Appendix B : Implementing ARIES on RPVM

Moving pages that are not backed up on real NVRAM to disk-based storage and ensuring that they are accessible will involve at least one I/O, if not more. As a result, the I/O performance of a BPDB that attempts to exploit NVRAM may even be worst than a similar BPDB system that does not do so if the amount of NVRAM available is less than that assumed by the former system. This performance anomaly may be due to two reasons. First, a page may be propagated to NVRAM instead of regular nonvolatile storage to avoid performing one write I/O. Doing so may incur more than one write I/O operation on the average as a result of pages being moved from NVRAM to the disk. The resultant number of I/Os incurred in this way can become greater than the number avoided by writing to the NVRAM. Second, the availability of fast NVRAM may result in pages being propagated there more frequently and freely without using special techniques that delay and group writes to reduce I/O costs. If less NVRAM is allocated, then these propagations may actually involve I/O operations that write to the disk, further increasing overall I/O costs.