Brian et al,
I'll throw a first volley at your question.
MSCS uses a 'quorum resource' to determine of the cluster 'exists' and is up and running. When one of the members is powered on it will check for a boatload of services and resources. If these services are not seen and resources are not owned by other machines, this machine assumes the cluster does not exist and begins the task of creating it. It will do things like 'take ownership' of the shared SCSI disk (the 'quorum') where all the cluster DLLs and log files reside. Thus, when other machines come online and 'look for the cluster', it exists and they go through the process of joining rather than creating the cluster. The whole idea is kind of like taking someone else's seat. ("As a matter of fact this disk _does_ have my name on it!!").
If I remember correctly one should even be careful of powering on two cluster members at the same time since they _may_ both try to own the quorum. This may have been resolved in W2K.
A few points:
FC is supported since it's (basically) SCSI protocol in the FC payload and generally a SAN implementation will have one machine mapped to one volume.
You have to ask yourself what you want on the Filer: - You can't have the quorum resource there because (in MSCS) the cluster will never be created. The first machine will come up and can't physically own the volume. So, it only tries to join a cluster (which doesn't exist yet).
- Many Cluster aware apps will want their application on either local or shared storage. For example, MS SQL 7.0 _highly_ recommends that the binaries and Master DB (the DB about the other DBs) are on direct attached storage. If these two things aren't available then you can't start the SQL service (novel concept, huh?)
- This leaves only logs and application data on the Filer - which is the meat of the app anyway - and probably what you're more intersted in.
The problem comes whe you start talking price... Most people ask why they would consider using Filers for data storge if they _have_ to purchase and configure a shared SCSI disk shelf (and associated peripherals) to use MSCS anyway.
That's my two cents.
BTW - I understand there are cluster implementations (Veritas for one) that don't use the 'quorum resource'. But you have to use their Volume manager. (Don't quote me though).
Hope this helps,
Thomas Heffron - SE DataWise, Inc.
"When you're going through hell, keep going..." - Winston Churchill
-----Original Message----- From: owner-toasters@mathworks.com [mailto:owner-toasters@mathworks.com]On Behalf Of Brian Tao Sent: Tuesday, June 19, 2001 11:28 PM To: toasters@mathworks.com Subject: Microsoft Cluster Server and Netapps
First thing I'll say is that I know almost nothing about the way Windows NT or 2000 works, and when someone tries to explain it to me, I always wonder why a simpler and more direct UNIX-like approach wasn't used. But anyway... :)
Is anyone using Netapps as a data storage backend for MSCS-aware applications? I've read the white paper on SQL Server 7 and MSCS at http://www.netapp.com/tech_library/3084.html, but do those principles apply to any clustered application? I read Microsoft's own MSCS FAQ, and although it starts off saying it needs a shared storage backend, it goes on to mention only shared SCSI devices... nothing about FibreChannel or CIFS. It also needs a shared "quorum" disk, which somehow allows MSCS to determine which servers in the cluster are down (but the FAQ doesn't say how exactly the negotiation is done).
More specifically, has anyone deployed the e-Assist suite of CRM applications on filer backends? The vendors says they use MSCS for clustering, and recommend local RAID storage for all the servers. Given that they say 35 dual-CPU servers are needed to handle 1000 users (ridiculous, but that's another story), I'm not particularly thrilled to have to maintain locally-attached storage for all of them, and I would rather take advantage of the existing pair of F740s.
Hoping someone can help me differentiate conventional wisdom ("you can't run databases on NAS") from true wisdom ("sure you can, if you set things up right") when it comes to NT/2000 clustering... thanks. -- Brian Tao (BT300, taob@risc.org) "Though this be madness, yet there is method in't"