DSXchange

Hello,

I have multiple network interfaces and would like to configure different interfaces for each node in the APT_CONFIG_FILE.

How do I configure for multiple network interfaces on SMP?

When I change the fastname to one of the other interfaces, jobs abort with timeout messages.

Thanks

Welcome aboard. :D

I'm not sure what you mean by "network interfaces" in this context. SMP to me implies one machine with multiple CPUs. If you want to extend processing to multiple machines, then you're effectively talking MPP.

Read the chapter on configuration files in the Manager manual. In it you will learn that a processing node's fastname is the name by which the machine on which processing is to occur is known on the fastest network connection to that machine. In a (single machine) SMP configuration file, every processing node has that machine's name as its fastname.

I'm not sure what you mean by "network interfaces" in this context. SMP to me implies one machine with multiple CPUs. A single machine can have multiple network interfaces. For example, you configure many database systems to use multiple network interfaces on one physical node. You do this to load balance network IO across the network interfaces for very network IO intensive applications.

Now, I have four network interfaces in ONE physical server. I want to load balance NETWORK IO across those network interfaces. This is basic unix stuff. Network IO load balancing. Can it be done! From documentation, I would say NO. Thus, I think this will need to become an enhancement request.

Thanks.

The PX configuration file is about "processing nodes", which are logical entities, but which each refer to some processing capacity and associated resources. You may investigate distributing resources onto mounted file systems, and you can employ specific resources (for example the db2 resource) onto specific processing nodes, through which you may be able to accomplish your ends.

Thanks for hijacking the thread and not answering my question.

I don't think Ray quite understood your original issue. I believe you've answered your own question anyway and just needed some confirmation. I believe you are correct, there is no configuration ability within the config file. I believe PX is network ignorant.

DataStage PX doesn't know about network interfaces. What you can do is ensure that when trying to balance activity across multiple networks that you use the appropriate IP so that the OS and/or application layer chooses the interface to go through. Most often this is done when configuring your databases in PX or Server and you can control through which connection data will flow when using that connection. But the developer needs to do any balancing themselves, there is no facility within DataStage that can monitor and control this activity.

I think that this type of tuning is solely within the realm of the OS and those databases that explicitly support multiple links. DataStage itself knows only about sources & targets as file system objects or virtual connections, so there really is no way of implementing anything at the ETL side of things.

I would be happy with the ability to just round robin the interfaces by choosing the appropriate .apt file. However, I get an error saying that software is not running on that network. Even though all the networks are for the same server.

The error is coming because of what Ray and Ken have been posting - the nodes in the configuration file tell DataStage about virtual systems, not actual remote system CPUs (which would then be MPP).

To go back to your statement that you want PX to 'round robin' through the interfaces. If you connect to a database (we'll use Oracle) you give it a DSN which in turn implies an IP address. The OS layer is responsible on deciding what path a give address needs to take, so even if DS could know about the networks the OS would decide what-goes-where. Now if you configured the same instance with two separate addresses, going through different subnets and thereby hopefully through different controllers you could end up achieving a bit of balancing. I've done this before (going to an IBM mainframe and using 2 different subnets); but PX had no idea which packet went where.

With Oracle (tnsname.ora) and Sybase (interfaces) on Unix I can configure what specific networks the database will start listeners for. I can configure the client software for Oracle (tnsname.ora) and Sybase (sql.ini) to use a specific network interface. That is the capability I would like.

Ultramundane,

This is what I have talking about in my previous posts. By declaring different logical databases in PX I go through different subnets to listeners on the same box and same database. It would be nice to see PX automatically detect each physical interface and see which paths to which listeners go the quickest and dynamically balance that load while running, but I don't see that IBM will consider this part of the package. Still, it costs nothing to submit an enhancement request and might get some people in engineering thinking about the possibilities. Heck, they'll probably write it and then upsell it as a "Px Extreme" product.

Good post.

Thanks.

So why can't you just call those other network interfaces in your config file? PX can handle it....

That is what I tried track_star. It gives error messages about PX not running for that interface and the jobs timeout. It may work for MPP and extending the PX environment, but it does not for SMP. If it does work for SMP, I don't know how to extend as the documentation does not explain this functionality.

Track_Star,

I didn't respond when you posted your comment, since this whole thread has been about why DataStage doesn't let you define specific hardware interfaces - I thought we had this subject discussed to a conclusion and then you so succintly stated that "PX can handle it".

Perhaps you could post a PX config file that illustrates your assertion about PX being able to define specific network interfaces per node and we will all profit greatly from that.