Back to files

Network technology evolves quickly, but available bandwidths are still not sufficient for fast transfers of big files to a bunch of clients (for example populating operating systems' installation images). Multicasting Master-Slave Network Filesystem (mmsnfs) is a new method of distributing entire filesystem among several client-machines simultaneously and efficiently. Moreover it provides mechanisms for modifying distibuted filesystem structure reliably and securing it against introducing inconsistencies. It is designed to be used in small networks of similarly-configured machines, for instance university labs or internet cafes.

Installation

To create and manage mmsnfs filesystems you need mmsnfs-utils package. Current modules are compatible with mmsnfs-utils version 0.7a. It's not guaranteed, that future releases will match. Utilities can be obtained from <http://www.sourceforge.net/projects/mmsnfs/>.

To install mmsnfs kernel modules you need Linux 2.6 kernel with its source tree. Make sure, that there is symlink /lib/modules/<kernel version>/build pointing to this source tree.

If you would like to build modules matching your running kernel, just run 'make' and these modules will be built _and_installed_ without any questions.

To build for some other kernel version, run 'make UNAME=<version>'.

Master and slave nodes

Refer to <http://mmsnfs.sourceforge.net/documentation/dov/node5.html>.

In short: master files and directories are read-only on client machines and are always synchronized with server. Slave files can be modified by clients, their new content is stored in cache, byt modified slave files are no more synced with server.

Implemetation design

Mmsnbfs filesystem consist of several kernel modules. One core module called 'netfs' and several cahcing policies.

'dummyc' caching policy:

        it's very simple no-op caching policy, which can be used only as
        policy for client machines. It can be used, if no caching is
        requested.

'diskc' caching policy:

        it's just a normal filesystem implementation, which can be used
        as standalone filesystem, network server filesystem, which is then
        populated among clients, or as an additional cache for clients.
        It's needed for support of modifiable slave files

Modules parameters

netfs_no_debug

turn off debugging output (recommended)

netfsd_hostid

        assign arbitrarily specified host identifier. Every host in
        mmsnfs-ed network must have unique host identifier (regardless of
        its IP). If this option is not specified, hostid is generated
        randomly when needed.

netfsd_min_reqds

minimum number of requests handling kernel threads

netfsd_max_reqds

maximum number of requests handling kernel threads

netfsd_multiaddr

        multicast address used for transferring mmsnfs protocol packets.
        If not specified, broadcast address 255.255.255.255 is used.
        Default interface is used for communication (ie. this one, out of
        which your default route points).

netfsd_localaddr

        local address to bind to. If not specified, your default IP
        matching outgoing interface is chosen.

netfsd_multiif

        interface number for outgoing multicast traffic. Look at output of
        'ip link' for appropriate number.

netfsd_multittl

        multicast packets TTL (time-to-live). Setting it to 1 ensures no
        packet will go outside your local network.

Mounting

Please refer to mmsnfs-utils package documentations for instructions regarding creating mmsnfs filesystems.

Supported mount parameters:

cache

caching policy ('diskc' or 'dummyc')

server

act as server instance.

client

        act as client instance. Specifying both server and client causes
        unspecified behaviour.

mntid

        unique mount id to serve (SVR mode) or to join (CLI mode). Can be
        omitted if you specify mntname.

mntname

user-friendly mount name. Optional.

svrid (CLI only)

DEPRECATED and no more available, is now determined automatically

liveupd (SVR only)

push all changed to clients immediately when in edit-mode.

editmode (SVR only)

        enter edit-mode at mount, enable changing filesystem. Not
        recommended, please use mmsnfs-utils.

evict (CLI only)

        enables eviction mechanisms, which should be used when cache size
        is less than total filesystem size. This enables replacement of
        old content. It's highly experimental and may cause deadlocks.
        Supported only by 'diskc' caching policy.

dontwait (CLI only)

        never wait for server when it's unreachable. Mmsnfs default
        behaviour is to forcibly wait for server when no data (even
        cached) is available for some file/directory. Passing tis option
        causes mmsnfs to fail with IO error in such cases.

svrtimeout (CLI only)

        amount of time (arbitrary units, see include/linux/netfs_fs.h,
        NETFS_POLLD_INTERVAL) after which netfs instance enters detached
        mode when no server reachable.

sleepevery (CLI only)

        force 'updater' process to sleep every this many requests
        processed.

sleeptime (CLI only)

force 'updater' to sleep this many jiffies.

noupdate (CLI only)

never run 'updater', useful with 'dummyc' caching policy.

Examples

mount -o loop,liveupd,mntid=3,server,mntname=test,cache=diskc \

-t netfs bigfs1 foo

mount -o loop,mntid=3,client,cache=diskc \

-t netfs pfs2 bar

Reporting bugs

All observed bugs/problems and feature requests please direct to me using SourceForge <http://sourceforge.net/tracker/?group_id=99790>. Don't hesitate to contact me privately on <mailto:accek@poczta.gazeta.pl> if you have problems with this software. But before doing that, please review all revealant documentation.

Please refer to REPORTING-BUGS included with your kernel tree source when reporting bugs.

Author

Szymon Acedanski
ul. Sloneczna 81D/1
40-136 Katowice, POLAND
phone: +48 506 202715
email: accek@poczta.gazeta.pl

# vim:tw=70