Oracle finally introduced its latest SPARC server line, the T4, in September. Contrary to the boasts of Oracle CEO Larry Ellison over the past several years, the new server line is no System z or System p killer. To the contrary, it might be too little too late to put a significant dent in the ongoing migration away from the Oracle/Sun SPARC hardware platform.
For sure, the new latest SPARC offers solid performance. Optimized for single-threaded workloads and for running Oracle database applications, the T4 apparently is intended for the consolidation of multiple application tiers onto a single server. Oracle claims SPARC T4 servers deliver outstanding performance against IBM POWER7 and HP Itanium-based systems. In data warehousing, Oracle reports the SPARC T4-4 server delivered over 2.4x better performance per socket than the IBM Power 780, 33% better price/performance.
Since the acquisition of Sun two years ago, Oracle has promised that it was committed to and capable of optimizing the hardware platform for its software. That hardware-software optimization—something IBM has been doing for years—would deliver the platform payoff.
With the SPARC T4 Oracle announced new versions of three SPARC T4 server-based Oracle optimized products. The products promise superior performance and value from Oracle’s fully integrated, pre-tested offerings. These include an Oracle database package, a JD Edwards EnterpriseOne package, and an Oracle WebLogic suite. For each, Oracle boasted better performance at lower cost than the systems running on IBM POWER7.
Any vendor-published benchmarks must always be suspect. There are just too many ways a vendor can bias benchmarks by tweaking the configuration, such as increasing cache or slipping in some SSD. Even the fine print that should accompany the results won’t necessarily reveal the bias. Dancingdinosaur once trained as a benchmark auditor and even then it’s easy to slip slight configuration tweaks past.
Here is a response by Conor O’Mahony, an IBMer, to one of Oracle’s many benchmarking announcements in the past couple of weeks. It shows how you have to parse these kinds of announcements. Let’s look at the announcement around Oracle’s SPARC SuperCluster and their latest SPECjEnterprise2010 benchmark result. First off, it turns out that Oracle did not actually use their SPARC SuperCluster for either that benchmark result or their TPC-H benchmark result. They assembled a system that is similar to the SuperCluster, but it is not the SuperCluster. (Instead of using the ZFS storage appliance, this benchmark uses F5100 flash arrays to store the database completely on solid state drives!!)
In that announcement, a SPECjEnterprise2010 benchmark result beat IBM’s best performance result in the same benchmark by over 2.4 times, delivered 20% higher performance per processor than IBM POWER7, and supposedly delivered 6.7 times better price/performance. Yes, Oracle does have a better benchmark result, but they also used a lot more hardware to get there. O’Mahony’s full analysis can be found here.
Oracle, for example, didn’t just use twice as many processor cores for this benchmark. They also used substantially more server memory, additional storage processors, more storage memory, more hard disk drives, and solid state drives. And they disabled integrity checking to boost performance. Would you do that in a production system?
The other thing to be aware of is how Oracle derived their pricing analysis. That analysis did not compare the total costs for these systems; instead it compared only the application server portion of the costs. In other words, they chose to not consider the cost of the database software (which includes licensing both Oracle Database and Oracle RAC on 64 processor cores). If they did add that cost, the conclusions would have been very different.
The best you can say about the SPARC T4 at this point is that Oracle SPARC is back as a serious hardware competitor. That’s actually good for System z and p shops because it will keep price pressure on IBM. It will, however, take case studies of legitimate clients running actual configurations for real workloads and paying true prices before anyone can really gauge the price/performance of the T4 servers.