Introducing Lustre

Those of us who live in the world of open systems spend a lot of time working with NFS, the network file system that began with Sun back in the 1980s and is now widespread (if not ubiquitous) in those commercial IT environments that use Unix.  NFS allows networked computers with different operating systems – including Windows and Macintosh – to share files and disk storage over TCP/IP.

Broadly speaking, the commercial applications most of us use rely on integer math – as do the operating systems on which they sit -and we measure the power of such systems in terms of millions of instructions per second (MIPS). 

Scientific applications on the other hand, are concerned with applications that mostly use floating point math, which enables them to handle extremely long numbers with ease.  In such circumstances the capabilities of a computer to do floating point operations (FLOPS) are much more important than the ability to execute machine instructions.  Gigaflops, teraflops and perhaps even petaflops are the measures they use.

The most popular operating system among researchers these days is Linux.  In computationally intensive environments – particularly where data must be processed in parallel rather than as discrete pieces – the Linux machines are frequently clustered to work on complex problems.  Processing data in this sort of high performance environment thus demands high speed I/O of very large files. 

Because the engineering and scientific community places a different set of demands on its processors than do most commercial users, it should come as no surprise that research sites involved with high performance computing frequently have a different set of requirements for the storage devices and file systems that support their processing.  At many such sites, NFS does not fill the bill.

The research community is turning to an open source clustered file system to address its needs.  Called “Lustre,” it is object-oriented, Linux-based, Posix-compliant, and offers very high performance.  Consider for example that Lustre is a file system capable of creating 5,000 files per second in a single directory, that it offers I/O at a rate of over 600M byte/sec to individual clients (with an aggregate I/O of 11G byte/sec for the system), and that it is presently scalable to thousands of nodes.  Now also consider that Lustre runs on industry-standard hardware, can be used with direct-attach or networked storage, and is supported by such

disparate vendors as Cray, Dell and HP.

Did I mention that it is already in Release 1, and is in use at a number of major North American research sites, including Lawrence Livermore National Laboratories?

We’ll dive a little deeper next time.