Making the grid invisible

Back to index page

Previous article: Grid computing made Rapid

Next article: Bioinformatics, e-Science and the Grid: a symbiotic relationship

Historically, most people have found grids to be non-intuitive and hard to use. There is a steep learning curve, and applications must be modified to use the grid. Rather than make users adapt to the grid, we in the Genesis II project adapt the grid to the users. Our motto drives the project: ‘by default the user should not have to think’.

We set out to make grids easier to use by exploiting a paradigm already familiar to most users: the file system, with its notions of directories and files, copying, drag and drop, and so on. The basic idea is simple: map grid resources such as compute clusters, files, directory trees, security groups, and relational databases, into a grid Directory System (GDS) that behaves like a typical Windows or Linux directory. Then map the GDS into the local operating system’s file system – and voilà – users’ applications can access remote grid resources as if they were local. Applications, whether they are a binary, a shell script, or a Perl program, do not have to be changed or recompiled. No muss, no fuss, no Web Services programming and no new APIs.

OGRSH (Open GRid Shell), a part of Genesis II and developed in partnership with OMII-UK, is a software shim for Linux that intercepts application I/O calls and redirects them to a grid proxy that accesses grid resources on the user’s behalf. The result is that the GDS appears to client applications, like bash, vi, or your-application, as a mounted file system. This eliminates the need to port applications to the grid or use Web Services.

The treatment of grid file and directory resources in OGRSH is obvious, but what about compute resources such as clusters or desktops? Once again, we exploit the file system paradigm. We represent computational resources as directories. To start a job, a JSDL file that describes the job is copied or dropped into the directory that represents a computation resource. To see the jobs running, a directory listing (e.g. ls) is performed and the names of the jobs show up as directory entries. To see the status of a job, the user simply reads or cats the corresponding file in the directory. We use a similar strategy in Windows, where we have developed G-ICING, a grid-aware Windows-installable file system that maps the GDS into Windows.

The point of our work is to make grids easier to use – in fact, it is to make users unaware that the grid is even there.

Andrew Grimshaw, OGRSH.

www.tinyurl.com/6lk87o

www.tinyurl.com/5fuy6t