Re: [Gluster-devel] running RDBMS on top of GlusterFS

[Anand Babu Periasamy]

Marcin Krol writes:

> Hello everyone,
>
> First, thanks to developers for GlusterFS - we are considering
> transition to cluster filesystem and so far I have narrowed the
> possible choices down to only two usable options, namely GlusterFS and
> LUSTRE.
>
> I have searched archives for issues related to running SQL engine on
> top of GlusterFS, but haven't found explicit answers, so here it
> goes:
>
> 1. Are there any drawbacks to running RDBMS such as PostgreSQL or
> MySQL on client node while the data like .MYI and .MYD files are
> stored on GlusterFS brick?
>
> Conversely, are there any advantages? E.g. any thoughts about
> configuration that could parallelize reads of different sections of
> index and data files and thus boost performance of SELECT queries?

I don't see any drawbacks as such, but here are my thoughts.
I consider RDBMS and FS as parallels. RDMBS for structured data and FS
for unstructured data. Some RDBMSs assume that they are smarter than
the OS and do everything by itself (such as raw partitions). In such
cases, I recommend Infiniband SRP (SCSI RDMA Protocol) based
SAS/SATA RAID storage. 

For filesystems based RDBMS, GlusterFS can still provide benefits. 
If multiple smaller databases are served from a single server,
database files will be scattered and cached for performance by
default. For large databases, you can choose to stripe each database
file across the cluster. 

GlusterFS is a POSIX compliant file system. Databases should work
fine. How ever we have not evaluated any. I won't be surprised if they
are any glitches. If you face any issues, please write to us.

> 2. Any hard data / tests / hints about performance in comparison to
> running RDBMS engine on local fs?

We will be interested to see if you make one. We will also plan for
RDBMS benchmarks after the next release. Also I strongly recommend
Infiniband RDMA (verbs transport driver) for such needs. GigE/TCPIP is
too slow (bandwidth/latency) for such needs. 

> 3. Can you think of any problems regarding reliability, locking
> issues, etc?

Reliability is better with clustered file systems. GlusterFS supports
complete POSIX mandatory and advisory locking.

> It would be nice to add an item like this to FAQ - I'd write smth on
> the topic myself, but frankly I have very little experience with
> cluster filesystems, so it's not like I feel knowledgeable on the
> subject.

I will add them to the FAQ. Thanks :)
--
Anand Babu
GPG Key ID: 0x62E15A31
Blog [http://ab.freeshell.org]
The GNU Operating System [http://www.gnu.org]