Two Socket Database Servers vs. Four Socket Database Servers

Fujitsu has submitted a new TPC-E benchmark for a two socket, Sandy Bridge-EP system that had a tpsE score of 1,871.81.  It is no real surprise that this system is using the very capable Intel Xeon E5-2690 processor, which has eight physical cores per processor and is the top of the Xeon E5 line in terms of performance and scalability. This is also the third TPC-E benchmark that has been submitted for SQL Server 2012.

The TPC-E OLTP benchmark is primarily limited by processor performance, assuming that you have enough I/O capacity in terms of IOPS and sequential throughput to properly drive the workload for the benchmark process. For official TPC-E submissions from a hardware vendor, this is a pretty safe assumption, given the cost and time required to submit test results to TPC.

SQL Server 2012 Enterprise Edition is licensed by the physical processor core instead of by the physical socket unlike previous versions of SQL Server. This means that it is even more important to carefully choose which processor to use for your database server. You want to get the best single-threaded performance possible for each physical processor core, so that you get the full value of your SQL Server 2012 core licenses. Choosing poorly when it comes your processor choice, and choosing poorly when it come to choosing between a two socket server and a four socket server can cause you to spend a lot more money on SQL Server 2012 Enterprise Edition licenses and actually get less performance.

This means that it makes sense to look at actual TPC-E scores divided by the total number of physical processor cores in the database server to get an idea about which processor will give you the most OLTP performance for each SQL Server 2012 core license.

Because of the way that Intel develops and releases processors to adhere to their “Tick-Tock” development schedule, and because of the more frequent processor revisions for the higher volume two socket space, a modern two socket server with the Intel Xeon E5-2690 processor is pretty much unbeatable for TPC-E performance from a score per core perspective.

Looking at Table 1, you can see a system with a Xeon E5-2690 having a 116.99 Score/Core rating, followed pretty closely by a previous generation Xeon X5690 with a 107.01 Score/Core. Even the older Xeon X5570 does quite well with a 102.14 Score/Core.

The best four socket system by this calculation is a Xeon X7560, coming in at a 63.97 Score/Core, followed pretty closely by a newer Xeon E7-4870 system with a 61.35 Score/Core. It seems that those two extra cores in a Westmere-EX are not giving as much as an advantage over the older Nehalem-EX as you would expect for TPC-E.

Processor tpsE Sockets Cores Score/Socket Score/Core
Xeon X5460 317.45 2 8 158.73 39.68
Xeon X5570 817.15 2 8 408.58 102.14
Opteron 6176 SE 887.38 2 24 443.69 36.97
Opteron 6282 SE 1232.84 2 32 616.42 38.53
Xeon X5690 1284.14 2 12 642.07 107.01
Xeon E7-2870 1560.7 2 20 780.35 78.04
Xeon E5-2690 1871.81 2 16 935.91 116.99
Xeon X7350 492.34 4 16 123.09 30.77
Xeon X7460 729.65 4 24 182.41 30.40
Opteron 6176 SE 1400.14 4 48 350.04 29.17
Xeon X7560 2046.96 4 32 511.74 63.97
Xeon E7-4870 2454.51 4 40 613.63 61.36

Table 1: TPC-E Performance by Processor

As I have said before, it is really hard to argue against choosing a modern, two socket database server for most OLTP workloads. The latest two socket Sandy Bridge-EP systems can support up to 384GB of RAM (with affordable 16GB DIMMs) and have PCI-E 3.0 support, with a pretty high number of PCI-E expansion slots. This should be more than enough CPU, memory, and I/O capacity for most workloads.

This entry was posted in Computer Hardware, Processors, SQL Server 2012 and tagged , . Bookmark the permalink.

5 Responses to Two Socket Database Servers vs. Four Socket Database Servers

  1. Pingback: Something for the Weekend - SQL Server Links 13/07/12

  2. kevinclosson says:

    Great post. Very few people understand that big systems are generally not fast systems. More sockets means more latency. At two sockets a process is 50% local and the remainder (on modern QPI systems) of the references falling remote suffer only a ~20% penalty. At 4, it’s 25% local and the remote tax is the same for single-hop 4S servers (e.g., EX). I’ll be surprised to see a 4S E5-4600 result because it’s not a 1-hop box as each socket has only 2 QPI links leaving the unfortunate “third-legged socket”.

    Take a peek at the E5-2643. Being 4S and top-bin frequency (and 130W TDP) it should easily top the per-core rating.

    As an aside, Oracle has always be licensed per core but for reasons I could never understand Oracle customers usually don’t go out of their way to save money–a trend I expect is seeing its sunset.

  3. John says:

    Thanks for the good information. I’d like to see the apples to apples comparisons of the same CPU. What’s the benefit of going from 1 socket to 2 sockets, or 2 sockets to 4 sockets, with the same processor. Do you have that info?

  4. I blog likewise and I’m creating a little something alike to
    this specific posting, “Two Socket Database Servers vs.
    Four Socket Database Servers | Glenn Berry’s SQL Server Performance”. Will you care in case I reallyimplement a number of your concepts? Thanks ,Christine

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s