I find this particular paper by Nightingale (Microsoft Research) quite interesting (will be presented at EUROSYS '11); it analyses data collected by the Windows Error Reporting system (crash dumps) to look at CPU, DRAM and disk failure rates.

Some of the things learnt:

• don't overclock (failure rate increases > 10x)
• brand-name machines are more reliable
• laptop computers are more reliable

Another interesting piece of work is Google's Failure Trends in a Large Disk Drive Population, FAST '07.

posted at: 16:11 | path: / | permanent link