NFSv4 Proxy in User Space on a massive cluster architecture: issues
and perspectives

Compute architectures evolved from single machines to big SMP
clusters. Further evolutions
will see larger machines designed as massive "clusters of
clusters". In this new
architecture, using NFS exported file systems is an issue: the server
won't be able to
support the huge number of clients from the compute machine.
Use of proxy servers is a classical way of leveraging and reducing
load on the
server. The implementation of several NFS proxy servers is a solution
to serve
correctly the clients from a "cluster of clusters".
CEA/DIF did develop NFS-GANESHA (https://nfs-ganesha.sourceforge.net/),
a NFSv4 server
running in user space able of managing very big caches (up to millions
of records).
It has different back-end modules and one of them turns the product
into a User Space NFSv4 proxy server.
This WiP topic will present an architecture deployement with several NFSv4
proxy servers in User Space (via NFS-GANESHA). Discussion will be made
on related issues that
appears in this scope (mostly coherency issues among the servers) and on
various solutions (including slight protocol additions to NFSv4) to manage
these issues.