The speaker is a physicist, not a computer scientist, and knows almost
nothing about file systems. Interaction from the audience in the
form of questions, comments and flames is desirable. Without this, due to
lack of genuine content, the talk will be dull and will probably end ahead
of schedule
My research group at the MPI for Gravitational Physics, Hannover, operates
a large computer cluster used for data analysis. In principle, the most
cost-effective and highest-bandwidth data storage available is the disks
local to the compute nodes. In the case of Atlas, there are 1680 disks
with an aggregate capacity of 840 TB. With a suitable file system the
array of these disks could form a highly reliable storage system, however
at the moment there appears to be no open-source distributed file system with
the necessary RAID-like properties. In this talk I present a list and
description of the properties that such a file-system should have, and
arguments to support the case that it should be possible to achieve this
in the real world. If such a file system were available, I believe that
thousands of large computer clusters around the world would employ and
benefit from it.