Diagnosing the Efficiency of Disk Striping and Load Balancing

摘要

You would expect the striping, even of disk slices, to yield almost perfect disk I/O load balancing, would you not? However, the numbers disprove that for our environment. This could be due to the way the striping algorithms work. It is very likely that the striping algorithm for AIX is using the end disks for some sort of calculation with significant writing overhead, somewhat similar to the "outer edge, outer middle, center, inner middle, inner edge, etc." AIX philosophy, which would make the needle always write to the outer edge to service the mirror write consistency cache, etc., for all write requests. There is some documentation on these aspects. A search conducted on "NEC Research Institute CiteSeer" (http://citeseer.nj.nec.com) will result in a large number of authors, like Arif Merchant, Philip Yu, Stavros Christodoulakis, Fenia Zioga, Mustafa Uysal, Prashant J. Shenoy, and Harrick M. Vin, conducting have conducted mathematical and probabilistical modeling studies on the striping configurations. They have compared "partitioned striping" to "non-partitioned striping." Their results are theoretical and seem to conclude that the probability of having a better "access distribution" in request concurrency is favored by partitioned disks striping. In doing this, they assumed from the start, in building their mathematical models, that disk access is perfectly balanced, the very fact that we debate here.