I would create a web server farm with lvs (www.linuxvirtualserver.org) under linux (logically). This should be a HA load balanced enviroment for shared hosting services. The main problem is about storage. Supposing that it grows, how can I "add" disk space to the entire cluster? Is there any filesystem that support that feature? I know PVFS (Parallel virtual File system), but it hasn't got fault tolerance, and CODA uses server-client infrastructure, and I want a decentralized filesystem. :confused:

Finally, I have the same problem for database load balancing, replication and extension (in case it gets small). :bawling:

Any suggestions?

THX.

PD: Sorry if my english hurts :P

Is it possible that nobody knows nothing about this?
I am simply asking for a decentralized, fault-tolerant distributed file system or similar, and a way of scale mysql databases!! :P

Again, sorry for my English

Use openafs cluster for a decentralized distributed filesystem. . As for mysql, you can setup multiple slave servers to do reads from. As of right now, mysql has no way of doing locking between master and slave so all writes must be on the master.

I'll take a look to OpenAFS, but i think it uses server-client structure.

The mysql solution is not the right one, as it provides load balancing and HA, it doesn't allow horizontal scalability.

NFS :)

Originally posted by Papipo
Is it possible that nobody knows nothing about this?
I am simply asking for a decentralized, fault-tolerant distributed file system or similar, and a way of scale mysql databases!! :P


Ofcourse people about it... but it's not something you "simply" do. It's a quite complex setup you're looking for.

To get such a file system setup without a single point of failure (like with NFS, AFS, etc.) -- I would probably go for an install with GFS -- or perhaps OpenGFS.

Remember that you will need a shared storage media, like a shared SCSI disk for this to work. You can use iSCSI or HyperSCSI as well, in that case you will not need a shared SCSI disk in the traditional sense. However be aware that HyperSCSI does not yet support SMP.

I am not quite sure that OpenGFS or GFS can handle VERY large amount of IO operations (for example SMTP, mail servers), but for hosting it may be a good choice.

A.

Originally posted by apollo
I am not quite sure that OpenGFS or GFS can handle VERY large amount of IO operations (for example SMTP, mail servers), but for hosting it may be a good choice.

A.

What do you consider "very large"?

And why do you think GFS can't handle it?

GFS handles 2 gbps Fibre Channel, so standard SCSI or IDE disks used in normal mail servers should be absolute no problem. GFS is optimized for Oracle9, so storage of large amount of data should be no problem either.

we had a large project coming up with very heavy IO, we had SAN fibre storage over 1GB fiber... they (GFS sellers), never said it will handle the volume we required....

but who knows.... you can always order and try out :) by the way, NFS is your friend. Old. Stable. It simply delivers.

Originally posted by apollo
we had a large project coming up with very heavy IO, we had SAN fibre storage over 1GB fiber... they (GFS sellers), never said it will handle the volume we required....

but who knows.... you can always order and try out :) by the way, NFS is your friend. Old. Stable. It simply delivers.

Then you're probably not going to run the odd mail server on the system...

That kind of I/O is probably not what the requester expects to have on his system.

By the way: Which GFS seller did you talk to? (Sistina?)

There's a world of difference between GFS and NFS. You can't really compare them like that. They have very different architectures, intents, semantics, etc.

nfs is severely broken on linux. try doing this: open a bunch of files one the client on the nfs mount with fcntl locks. crash client. bring client up. now try opening the files with fcntl locks on an nfs mount from a diff client. boom, files still locked. lockd in 2.4.x does not handle SM_NOTIFY correctly. damn olaf =]

there is little to no way to scale mysql without the application's cooperation. your best bet would be to reqrite the mysql client libraries to perform write queries against the master and the read queries against the slaves.

paul