Re: Speaking of Clusters:
- From: Bill Todd <billtodd@xxxxxxxxxxxxx>
- Date: Thu, 03 Aug 2006 19:18:04 -0400
Keith Parris wrote:
Doug Phillips wrote:curiosity leads me to ask if
anyone here has experience with both Veritas and OpenVMS clusters and
how they compare?
As I understand it, the basic capability that Veritas Cluster Server provides is a cluster-wide file system, which is a prerequisite for any sort of failover capabilities within a cluster (e.g. you have to be able to see the files before you can start up an application which uses those files), and some support for application failover operations between nodes.
MC/Serviceguard did not have a cluster-wide file system before, so for them it undoubtedly seemed the logical next step.
What it does not provide is a Distributed Lock Manager, or any of the file-sharing and coordination functions within a cluster that are built on top of a DLM.
So access to disks is still on a one-node-at-a-time basis -- you'd need to scale up within an SMP box for more horsepower, since you can't scale out across multiple nodes running the same workload and sharing the same files (except presumably for Oracle RAC, which has its own DLM built in).
Well, not using a conventional, monolithic file system, anyway.
What Unix environments sometimes do in such cases is to partition the content of the file system into multiple file systems mounted at strategic common locations - so everyone sees the same 'file system' hierarchy but the sub-components of that hierarchy are spread out over multiple servers.
Now, it would certainly be possible to design a file system which did this automatically, internally, and transparently, such that even the contents of an individual file could be spread across the entire complement of cluster systems. This approach would still not require any centralized lock management (at least for file-access purposes) and could still use simple fail-over availability mechanisms (typically at the server rather than the disk level, though local disk-sharing would still be an option if inexpensive facilities were available - SAS may qualify before too long), since any given datum would be accessed at a single server which would coordinate globally-shared access to it. This is a bit reminiscent of DB2's mechanism for distributing a table across multiple clustered systems by hashing, for example.
No, it wouldn't (at least in the absence of SAS) have shared storage at the same level that VMS does, but then again it wouldn't require storage that costs nearly as much as shared VMS storage does. And while non-shared, fail-over-managed storage would concentrate loads for specific data on a single server, that server could still at least off-load some of the bandwidth by instructing its fail-over partner to serve up a portion of the data (and if that's still not enough, it's possible to distribute the copies of the data on one server across *all* the other servers so that they can help carry even the most concentrated loads, though this has availability implications unless there's also a remote mirror site - or backup copies can serve, as in the case of something like a video-on-demand server - to keep mean time to data loss under control).
I used to be a strong believer in VMS-style shared storage, but that was before commodity storage became so incredibly inexpensive and at least started to approach FC/SCSI MTBFs and features (such as command queuing) - plus for a few SATA enterprise-rated drives high-duty-cycle ratings as well. Nowadays, doubling or tripling up on direct-attached commodity storage is far less expensive than even the cheapest shared-storage array solution, plus provides superior performance to boot to heavy parallelizable loads.
So I'm now a believer in the kind of distributed, shared-nothing file server cluster that I just described, and think I've finally worked out most of the kinks in such an approach.
- bill
.
- Follow-Ups:
- Re: Speaking of Clusters:
- From: Malcolm Dunnett
- Re: Speaking of Clusters:
- References:
- Speaking of Clusters:
- From: Doug Phillips
- Re: Speaking of Clusters:
- From: Keith Parris
- Speaking of Clusters:
- Prev by Date: Re: OpenVMS, Alpha still rule roost in Intel fabs (2005-09)
- Next by Date: Re: Apache in COLPG state
- Previous by thread: Re: Speaking of Clusters:
- Next by thread: Re: Speaking of Clusters:
- Index(es):
Relevant Pages
|
