Greg Perry wrote:
I have been researching GFS for a few days, and I have some questions
that hopefully some seasoned users of GFS may be able to answer.
I am working on the design of a linux cluster that needs to be
scalable, it will be primarily an RDBMS-driven data warehouse used
for data mining and content indexing. In an ideal world, we would be
able to start with a small (say 4 node) cluster, then add machines
(and storage) as the various RDBMS' grow in size (as well as the use
virtual IPs for load balancing across multiple lighttpd instances.
All machines on the node need to be able to talk to the same volume
of information, and GFS (in theory at least) would be used to
aggregate the drives from each machine into that huge shared logical
volume).
With that being said, here are some questions:
1) What is the preference on the RDBMS, will MySQL 5.x work and are
there any locking issues to consider? What would the best open source
RDBMS be (MySQL vs. Postgresql etc)
Someone more qualified than me will have to answer that question.
2) If there was a 10 machine cluster, each with a 300GB SATA drive,
can you use GFS to aggregate all 10 drives into one big logical 3000GB
volume? Would that scenario work similar to a RAID array? If one or
two nodes fail, but the GFS quorum is maintained, can those nodes be
replaced and repopulated just like a RAID-5 array? If this scenario
is possible, how difficult is it to "grow" the shared logical volume
by adding additional nodes (say I had two more machines each with a
300GB SATA drive)?
GFS doesn't work that way. GFS is just a fancy filesystem. It takes
an already shared volume and allows all of the nodes to access it at
the same time.
3) How stable is GFS currently, and is it used in many production
environments?
It seems to be stable for me, but we are still in testing mode at the
moment.
4) How stable is the FC5 version, and does it include all of the
configuration utilities in the RH Enterprise Cluster version? (the
idea would be to prove the point on FC5, then migrate to RH
Enterprise).
Haven't used that one.
5) Would CentOS be preferred over FC5 for the initial
proof of concept and early adoption?
If your eventual platform is RHEL, then CentOS would make more sense
for a testing platform since it is almost identical to RHEL. Fedora
can be less stable and may introduce some issues that you wouldn't have
with RHEL. On the other hand, RHEL may have some problems that don't
appear on Fedora because of updated packages.
If you want bleeding edge, use Fedora.
If you want stability, use CentOS or RHEL.
6) Are there any restrictions or performance advantages of using all
drives with the same geometry, or can you mix and match different size
drives and just add to the aggregate volume size?
As I said earlier, GFS does not do the aggregation.
What you get with GFS is the ability to share an already networked
storage volume. You can use iSCSI, AoE, GNBD, or others to connect
the storage to all of the cluster nodes. Then you format the volume
with GFS so that it can be used with all of the nodes.
I believe there is a project for the aggregate filesystem that you are
looking for, but as far as I know, it is still beta.
