Why Computer Systems Fail

The last years have seen dramatic changes in the way the world stores, manages, and uses its data. Whereas the end of the previous millennium saw an emphasis on decentralized, desktop-based storage and manipulation of data, we are now seeing information technology returning in great part to a model of centralized, professionallymanaged computation and data delivery services, using increasingly more ubiquitous networks. This is however not a return to the era of mainframes, but rather a segregation of the IT world into small, end-consumer devices on one hand, and the infrastructures to support and interconnect these devices on the other. One might argue this is a more natural division of functionality. This model of infrastructure-provided compute services has enabled a rich landscape of technological enhancements to our daily lives, ranging from e-mail, SMS, to e-banking, e-government, and e-voting. At the same time, the new usage models have encouraged outsourcing, have brought about large clusters of commodity hardware, and generally have pushed the providers to optimize for cost and peak performance, while initially ignoring traditional mainframe virtues like reliability and manageability. In such a service-oriented world, dependable computing infrastructures become indispensable. For example, organizations rely increasingly more on e-mail and instant messaging for communicating and making decisions on business-critical matters; Osterman [2001] found that loss of e-mail access (whether due to downtime, viruses, or other reasons) can result in up to a 35% decrease in worker productivity. Infrastructures are critical to the viability of many sectors of the economy; according to Patterson [2002], service outages can cost tens of thousands to millions of dollars per hour, with collateral damage ranging from loss of customers to drops in market valuation, as described by Fox and Patterson [2002]. As witnessed just a few years ago (Poulsen [2004]), the risks of failure