SQL Server 2012 : Demystifying Hardware - Processor Vendor Selection

The critical first question is whether you want an Intel processor or an AMD processor for your database server. Unfortunately, it is very hard to make a viable case for choosing an AMD processor-based server for SQL Server 2012, for two main reasons. The first reason is performance. The cold, hard fact is that AMD has simply been unable to compete with Intel from a single-threaded performance perspective since the introduction of the Intel Nehalem microarchitecture in 2008. This gap has only increased over the past several years with the introduction of the Westmere, Sandy Bridge, and Ivy Bridge processors. The second reason is the licensing cost for SQL Server 2012 Enterprise Edition. AMD processors have higher physical core counts in their processors compared to Intel, and they provide lower performance per physical core. This forces you to pay for more SQL Server 2012 core licenses but get lower single-threaded performance, which is not a very good combination.

Because SQL Server 2012 Enterprise Edition is licensed by physical core, this makes it much more expensive to use a relatively poorly performing AMD processor for SQL Server 2012. One argument in favor of AMD is that their high-end processors are significantly less expensive than the high-end Intel models. If your primary consideration is getting the absolute lowest hardware cost, regardless of the effect on performance or scalability, then you should be considering a low core count, AMD processor-based system. In fairness to AMD, many typical SQL Server workloads would run perfectly fine on a modern AMD system; therefore, if low hardware cost is your first priority, you can buy an AMD server with a low core count processor to save some money.

1. Intel Processors

Until the introduction of the Intel Xeon E7 processor family in 2011 and the Intel Xeon E5 processor family in 2012, Intel had different processor families for different socket count servers. For example, the Intel Xeon 3xxx family was for single-socket servers, the Intel Xeon 5xxx family was for two-socket servers, and the Intel Xeon 7xxx family was for four-socket (or more) servers. Now you can get an Intel Xeon E5 family processor for a one-, two-, or four-socket server. You can choose a Xeon E5–2400 series processor for a one- or two-socket server, a Xeon E5–2600 series processor for a two-socket server, or a Xeon E5–4600 series processor for a four-socket server. You can also get an Intel Xeon E7 family processor for a two-, four-, or eight-socket server. You can choose a Xeon E7–2800 series processor for a two-socket server, a Xeon E7–4800 series processor for a four-socket server, or a Xeon E7–8800 series processor for an eight-socket (or more) server. These new options from Intel can be quite confusing to sort out unless you pay attention to the details.

Prior to the release of SQL Server 2012, paying the price premium for the absolute best processor available for each socket in your database server was an effective strategy for database server processor selection. The SQL Server processor license cost was pretty high (even for Standard Edition), so you wanted to get as much performance and scalability capacity as possible for each expensive processor socket license that you purchased.

This is still a valid strategy for SQL Server 2008 R2 and earlier, but the licensing changes in SQL Server 2012 Enterprise Edition dictate a few modifications to this line of thinking. In early November 2011, Microsoft announced some rather fundamental changes regarding how SQL Server 2012 will be licensed compared to previous versions. SQL Server 2012 has three main editions: Enterprise Edition, Business Intelligence Edition, and Standard Edition. The old Data Center Edition and Workgroup Edition have been eliminated, which is probably no big loss. The existing Developer and Express Editions are still available, along with Web Edition for hosting providers.

Rather than the old, familiar socket-based licensing used in SQL Server 2008 R2 and earlier, SQL Server 2012 uses a combination of core-based and Server + Client Access License (CAL) licensing, depending on which edition you buy, and which choice you make for Standard Edition. With Standard Edition, you can choose core-based licensing or Server + CAL-based licensing. With Business Intelligence Edition, you have to use Server + CAL-based licensing, while Enterprise Edition requires the use of core-based licensing. Standard Edition is the base edition, with a limit of 16 physical processor cores. Microsoft decided to maintain the 64GB RAM limit for SQL Server 2012 Standard Edition (just like the 64GB RAM limit in SQL Server 2008 R2 Standard Edition). Business Intelligence Edition includes all the functionality of Standard Edition, plus extra BI features and functionality. Enterprise Edition includes everything in BI Edition, plus all the extra Enterprise Edition features and functionality. Enterprise Edition is the top-of-the-line edition of SQL Server 2012, now including all the features that were available in SQL Server 2008 R2 Data Center Edition. As a DBA, you really want to use Enterprise Edition if you have any choice in the matter, as it offers so many useful features, such as online index operations, data compression, and AlwaysOn availability groups, to name a few.

If you are using core-based licensing (as you must for SQL Server 2012 Enterprise Edition), each physical socket in your server must use a minimum of four core licenses. That means if you have old hardware that uses dual-core processors, you still have to buy four core licenses for each socket. That is yet another reason to not use ancient hardware for SQL Server 2012. Any Intel Xeon or AMD Opteron processor that has only two physical cores was at least four to five years old by the time SQL Server 2012 was released, so it really should be retired. Keep in mind that only physical cores count for licensing purposes (on non-virtualized servers), so Intel hyperthreading is free from a licensing perspective.

Core licenses are now sold in two-core packs, again with a minimum of four cores per physical socket. The full retail license cost per physical core is $6,874 for SQL Server 2012 Enterprise Edition. This is pretty grim news for AMD, with their higher physical core counts and lower per-socket performance compared to Intel. This situation was so obvious that Microsoft released a SQL Server 2012 Core Factor Table on April 1, 2012, that reduces the per-core license cost by 25% for a number of modern AMD processors that have six or more cores. Even with this change, the latest AMD processors are not a very cost-effective choice for SQL Server 2012. The numbers in Table 1 show the cost differential in pretty graphic detail, even with the .75 AMD Core Factor (see the “AMD Processors and Numbering” section later) applied to the license costs for the AMD processors.

TABLE 1: SQL Server 2012 License Cost Comparison

For an OLTP workload on a two-socket server, an Intel Xeon E5–2690 processor would be preferable to an Intel Xeon E7–2870 processor because of its better single-threaded performance, a result of being a newer-generation model (Sandy Bridge-EP vs. Westmere-EX), higher clock speed, better memory bandwidth, and PCIe 3.0 support. For a DSS/DW workload, the E5–2690 would be preferable for the same reasons, even though it has a lower core count and a smaller L3 cache size.

For most OLTP workloads, you would also be far better off, from a performance perspective, with an older two-socket Intel Xeon X5690 server or a two-socket Intel Xeon E5–2690 server than you would be with a four-socket AMD Opteron 6282SE server. The extremely large difference in license cost between those two options makes Intel an even more compelling choice. As shown in Table 2, one way to partially confirm this assessment is to look at TPC-E scores for different systems and divide them by the total physical core count for the system (not by the thread count).

TABLE 2: TPC-E Scores by Total Physical Cores

It is very unlikely that you would ever upgrade to a better processor in an existing database server, so you will be stuck with your processor choice for the life of the server. If you have “excess” processor capacity, consider using it to trade CPU utilization for I/O utilization by using backup compression and data compression (if you have the Enterprise Edition of SQL Server 2008 or newer). Unlike a laptop or web server, it is a mistake to buy a processor that is a couple of steps down from the top-of-the-line model for database server usage. Trading some extra CPU utilization for less I/O utilization is usually a net win, especially if you have a modern, multi-core processor that can readily handle the extra work.

Of course, a new two-socket server will have a lower total RAM limit than a new four-socket server. For example, a two-socket Xeon X5690 would be limited to 288GB of RAM, which is probably enough for most workloads. A two-socket server will also have less total I/O capacity than a new four-socket server because it has fewer PCIe expansion slots. Still, you can easily get 5–6GB/sec of sequential throughput out of a modern two-socket server, which should be plenty for most workloads. After the Intel 32nm Sandy Bridge-EP Xeon E5–2600 series was released in early 2012, the wisdom of choosing a two-socket Intel-based server was even clearer, as it has higher memory density, more I/O bandwidth, and even better per-core performance than the Xeon 5600 series did.

If you are looking at the lower end of the cost and workload spectrum, you have several options. The one-socket 22nm Intel Xeon E3–1290 v2 processors (which are basically the same as the desktop Ivy Bridge Core i7 processor) are limited to 32GB of RAM, which somewhat limits their utility for larger database usage. If 32GB of RAM is not enough for your workload, a single-socket Dell R320 server with one Intel Xeon E5–2400 series processor and up to 96GB of RAM is available. Keep in mind that the memory limit for SQL Server 2012 Standard Edition is still 64GB, which is too low considering the memory density of modern hardware. One possible way around it with good hardware (with more than 128GB of RAM) is to install more than one instance of SQL Server 2012 Standard Edition on the same physical server.

Classic Intel Processor Numbering

In order to understand older Intel processor numbers, you need to know how to decode “classic” Intel processor numbers. By classic we mean Intel Xeon processors produced from about 2006 until April 2011 (when Intel introduced a new processor numbering system for new and upcoming processors).

Knowing how to decode the processor model number is a very handy skill to have when you want to understand the capabilities, relative age, and relative performance of a particular processor. An example of an Intel processor number is shown in Figure 1.

FIGURE 1

Intel Xeon processor numbers are categorized in four-digit numerical sequences, plus an alpha prefix that indicates whether it is optimized for electrical power usage or performance. The alpha prefixes are as follows:

• X, meaning performance
• E, meaning mainstream
• L, meaning power optimized

The model number starts with 3, 5, or 7, depending on the server form factor for which the processor is designed. If the processor number starts with a 3, it is designed for a single-socket server; if it starts with a 5, it is designed for a two-socket server; and if it starts with a 7, it is designed for a four-socket or more server. The second digit of the model number designates the generation, or relative age, of a processor. For example, the Xeon 5100 series was launched in Q2 2006, while the Xeon 5300 series was launched in Q4 2006, and the Xeon 5400 series was launched in Q4 2007.

For a more complete example, a Xeon X7560 is a high-end performance processor for multi-processor systems, an Intel Xeon E5540 is a mainstream processor for dual-processor systems, while an Intel Xeon L5530 is a power-optimized processor for dual-processor systems. The final three digits denote the generation and performance of the processor; for example, a Xeon X7560 processor would be newer and probably more capable than a Xeon X7460 processor. Higher numbers for the last three digits of the model number mean a newer generation in the family — for example, 560 is a newer generation than 460.

You should always choose the performance models, with the X model prefix, for SQL Server usage. The additional cost of an X series Xeon processor, compared to an E series, is minimal compared to the overall hardware and SQL Server license cost of a database server system. You should also avoid the power-optimized L series, as these can reduce processor performance by 20% to 30% while only saving 20 to 30 watts of power per processor, which is pretty insignificant compared to the overall electrical power usage of a typical database server (with its cooling fans, internal drives, power supplies, etc.). Of course, it would be a different story if you needed dozens or hundreds of web servers instead of a small number of mission-critical database servers, as the overall power savings would be pretty significant in that case.

Current Intel Processor Numbering

This section explains the current processor numbering system for Xeon processors that Intel introduced on April 5, 2011. This new system, shown in Figure 2, is used for the new processor families that Intel released on that date (the E3 series and the E7 series) and the E5 series that was released in March of 2012. The model numbers for the older existing Xeon processors remain unchanged in this system.

FIGURE 2

The first two digits in the processor number represent the Product Line designation, which will be E3, E5, or E7, depending on their place in the overall product lineup. After the Product Line designation is a four-digit number that provides more details about the particular processor. The first digit is the “wayness,” which is the number of physical CPUs that are allowed in a node (which is a physical server). This first digit can be 1, 2, 4, or 8. The second digit is the socket type, in terms of its physical and electrical characteristics. The last two digits are the processor SKU, with higher numbers generally indicating higher performance. Finally, an L at the end indicates energy-efficient, low electrical power processors. For SQL Server database server usage, you should avoid these power-optimized processors, as the performance impact of the reduced power usage is pretty dramatic.

The E3 Product family is for single-processor servers or workstations. The first generation of this family (E3–1200 series) is essentially the same as the desktop 32nm Sandy Bridge processors that were released in January 2011. The second generation of this family is the E3–1200 v2 series, which is basically the same as the desktop 22nm Ivy Bridge processors that were released in May 2012. They are both limited to 32GB of RAM.

The E5 Product family (the 32nm Sandy Bridge-EP) includes the E5–2600 series that was released in March 2012, and the E5–2400 series (32nm Sandy Bridge-EN) and E5–4600 series that were released in May 2012. You should probably avoid the entry-level Sandy Bridge-EN series, which has less memory bandwidth and lower clock speeds compared to the Sandy Bridge-EP series.

The E7 Product family (the 32nm Westmere-EX) has different models that are meant for two-socket servers, four-socket servers, and eight-socket and above servers. The E7–2800 series is for two-socket servers, the E7–4800 series is for four-socket servers, while the E7–8800 series is for eight-socket and above servers. Just in case you are wondering, the “EP” designation at the end of the family code word (such as Westmere-EP) stands for “efficient performance,” while the “EX” designation stands for “expandable.”

Intel’s Tick-Tock Release Strategy

Since 2006, Intel has adopted and implemented a Tick-Tock strategy for developing and releasing new processor models. Every two years, they introduce a new processor family, incorporating a new microarchitecture; this is the tock release. One year after the tock release, they introduce a new processor family that uses the same microarchitecture as the previous year’s tock release, but using a smaller manufacturing process technology and usually incorporating other small improvements, such as larger cache sizes or improved memory controllers. This is the tick release. This Tick-Tock release strategy benefits the DBA in a number of ways. It offers better predictability regarding when major (tock) and minor (tick) releases will be available. This helps you plan hardware upgrades to possibly coincide with your operating system and SQL Server version upgrades.

Tick releases are usually socket-compatible with the previous year’s tock release, which makes it easier for the system manufacturer to make the latest tick release processor available in existing server models quickly, without completely redesigning the system. In most cases, only a BIOS update is required to enable an existing model system to use a newer tick release processor. This makes it easier for the DBA to maintain servers that are using the same model number (such as a Dell PowerEdge R710 server), as the server model will have a longer manufacturing life span. For example, the Dell PowerEdge R710 was able to use the original 45nm Nehalem-EP Xeon 5500 series processors and the newer 32nm Westmere-EP Xeon 5600 series processors, so that model server was available for purchase for over three years.

As a DBA, you need to know where a particular processor falls in Intel’s processor family tree in order to meaningfully compare the relative performance of two different processors. Historically, processor performance has nearly doubled with each new tock release, while performance usually increases by around 20–25% with a tick release. Some of the recent and upcoming Intel Tick-Tock releases are shown in Figure 3.

FIGURE 3

The manufacturing process technology refers to the size of the individual circuits and transistors on the chip. The Intel 4004 (released in 1971) series used a 10-micron process; the smallest feature on the processor was 10 millionths of a meter across. By contrast, the Intel Xeon “Ivy Bridge” E3–1200 v2 series (released in May 2012) uses a 22nm process. For comparison, a nanometer is one billionth of a meter, so 10 microns would be 10,000 nanometers. This ever-shrinking manufacturing process is important for two main reasons:

• Increased performance and lower power usage — Even at the speed of light, distance matters, so having smaller components that are closer together on a processor means better performance and lower power usage.
• Lower manufacturing costs — This is possible because more processors can be produced from a standard silicon wafer. This helps to create more powerful and more power-efficient processors available at a lower cost, which is beneficial to everyone but especially the database administrator.

The first tock release was the Intel Core microarchitecture, which was introduced as the dual-core “Woodcrest” (Xeon 5100 series) in 2006, with a 65nm process technology. This was followed up by a shrinkage to 45nm process technology in the dual-core “Wolfdale” (Xeon 5200 series) and quad-core “Harpertown” processors (Xeon 5400 series) in late 2007, both of which were Tick releases. The next tock release was the Intel “Nehalem” microarchitecture (Xeon 5500 series), which used a 45nm process technology, introduced in late 2008. In 2010, Intel released a Tick release, code-named “Westmere” (Xeon 5600 series) that shrank to a 32nm process technology in the server space. In 2011, the 32nm “Sandy Bridge” tock release debuted with the E3–1200 series for single-socket servers and workstations. This was followed up by the “Ivy Bridge” tick release of the E3–1200 v2 series for single-socket servers and workstations that had a process shrink to 22nm. Table 3 shows the recent and upcoming Tick-Tock releases in the two-socket server space.

TABLE 3: Intel Tick-Tock Release History for Two Socket Servers

Intel Hyperthreading

Intel originally implemented a feature called hyperthreading back in 2002, as part of the NetBurst architecture in the Northwood-based Pentium 4 processors and the equivalent Xeon family. Hyperthreading was created to address the frequently wasted processor cycles that resulted when the central processor in a system waited on data from main memory. Instead of wasting processor cycles during this wait time, the idea was to have two logical processors inside a single physical core that could each work on something different when the other logical processor was stalled waiting on data from main memory.

Hyperthreading is Intel’s marketing term for its simultaneous multi-threading architecture whereby each physical processor core is split into two logical cores. The “simultaneous” term is a little misleading, as you cannot actually have two threads running simultaneously on the two logical cores in a single physical core of the same physical processor. What actually happens is that the threads run alternately, with one working while the other one is idle.

Hyperthreading works quite well for desktop applications. The classic example is running a complete anti-virus scan while the user is still able to work interactively with another application in the foreground. Unfortunately, the initial implementation of hyperthreading on the Pentium 4 NetBurst architecture did not work very well on many server workloads such as SQL Server. This was because the L2 data cache for each physical core was shared between the two logical cores, which caused performance issues because the L2 cache had to be constantly refreshed as the application context switched between the two logical processors. This behavior was known as cache thrashing, and it often led to a decrease in overall performance for SQL Server workloads. Another factor that made this situation even worse was the very deep processor pipeline that was used in the Pentium 4 architecture, which made it even more costly when the data needed by the logical processor was not found in the L2 cache.

Because of these factors, it became very common for database administrators to disable hyperthreading for all SQL Server workloads, which is really a mistake. Different types of SQL Server workloads react differently to having hyperthreading enabled, with OLTP workloads generally performing better with hyperthreading enabled, and data warehouse workloads sometimes performing worse with hyperthreading enabled. Before you decide whether to enable or disable hyperthreading, test it both ways with your actual workload.

Modern Intel processors (Nehalem, Westmere, Sandy Bridge, and Ivy Bridge) seem to work much better with hyperthreading because of larger L2 and L3 cache sizes, newer processor architectures, and faster access to main memory. Because of this, we advise you to enable hyperthreading for SQL Server, especially for OLTP workloads, unless you have done testing that actually shows a performance decrease with your workload. It is significant that every single TPC-E OLTP benchmark submission for these modern Intel processors has been done with hyperthreading enabled on the database server, which is certainly intentional.

2. AMD Processors and Numbering

This section discusses AMD Opteron processor numbering. Advanced Micro Devices (AMD) has various versions of the Opteron family that are meant for server use. When assessing AMD processors, it is very helpful to understand what the model numbers actually mean. Recent AMD Opteron processors are identified by a four-character model number in the format ZYXX, where the Z character indicates the product series:

• 1000 Series = 1-socket servers
• 2000 Series = Up to 2-socket servers and workstations
• 4000 Series = Up to 2-socket servers
• 6000 Series = High performance 2- and 4-socket servers
• 8000 Series = Up to 8-socket servers and workstations

The Y character differentiates products within a series:

• Z2XX = Dual-Core.
• Z3XX = Quad-Core.
• Z4XX = Six-Core.
• First-generation AMD Opteron 6000 series processors are denoted by 61XX.
• Second-generation AMD Opteron 6000 series processors are denoted by 62XX.

The XX digits indicate a change in product features within the series (for example, in the 8200 series of dual-core processors, you can find models 8214, 8216, 8218, and so on), and are not a measure of performance. It is also possible to have a two-character product suffix after the XX model number, as follows:

• No suffix — Indicates a standard power AMD Opteron processor
• SE — Performance optimized, high-powered
• HE — Low-powered
• EE — Lowest power AMD Opteron processor

For example, an Opteron 6282 SE would be a 6000 series, 16-core, performance-optimized processor; an Opteron 8439 SE would be an 8000 series, six-core, performance-optimized processor; while an Opteron 2419 EE would be a 2000 series, six-core, energy-efficient processor. For mission-critical database servers, we recommend selecting an SE suffix processor, if it is available for your server model. The reason why it isn’t available in every server model is due to its higher electrical power requirements.

It should also be noted that AMD has broken their own processor numbering rules with the most recent versions of the Opteron (including the 4100, 4200, 6100, and 6200 series), as they do not follow the standard numbering scheme just described.

Recent Opteron AMD releases, plus planned releases, are summarized in Table 4. Since 2011, the 16-core Interlagos processor has been AMD’s best-performing model, even though it did not live up to expectations for that release.

TABLE 4: Recent AMD Processor Releases

SQL Server 2012 Core Factor Table

Looking at recent TPC-E benchmark submissions for both AMD and Intel processors, it is pretty difficult to avoid noticing how poorly the few tested AMD systems have done compared to the latest Intel systems. For example, in January 2012, a new Hewlett-Packard TPC-E benchmark submission showed a 1232.84 TpsE score for a two-socket AMD system with 32 physical cores, compared to a 1284.14 TpsE score for a two-socket Intel system with 12 physical cores. Both of these TPC-E benchmark submissions were on SQL Server 2008 R2. With these results, you would be paying 2.66 times as much for SQL Server 2012 Enterprise Edition core licenses for the AMD system compared to the Intel system (32 physical cores vs. 12 physical cores). This is rather dire news for AMD, with their higher physical core counts and lower per physical core OLTP performance compared to Intel.

Likely in response to this situation, on April 1, 2012, Microsoft released a new SQL Server 2012 Core Factor Table for AMD processors, which is shown in Table 5. Note that not all processors are included in the table.

TABLE 5: SQL Server 2012 Core Factor Table for AMD Processors

The most relevant part of this table regards the newer AMD 31XX, 32XX, 41XX, 42XX, 61XX, and 62XX series of processors with six or more cores that have a core factor of 0.75. Having a core factor of 0.75 means that you multiply the actual number of physical cores times the core factor to arrive at the number of cores for SQL Server licensing purposes; for example, if you had a four-socket server, where each socket was populated with an AMD Opteron 6284 SE processor. That particular processor has 16 physical cores, so 4 times 16 would give you a result of 64 SQL Server 2012 core licenses that would be required for that server (before the Core Factor table was introduced). Using the new licensing rules from the Core Factor table, you would be able to multiply 64 times 0.75 to get a new result of 48 SQL Server 2012 core licenses that would be required for that server (after the Core Factor table was introduced). This means that AMD cores for some processors are somewhat more affordable now for SQL Server 2012 than they would be without the core factor calculation.

Based on the SQL Server 2012 Core Factor Table, you would only be paying twice as much for SQL Server 2012 Enterprise Edition licenses for the 32-core AMD system compared to the 12-core Intel system (32 AMD physical cores times 0.75 vs. 12 Intel physical cores). That is a slightly better story for AMD, but it is still a pretty hard sell.

Based on the TPC-E benchmark results, both the older Intel Xeon X5600 Westmere-EP series and the new Intel Xeon E5–2600 Sandy Bridge-EP series perform much better per physical core on OLTP workloads than the latest AMD Opteron 6200 series processors. These Intel processors simply have significantly better single-threaded performance, which is very important for OLTP workloads.

As a result of this new Core Factor Table, SQL Server 2012 processor licenses will be a little less expensive than they were previously for those AMD processor families that have more than six cores, but they will still be much more expensive in total than a better-performing Intel solution. The somewhat lower hardware cost for the AMD processor compared to the hardware cost of the Intel processor is rather trivial compared to the difference in the licensing cost. Hopefully AMD can do better with the upcoming Piledriver core-based Opteron series expected in 2013.