| |
For over two decades, the interface between file systems and storage
systems has consisted of a simple, linear array of blocks that can be
read and written. This abstraction has enabled innovations in each of
the surrounding layers, but it also obscures implementation details that
may compromise the performance and reliability of the system as a whole.
In this talk I will discuss two pieces of work that address this
limitation in the storage interface.
First, I will describe Shear, a software system that automatically
discovers
the important RAID properties hidden by the block interface. This
information can be used to improve performance by tuning the file system
to particular RAID characteristics. Shear can also be used in a
management context to verify RAID construction and monitor the operation
of an array. Second, I will describe journal-guided resynchronization,
which uses file system knowledge and an expanded interface to improve the
process of software RAID crash recovery. This collaborative approach
reduces recovery time by several orders of magnitude, improving
reliability and availability while suffering little performance loss
during normal operation.
|
