[gfarm-discuss:17239] Re: Gfarm version 2.0.0 released

[Naotaka YAMAMOTO <naotaka@xxxxxxxxxxxxx>]

Dear Tatebe-san and all,

Congratulations on gfarm v2! I'm looking forward to imporove
metadata accesses. However, I wonder how dose Gfmd need memory
space on metadata node? So, I'm using Gfarm v1 with gfarm_agent
which use over 4GB (4.2GB now). My metadata server (PostgreSQL)
has about 26 milion pathname with about 70 Bytes (70 charactors of pathname).
Dose Gfmd(v2) require memory less than gfarm_agent(v1)?

thanks,

- Naotaka

> Hi Matt,
> 
> (I mailed this to gfarm-discuss ML not gfarm-announce ML.)
> 
> Gfarm v1 directly uses a postgresql (or a openldap) server as
> a metadata server.  On the other hand, Gfarm v2 uses a gfmd
> as a metadata server.  Gfmd in Gfarm v2 basically caches all
> metadata in memory, and accesses a backend database (postgresql
> or openldap) server in background.  When responding a request
> there is no disk wait.  Also, we minimuze the number of
> interactions with the metadata server in Gfarm v2.  That is why
> performance for small files is improved.
> 
> Latency of a file creation and other file system operations is
> almost the same as that of NFS exported in async mode.
> Regarding I/O bandwidth, Gfarm v2 can achieve scalable
> performance when you increase the number of clients (and the
> number of file system nodes).
> 
> Regarding a failover of the metadata server, you can set up
> a stand-by metadata server when a backend database is shared.
> Moreover, the backend database (postgresql or openldap)
> itself can be configured with some redundancy by many ways.
> 
> Thanks,
> Osamu
> 
> On Wed, 28 Nov 2007 12:32:20 +0000
> Matt <mattmoran76@xxxxxxxxx> wrote:
> 
> > Osamu,
> > 
> > Congratulations on releasing Gfarm v2.
> > 
> > Can you advise on how you have improved the handling of small files, as I
> > noticed you are still using either a ldap or PostgreSQL backend for the
> > mds.  When I was testing gfarm, I found when reaching millions of files that
> > this was my bottleneck.
> > 
> > It's possible that I didn't have the mds configured properly to grab all
> > available memory (4gb), is there specific settings for this? will a 64bit
> > version make more memory available to the mds?
> > 
> > Also, can you advise if performance has been improved, as I found results
> > using iozone disappointing compared to a standard NFS server and lustre
> > cluster.
> > 
> > Thirdly, have you any thoughts or suggestions about clustering/failing over
> > the mds server?
> > 
> > Thanks,
> > 
> > Matt
> > 
> > On 28/11/2007, Osamu Tatebe <tatebe@xxxxxxxxxxxxxxxx> wrote:
> > >
> > > Hi all,
> > >
> > > We are very pleased to announce the first public release of Gfarm file
> > > system version 2.0.0, aka Gfarm v2, which is a major new release of
> > > the Gfarm file system. It solves several security issues in Gfarm v1,
> > > and improves performance for small files dramatically.
> > >
> > >                 http://sourceforge.net/projects/gfarm/
> > >
> > > To mount the Gfarm file system, gfarm2fs is also required to be
> > > installed. Since it is a major new release, there is no compatibility
> > > in structure of a spool directory on each file system node. Easiest
> > > way to migrate files from a Gfarm v1 file system is
> > >
> > > 1. Install Gfarm v2 to a different target directory, and configure a
> > >    Gfarm v2 file system using a different spool directory and
> > >    different ports from Gfarm v1.
> > >
> > > 2. Mount Gfarm v1 file system and Gfarm v2 file system by gfarmfs and
> > >    gfarm2fs, respectively, on different mount points.
> > >
> > > 3. Copy files possibly in parallel using multiple clients.
> > >
> > > Enjoy!
> > > Osamu
>