SESSION LEADER(S):Christopher Morrone, Brian Behlendorf
ABSTRACT: The Lustre parallel filesystem is widely used on large systems in
HPC. In order to continue scaling up Lustre filesystem at a price
that is not prohibitive, we believe it is vital to replace the
underlying ext3/4-based backend filesystem with ZFS. ZFS takes a
fundamentally different approach to disks than previous Linux
filesystems. It assumes that cheap commodity disks are unreliable,
which is a given at the scales used in HPC. It takes measures to detect
and recover from silent errors and corruption. Additionally, at HPC scales with
Lustre, even large regular IO from applications looks like random IO to
disks, making disk IOPs rates the limiting factor in filesystem performance.
Thus another beneficial feature of ZFS is its ability to sequentialize
random IO streams, reducing the number of IOPS that drives must handle.
We will discuss the current state of the ZFS port
to Linux, what work remains to be done, and other general issues of
running ZFS on Linux with large drive counts. While our main goal
of supporting Lustre only requires ZFS up through the DMU layer,
we welcome those interested in porting the ZPL (POSIX layer) and
those interested in including a FUSE based version of ZFS in the same
We are eager to form a community of users and developers of ZFS on Linux.
Session Leader Details:
Christopher Morrone (Primary Session Leader) - Lawrence Livermore National Laboratory
Brian Behlendorf (Secondary Session Leader) - Lawrence Livermore National Laboratory