active/active or active/passive ?

[MSA1000 advice]

Hi,
I've inherited a pair of rx2600 Integrity servers running OpenVMS I64 with dual MSA1000 smart array controllers (in cluster config). Currently the MSA1000 is setup as active/passive but I'm not sure this is best for my application.

My plan is to run an RDB database for loading and serving data. I had hoped to use one server to process incoming data+load the OracleRDB, and the other to serve users/apps querying the database. (hoping to utilize both servers).

Does anyone have experience with this?
I wonder if active/active will increase performance and allow both servers access to the same clustered raid.
Any ideas?

Thanks,

[Hein van den Heuvel]

You really want to go active/active.
OpenVMS has always been ready for that.

The MSA1000 I obtained for my RX2600 originally came with 4.48 and I upgraded to V7.0 without much trouble,

I would recommend to read "HP StorageWorks
1000 Modular Smart Array application note
Migrating to active/active controllers in OpenVMS environments"

AI can not readily find a link to that 2006 document online, so I attached it here.

hth,
Hein.


http://h20000.www2.hp.com/bizsupport/TechSupport/DocumentIndex.jsp?contentType=SupportManual〈=en&cc=us&docIndexId=64179&taskId=101&prodTypeId=12169&prodSeriesId=377751

[Andy Bustamante]

If I understand your situation, you have two active/active decisions. The MSA, as originally released supported two controllers in active/fail-over-ready mode. Current firmware allows active/active.

The second option is how to configure the database and two servers. OpenVMS and RDB can share a database. Do you have an OpenVMS clustering license?

[John Gillings]

Active/Active is usually the best way to go in terms of using the resources you've paid for.

However, with a lock intensive workload, you may find that the cost of locking between cluster nodes may outweigh the benefit of having the resources of an additional system. Compare your CPU utilisation. If a single system has sufficient memory, CPU and I/O bandwidth to take the whole workload, you may get more throughput running on one because all lock traffic is local. In rare cases you may get lock trees flapping between nodes (unlikely in recent versions) Use MONITOR DLOCK to see how locking behaves for your workload.

The other issue is failover. Active/Passive is potentially simpler. There are only two cases, you either lose all processes or none. It's fairly obvious how to deal with either of those cases.

If you're running Active/Active, failover can get complex. Should you kill all the surviving processes and restart everything, or just reintroduce the lost processes? Will you have sufficient resources to run the whole workload on a single system?

As ever you need to understand your application, and the business imperatives to know which way to go.

[Rafiq Ahamed K]

Using Active/Active FW (MSA V7.x) with OpenVMS 8.3 and above will benefit. OpenVMS 8.3 (+) native multipath is enabled to use the ALUA supported on Active/Active arrays, which will facilitate moving the Non optimized paths to Optimized. But you should make sure to statically load balance the load across both the controllers, HBA ports and cluster nodes.

You should be able to find many best practices guide for configuring the A/A arrays on OpenVMS.

Once you optmize the loads, it would be good to monitor the peroformance on nodes using T4 and make sure load is balanced across nodes and ports.

Rafiq

A/A FW link for MSA1000 http://h20000.www2.hp.com/bizsupport/TechSupport/SoftwareIndex.jsp?lang=en&cc=us&prodNameId=377753&prodTypeId=12169&prodSeriesId=377751&swLang=8&taskId=135&swEnvOID=2082

[Robert Gezelter]

MSA1000 advice,

Welcome to the OpenVMS ITRC Forum!

As John, Andy, Hein, and Rafiq have noted, OpenVMS clusters are generally best used in the Active/Active sense. Unlike other clustering schemes, OpenVMS clusters are shared resource clusters, not clusters composed of an active member and a standby. In particular, it is common to configure an OpenVMS cluster with more than two nodes, with one reason being the ability to lose capacity in smaller units, which increases application uptime.

That said, as John noted, there are sometimes good reasons to keep certain processes within a single member of the cluster.

Much depends on the precise nature of your application. If "loading data" represents a significant workload (using metrics, not guesses) then perhaps it is worth a dedicated member. I have seen cases both ways.

Detailed analysis of precisely what is happening is the only way to resolve the best answer to this question.

- Bob Gezelter, http://www.rlgsc.com

[Uwe Zessin]

Bob,
"active/active or active/passive" refers to the way the storage array presents the SCSI LUNs, not how you use the servers.

[Jerry Alan Braga]

We currently use openVMS 8.3 with an MSA1500 active/active on 2 clustered Alpha DS25 nodes and we also run RDB. From the disk perspective I use the multi-path features to load balance across controllers for performance reasons. As far as RDB, I only run the databases on 1 since since RDB enables alot of optimizations of memory structures, transaction control and most notably locking. This is now the suggested method of running RDB team. I suggest you subscribe the RDB mail list forum. The actual RDB engineers monitor this group. Send an email to oraclerdb@jcc.com and place subscribe in the subject

[Steve Reece_3]

I preface this by saying that I don't have RDB experience so your mileage may vary on that part.

The MSA1000 isn't the fastest SAN array in the world. It's ok, but still reliant on SCSI for the disk paths and the cache on the controllers isn't huge.

If you're using something like PerfectCache then you're probably ok to go with using a clustered approach to the application/database. If you're reliant on XFC caching then you will want to test it and verify that you don't shoot yourself in the foot by bringing the application and database up on two nodes at the same time. XFC won't cache the files that are open on two nodes at the same time for writing. You'll have to rely on the speed of the MSA and any caching that RDB does for you.

For the disks, active/active firmware has to be the way to go in order to give the best possible throughput on the controllers.

Don't forget that you need a quorum disk on a two node cluster!

Steve

[Robert Gezelter]

Uwe,

There are alternative readings of the question. One of the challenges is that the term "Active/Active" and "Active/Passive" is used both in speaking of the disk controllers and in some quarters as a characterization of cluster technology.

The question alludes to both possibilities.

The problems of overlapping terminology and acronyms. If I recall correctly, I once saw a Digital memorandum that used the same acronym (without explication) for three different meanings. It was an interesting sentence structure, to say the least. Oddly enough, except for the oddity of the acronym usage, it was even intelligible.

- Bob Gezelter, http://www.rlgsc.com

[Hein van den Heuvel]

Bob wrote>> There are alternative readings of the question.

I beg to differ. The first question "Currently the MSA1000 is setup as active/passive but I'm not sure this is best for my application."

The Active/Passive is just a controller setup choice.
Whether the OS/application can exploit Active/Active is an separate question.
With the indication that this is an Itanium, I'm assuming it is running OpenVMS 8.3 which works well with Active/Active.

The second question is whether there is a benefit of operating as a real cluster with concurrent access from more than one node.
That is heavily application dependent and one may just have to try.
But many an RDB case study has been done which may help decide without trying. Some folks clone database (tables) to get a read-only copy to report from (from multiple nodes as need be).

The third question is whether "I wonder if active/active will increase performance and allow both servers access to the same clustered raid. "
That is an ill formed and bad question suggesting that our nameless friend "MSA1000 advice" has some reading up to do to grasps the basics involved.
active/active has the POTENTIAL to incread performance, if that performance is defined by actual IO... but often the LOCKS define the performance. Active/Active versus Active/Passive should have nothing to do with whether two servers can see the same LUN (where LUN is a raid device defined on the MSA, exposed to a connection)

hth,
Hein

[John Gillings]

Sorry,

I misunderstood the question, assuming "active" and "passive" in the more general sense, rather than just about the storage controller.

Having active/passive storage configuration makes no sense for an OpenVMS cluster. That's only needed for operating systems that can't coordinate access between systems sharing the storage.

Once you've got the storage set to active/active, you then have to decide if it also makes sense for the application.

[Uwe Zessin]

> That's only needed for operating systems that can't coordinate access between systems sharing the storage.

John, that is completely wrong.

It has nothing to do with the operating system and whether it supports shared file systems. Windows and Linux work fine with active/active arrays, can do round-robin or other load balancing methods across multiple paths and at least Windows can deal with ALUA presentations if the vendor-supplied DSM (the storage-array Device Specific Module) is properly implemented.

In the past the majority of mid-range storage arrays did do an active/passive LUN presentation (without going too much into detail, but it does not necessarily mean that the send controller 'does nothing'). Examples for "active/passive" arrays are the old DIGITAL HSG60/80 or the DIGITAL/Compaq EVA-3000/5000 running firmware below V4.