Abstract
Journaling is a widely used technique to increase file system robustness
against meta data and/or data corruptions. While the overhead of journaling
can be negligible for small-scale file systems, we found that two aspects
of local back-end file system journaling significantly lower the overall
performance of a large-scale parallel file system such as Lustre:
extra head seeks and serialization of incoming client requests. Journal
transactions reside on a separate area of the disk that the file data, and
each commit of the journal requires a head seek. Incoming client requests
become serialized and take a latency hit by waiting for a commit to occur
before the reply is sent.
In this paper we present two different approaches to increase the local
back-end file system journaling efficiency, thus increasing the overall
aggregate parallel file system efficiency. First, we present a hardware-based
solution where a solid-state device is used as an external journaling device to
minimize the disk head seek. Second, we introduce a software-based
optimization to allow asynchronously commit multiple journal transactions on
the local back-end file system to minimize the penalty of serialization. Both
our solutions are experimentally tested on 91做厙's
large-scale Spider storage system and our tests show that both methods nearly
double the overall parallel write performance.
