Trading Capacity for Performance in Disk Arrays (Thesis)

Abstract:

During the last two decades, while disk capacity, disk bandwidth
and CPU performance have been improving at a rate of about 60%
per year following Moore's law, disk access latency has only been
improving at a rate of about 10% per year. The performance gap
between disk access latency and the rest of the computer system
has been increasing exponentially. Moreover, Redundant Array of
Independent Disks (RAID) has become a standard approach to
improving the bandwidth and capacity of disk subsystems after over
a decade of research and development. As a result, disk access
latency becomes a performance bottleneck of many I/O intensive
applications. Although several techniques, such as striping,
mirroring and data replication within a disk track, have been
proposed to reduce disk access latency by sacrificing disk
capacity, it is not clear how to systematically configure disk
arrays to reduce disk access latency and improve disk throughput
for a variety of workloads and disk characteristics.

This dissertation proposes a novel way of designing disk arrays
that can flexibly and systematically reduce disk access latency
while improving disk throughput. We call this new disk array
configuration family "SR-Array" because it considers reducing
both seek time and rotational delay in a balanced
manner. The dissertation shows, via theoretical and experimental
studies, that the SR-Array approach can indeed improve disk access
latency significantly compared with existing disk array
configuration techniques.

The dissertation makes several contributions. First, we have
developed analytical models for disk array configurations, and we
show how to use the models to guide disk array designs towards
optimal configurations by considering both disk and workload
characteristics. Second, we have proposed a robust disk head
position prediction mechanism without any hardware support and a
new algorithm to reduce both seek time and rotational delay for
the SR-Array disk configurations. Third, we have implemented a
prototype disk array system together with an accurate simulator in
a layered approach that incorporates the configuration models.
Finally, we have demonstrated that the SR-Array approach can
indeed improve disk I/O performance significantly for several I/O
intensive workloads over existing disk array configuration
techniques.