I/O and Platform Technology
While the CPU and the chipset are undoubtedly the heart of the platform, there is more to the Bensley platform than just Dempsey and Blackford. Bensley is also the first implementation of Intel’s I/O Acceleration Technology. Describing I/OAT is a bit beyond the scope of this article, but the general idea is to offload as much of the TCP/IP traffic from the CPU to the chipset and NIC as makes sense. According to Intel, an I/OAT Gigabit Ethernet controller delivers 30% more effective bandwidth than a regular controller. While this 30% improvement was probably measured on pathological code, it is quite clear that this will be an advantage for web servers, and other network intensive uses. Bensley optionally incorporates iAMT (Active Management Technology), which was described here.
Intel Performance Estimates
Bensley packs quite a few interesting features which should enhance performance: dual front side buses, a snoop filter and an enhanced FB-DIMM memory subsystem. But how does it perform? Intel was kind enough to provide the reviewers invited to their Portland workshop with quite a bit of performance estimates (in fact, my hand cramped from writing everything down). However, there are a couple of caveats. First of all, Intel’s numbers for Paxville and Dempsey are estimates and therefore subject to change. Secondly, these are estimates for fully optimized and tuned systems, so the numbers may be hard to replicate. Lastly, the AMD Opteron performance numbers are for a 2.2GHz MPU; currently AMD is shipping 2.4GHz models and will be shipping 2.6GHz parts in the near future, while Bensley is 3-5 months away (and a lot can change in a few months). With that in mind, and a pinch of salt, it is time to look at Intel’s projected numbers.
Figure 3 – SPEC_rate performance comparisons, from Intel
The SPEC_rate (base) benchmarks are CPU benchmarks that generally reflect workstation performance, rather than server performance. Of the two, SPECint_rate is more important, since integer workloads are far more common than their floating point counterparts. More info about SPEC CPU can be found at www.spec.org/cpu2000.
Figure 4 – SPECjbb2000 performance comparisons, from Intel
SPECjbb2000 is a Java server benchmark that is modeled loosely on TPC-C. It is a very good measure of server performance, but also heavily relies on the JVM used. In this case, the JVM was BEA System’s JRockit 5.0 for the Intel systems and either JRockit 5.0 or Sun’s JRE for AMD’s system. SPECjbb2000 will be retired in early January of 2006, and in my performance measurements, I used SPECjbb2005, but it is still an excellent server benchmark. Information about SPECjbb2000 is at www.spec.org/jbb2000/.