As an experiment, I decided to import the current official TPC-E results spreadsheet into a SQL Server 2008 R2 database, so I could easily query and do some basic analysis against the results. I did a little data cleansing so that the data would always use the same terms for CPU model, SQL Server version, etc. I also added a few columns that are not in the original official results spreadsheet, such as the MemorySize, SpindleCount, and OriginalDatabaseSize, and added that data to each row in the table, using the information from the Executive Summary for each submitted TPC-E result. Luckily, there were only 41 TPC-E results at the time this was written.

After that effort, I was ready to write a few queries to see if anything interesting might reveal itself from the actual raw data. Since SQL Server is licensed by physical socket (when you buy a processor license), I thought it would be interesting to rank the results by TpsE per Socket, by simply dividing the TpsE score by the number of sockets. This gives you a pretty good idea of which processor does the best on that benchmark, assuming the rest of the system was properly optimized (which is probably a pretty safe assumption given the time and cost of doing an official TPC-E submission). This gives you an idea of which processor gives you the most “bang for the buck” on the TPC-E benchmark.

Looking at the results this way, the two systems at the top of the list both use the Intel Xeon X5680 processor. After that, the next seven systems use the Intel Xeon X7560 processor, and they show pretty good scaling as you go from four sockets to eight sockets (i.e. the eight socket scores are pretty close to double the four socket scores for that processor), which shows the effectiveness of NUMA. Next in line is a two socket AMD Opteron 6176 SE system, which does about 10% better than a trio of Intel Xeon X5570 systems that are about a year older than the AMD Opteron 6176 SE system. The next four systems are a mix of Intel “Nehalem” and AMD “Magny Cours” based systems that conclude the top tier of results in this query. The lower limit of this “top tier” is shown in green in the table below.

The interesting thing is how much of a drop we see with the next several systems that use the older, six core Intel Xeon X7460 (which was no slouch in its day), showing a drop of nearly 50% in the TpsE per socket score compared to the lowest system in the top tier. Even more interesting, is the difference between the four socket Intel Xeon X7460 systems and the four socket Intel Xeon X7560 systems, which shows what a huge improvement the Intel “Nehalem” is compared to the older Intel “Dunnington”. You can also see that the newer two socket Intel systems (X5680 and X5570) do much better than the older four socket X7460 systems, in both absolute and per socket terms. The two listed sixteen socket Intel Xeon X7460 systems do quite poorly, showing the weakness of the old shared front-side bus architecture in older Intel Xeon processors.

My takeaway from this relatively simple analysis is that any system with processors older than the Intel “Nehalem” or the AMD “Magny Cours” (Intel 55xx, Intel 75xx or AMD 61xx) will be pretty severely handicapped compared to a system with a “modern” processor. This is especially evident with older Intel four socket systems that use Xeon 74xx or older processors, which are easily eclipsed by two socket systems with Intel Xeon X5570 or X5680 processors.  This analysis also shows how well the 32nm Intel Xeon X5680 “Westmere-EP” processor does on the TPC-E workload, which in my mind, also shows that it is an excellent processor for OLTP workloads in general.