表明硬盤故障預(yù)測(cè)模型可以顯著提高存儲(chǔ)系統(tǒng)的可靠性。此外,本文還初步探索了主動(dòng)容錯(cuò)機(jī)制在實(shí)際大規(guī)模存儲(chǔ)系統(tǒng)中應(yīng)用的問(wèn)題。
[Abstract]:At present, large-scale storage systems generally use replica, erasure code and other technologies to provide high reliability. With the increasing scale and complexity of storage system, the traditional redundancy mechanism is difficult to provide enough reliability, so it has become a great challenge to build a high reliability storage system. At present, almost all hard drives support SMART fault prediction technology. smart, fully known as "Self-Monitoring, Analysis and Reporting Technology," monitors some of the health-related properties inside the hard disk if a property value exceeds the threshold. It sends out a fault warning message. However, the prediction accuracy of SMART technology is very limited, and only 3 鈮，

本文編號(hào)：2486636