LibLustre How-To Guide
Liblustre A How to Guide
Introduction to liblustre
For some Lustre versions, a library version of the Lustre client file system (liblustre) is now available. Liblustre gives a user application (linked with the library) access to Lustre file systems, without needing to mount Lustre (VFS) on the client. The key goals for the library are to provide a portable mechanism to access Lustre from different POSIX-compliant operating systems, and to provide access from microkernel-based systems and from the Windows operating system.
Currently, liblustre is still under development and only works on linux (i386 & x86_64, not tested on ia64).
In this document, we will discuss how to use liblustre.
Note that liblustre is NOT a required element of Lustre. liblustre is not required or even recommended for running Lustre on Linux. Instead, you should use the Lustre (VFS) client filesystem to mount Lustre directly.
Generally speaking, liblustre implements the Lustre client filesystem in user space. The liblustre component links LNET and libsysio together to form a shared library which can be used by applications to perform file I/O.
Building Clients and Servers for liblustre
When using liblustre, servers are first built/configured in the usual way as described in the How to Guide to Lustre. By default, liblustre is built unless "./configure --disable-liblustre" is specified.
The following liblustre files are located in lustre/liblustre:
How to Use liblustre
First, you must run the Lustre servers (MDS/OSTs) without ACLs and accept connections on insecure ports. For example, add options lnet networks=tcp(eth0) accept=all to /etc/modprobe.conf, then run the script sh llmount.sh (if you do not already have a mounted Lustre filesystem to test against). This sets up a MDS, OST and client on the local machine. You should umount the Lustre client at /mnt/lustre to avoid confusion between the liblustre client and the normal VFS client.
liblustre needs to know the mount target before connecting to a Lustre server. The format is similar to the following:
- mgs_nid is the actual hostname of the MGS (or IP address if you do not have proper name resolution set up).
- profile_name is the profile name of the client mount point, also called the filesystem name in many configurations.
For additional information on mount target, refer to Lustre documentation. This can be passed to most liblustre programs via the environment variable LIBLUSTRE_MOUNT_TARGET.
There is a statically built-in liblustre test program at lustre/liblustre/tests/sanity. You can use this test to verify if liblustre is working properly:
sanity --target mgs_nid:/profile_name
How It Works
If we run a program, e.g. iozone, then we will use LD_PRELOAD to load liblustre.so first. The start function of liblustre.so will mount a lustre partition on certain directories, e.g. /mnt/lustre. Furthermore, following loaded iozone's functions calls such as open/read/write... will dynamically linked with implementations in liblustre.so, instead of libc in the usual case. Thus, we can intercept filesystem-related system calls and translate them into Lustre commands.
Necessary Environmental Variables
There is a simple script lustre/utils/lrun which set some ENVs:
- mount point
This is where liblustre mounts the remote Lustre filesystem. The default location is /mnt/lustre. Make sure it exists on your system.
- mount target
Before using liblustre, you need to fill in the actual mount target (e.g. mdshost:/mdsname/client).
- shared library
Make sure you have installed lustre/liblustre/liblustre.so in this location.
Running Programs over liblustre
Until now, only a small number of applications have been tested with liblustre:
And also several standard UNIX commands:
You need to prepend 'lrun' before the programs that you intend to run:
lrun iozone -f /mnt/lustre/ioz_testfile -r 256k -s 1g lrun mkdir /mnt/lustre/testdir lrun touch /mnt/lustre/testdir/testfile lrun cp /etc/fstab /mnt/lustre/testdir lrun ls /mnt/lustre/testdir lrun find /mnt/lustre/ lrun .....
liblustre is somewhat sensitive to the glibc version. We are using glibc-2.2.5 and glic-2.3.2.