well, I read the academic research by Shroeder and Gibson. While it suggests (even states) there is no significant difference between disk types their conclusions are not supported by any scientific methodology or even base consistancy.
They did not collect their own results, rather they accepted information collected by different sources over widely different timescales. Of the '100,000' drives, around 20% were SATA drives but they had been in place less than 12 months, compared to various ages for other drives (5 and 8 years in two instances).
Additionally for some reason they ignored the fact that in one complete implementation, 11000 (around half of the SATA drives in the study) were replaced completely under warranty within the first year. This was due to unacceptable media error rates, but not considered to be disk failures. This is written off as a 'bad batch', and covered over by stating anecdotal evidence suggests other disk types have bad batches too.
The various applications are described briefly: The HPC apps contain mainly internal disks, with only one having external storage (the ill fated SATA storage that was replaced completely, and another FC array that wasn't replaced). The COM apps do not detail the storage at all, but since they are ISP's with huge number of servers, then this would appear to be internal storage too. There is no discription of the environmental factors, application loading, or relevant factors.
The inconsistancy in the raw data makes drawing conclusions futile, yet the authors present an impressive array of graphs, charts, and even conclusions. If you presented this to senior management they would probably be blinded by the academic gloss, accept the conclusions and base their decisions on a misleading artificial construction. This makes this paper far more dangerous than marketing material, where at least you can go back to the vendor for some accountability.
We do need an independent report on disk failures, but there needs to be controls in the data collection.
Others have already commented on the google report, which is of only slighly more use to companies that are not google.
Slashdot Mirror
well, I read the academic research by Shroeder and Gibson. While it suggests (even states) there is no significant difference between disk types their conclusions are not supported by any scientific methodology or even base consistancy.
They did not collect their own results, rather they accepted information collected by different sources over widely different timescales. Of the '100,000' drives, around 20% were SATA drives but they had been in place less than 12 months, compared to various ages for other drives (5 and 8 years in two instances).
Additionally for some reason they ignored the fact that in one complete implementation, 11000 (around half of the SATA drives in the study) were replaced completely under warranty within the first year. This was due to unacceptable media error rates, but not considered to be disk failures. This is written off as a 'bad batch', and covered over by stating anecdotal evidence suggests other disk types have bad batches too.
The various applications are described briefly: The HPC apps contain mainly internal disks, with only one having external storage (the ill fated SATA storage that was replaced completely, and another FC array that wasn't replaced). The COM apps do not detail the storage at all, but since they are ISP's with huge number of servers, then this would appear to be internal storage too. There is no discription of the environmental factors, application loading, or relevant factors.
The inconsistancy in the raw data makes drawing conclusions futile, yet the authors present an impressive array of graphs, charts, and even conclusions. If you presented this to senior management they would probably be blinded by the academic gloss, accept the conclusions and base their decisions on a misleading artificial construction. This makes this paper far more dangerous than marketing material, where at least you can go back to the vendor for some accountability.
We do need an independent report on disk failures, but there needs to be controls in the data collection.
Others have already commented on the google report, which is of only slighly more use to companies that are not google.