Disk Subsystem Load Balancing : Disk Striping vs . Conventional Data Placement

The I/O subsystem is becoming a major bottleneck in an increasing number of computer systems. To provide improved I/O performance, as well as to accommodate growing storage requirements, disk subsystems are increasing in size. A major hurdle to obtaining the performance available from these large disk subsystems is load imbalance, or disk skew. Dynamic data placement, the conventional load balancing technique, is usually inadequate to deal with load imbalance because it is forced to accept atomic data sets with rapidly changing access patterns. We name this rapid uctuation oating load imbalance and distinguish it from the conventional view of load imbalance, referred to as xed load imbalance. Dynamic data placement also becomes increasingly di cult as the number of disks in the subsystem grows. Disk striping at a high granularity is suggested as a solution to oating load imbalance, the atomic data set problem and the complexity of balancing large disk subsystems. Disk striping uniformly spreads data sets across the disks in the subsystem and essentially randomizes the disk accessed by each request. This randomization e ectively handles both xed and oating load imbalance. Unlike dynamic data placement, disk striping does not become more complex as the number of disks increases. While a more optimal load balance may be possible for some very well understood and well-controlled environments, disk striping should provide signi cantly improved load balance with reduced complexity for many applications. This improvement will result in shorter response times and