Peter Holditch

One definition of a commodity is something that you take for granted.

I'll bet there aren't many readers out there who wake in the morning and exclaim, "Thank goodness there's air in the room to breathe!" Likewise, computer users will seldom give thanks for their operating systems, a proclamation like "praise be to those at AT&T and BSD for giving me Unix!" would likely raise more eyebrows than nods around an average water cooler. Things that are taken so much for granted are ripe for methods such as open source as provisioning mechanisms - at the end of the day, if its interfaces are well known and an implementation works "well enough" and is supported, then who cares if it was written in a garage or a high-tech lab? That's one of the reasons why I think the application server has some way to go before it commoditizes to the point where open source provides a realistic alternative as an underpinning to a business critical deployment - the application server market is small, and the capabilities offered by the market leading app. Servers are too rich (and the standards are evolving too quickly) for that to happen just yet a while.

But I digress (although I'm not sure if it's possible to digress before you have actually started...). This month's article is about a piece of technology that exists somewhere in the twilight zone between commodity and valuable asset - the humble JDBC driver.

One of the main drivers for the widespread adoption of the J2EE platform was the existence of standards that provide a good degree of plug-and-playability in terms of the resources accessing and accessed by the platform. One such principal resource is the SQL database (whose adoption in its own life cycle was fueled by the standard relational model for data access providing some swapability at a different layer of the architecture). The Java standard that makes SQL databases pluggable at this level is, of course, JDBC. At a brief glance, you may be wondering what I am going to write about next, apparently being on the verge of saying that the JDBC driver is yet another commodity, but actually, the truth is, especially in the context of XA transactions, all JDBC drivers were not created equal. My intent this month is to dig a bit deeper into this, and help answer two questions: What's the overhead of XA and what JDBC driver should I use to access my Oracle database in an XA transaction?

What's the Overhead of XA?
Quite a few years ago, when X/Open first created the XA standard, there were no commercial databases that supported it. Then came Oracle 7 (and the other RDBMS products of around the same vintage) and XA moved from nice theory to realistic deployment option and the distributed transaction was finally liberated from the mainframe to roam free across the world of open systems. In those days, it was not unreasonable advice to suggest that architects build all their services to access the database via XA. The performance overhead of XA was small (I dimly remember figures of 5% bandied about) and the 1pc optimization built into Tuxedo (and later brought forward into BEA WebLogic Server's transaction manager) meant that the 5% was all you paid, if you didn't need a distributed transaction for any particular use-case.

These days, the world has moved on, and this advice is in need of amendment - it turns out that with recent versions of Oracle, and Java in the picture, the overhead of accessing the database via XA is considerably more than 5% - some benchmarks put the degradation in throughput caused by using XA as high as 15x in the worst case (and note, this is not the overhead of two-phase commit, simply overhead introduced by the DBMS engine as a result of using xa_begin and xa_commit to delineate transactions).

Clearly, the prudent advice to anyone architecting a J2EE system is, at the very least, to benchmark your application using the XA configuration you need for production to ensure that your exact setup will give the throughput you require from your application, and be prepared to segment your database accesses into those that can use database local transactions and those that require XA for synchronisation with resources external to the database (for example, once and once only guaranteed message delivery from transactional message queues).

What JDBC Driver Should I Use to Access Oracle?
The JDBC standard provides for four different "types" of JDBC drivers. The important ones for this discussion are type 2 (in which the database is accessed via JNI calls from the java layer to Oracle's native OCI libraries) and type 4 (in which a pure Java driver implementation talks across a socket to a remote server, which accesses the database as a proxy for the client)

Looking at this from a "commodity component" perspective, the choice between the type 2 and type 4 drivers might be expected to be academic (with perhaps a slight preference for type 4, given that this doesn't require installation of the database native components on your application server box). This is by and large the case for non-XA database access; however, for XA access the story is somewhat different. In this case, the throughput of the type 2 drivers is somewhat better than that of their type 4 equivalents, but more pronounced is the effect on the database CPU - XA access via the type 4 driver causes dramatically increased CPU utilization on the database server relative to the type 2 equivalent.

So, What's the Conclusion?
As with all performance-type considerations, the main conclusion has to be that you should performance test an exact replica of your intended production configuration in order to determine that its performance will meet your needs. Any benchmark or other comparison of that kind is inherently heavily influenced by the usage patterns of the application, and also any benchmark is a snapshot in time of the state of the system - patches or new releases of the driver code (or the database engine itself) could completely change (hopefully for the better!) the results of any comparison.

However, with the technology where it is today, it seems safe to conclude that you should try to stick to the type 2 driver for XA database access, unless you have a pretty compelling reason to do otherwise.