Cisco Data Center Virtualization Server Architectures

Processor, memory, and I/O are the three most important subsystems in a server from a performance perspective. At any given point in time, one of them tends to become a bottleneck from the performance perspective. We often hear of applications that are CPU-bound, memory-bound, or I/O-bound.

In this chapter, we will examine these three subsystems with particular reference to servers built according to the IA-32 (Intel^® Architecture, 32-bit), often generically called x86 architecture. In particular, we will describe the most recent generation of Intel^® processors compatible with the IA-32 architecture; i.e., the Intel^® microarchitecture (formerly known by the codename Nehalem). ¹

The Nehalem microarchitecture (see "Intel Microarchitectures" in Chapter 2, page 45) and its variation, the Westmere microarchitecture, include three families of processors that are used on the Cisco UCS: the Nehalem-EP, the Nehalem-EX, and the Westmere-EP. Table 2-1 summarizes the main characteristics of these processors.

Table 2-1. UCS Processors

The Processor Evolution

Modern processors or CPUs (Central Processing Units) are built using the latest silicon technology and pack millions of transistors and megabytes of memory on a single die (blocks of semiconducting material that contains a processor).

Multiple dies are fabricated together in a silicon wafer; each die is cut out individually, tested, and assembled in a ceramic package. This involves mounting the die, connecting the die pads to the pins on the package, and sealing the die.

At this point, the processor in its package is ready to be sold and mounted on servers. Figure 2-1 shows a packaged Intel^® Xeon^® 5500.

Figure 2-1 An Intel Xeon 5500 Processor

Sockets

Processors are installed on the motherboard using a mounting/interconnection structure known as a "socket." Figure 2-2 shows a socket used for an Intel^® Processor. This allows the customers to personalize a server motherboard by installing processors with different clock speeds and power consumption,

Figure 2-2 An Intel processor socket

The number of sockets present on a server motherboard determines how many processors can be installed. Originally, servers had a single socket, but more recently, to increase server performance, 2-, 4-, and 8-socket servers have appeared on the market.

In the evolution of processor architecture, for a long period, performance improvements were strictly related to clock frequency increases. The higher the clock frequency, the shorter the time it takes to make a computation, and therefore the higher the performance.

As clock frequencies approached a few GigaHertz, it became apparent the physics involved would limit further improvement in this area. Therefore, alternative ways to increase performance had to be identified.

Cores

The constant shrinking of the transistor size (Nehalem uses a 45-nanometer technology; Westmere uses a 32-nanometer technology) has allowed the integration of millions of transistors on a single die. One way to utilize this abundance of transistors is to replicate the basic CPU (the "core") multiple times on the same die.

Multi-core processors (see Figure 2-3) are now common in the market. Each processor (aka socket) contains multiple CPU cores (2, 4, 6, and 8 are typical numbers). Each core is associated with a level 1 (L1) cache. Caches are small fast memories used to reduce the average time to access the main memory. The cores generally share a larger level 2 (L2) or level 3 (L3) cache, the bus interface, and the external die connections.

Figure 2-3 Two CPU cores in a socket

In modern servers, the number of cores is the product of the number of sockets times the number of cores per socket. For example, servers based on Intel^® Xeon^® Processor 5500 Series (Nehalem-EP) typically use two sockets and four cores per sockets for a total of eight cores. With the Intel^® Xeon^® 7500 (Nehalem-EX), 8 sockets each with 8 cores are supported for a total of 64 cores.

Figure 2-4 shows a more detailed view of a dual-core processor. The CPU's main components (instruction fetching, decoding, and execution) are duplicated, but the access to the system buses is common.

Figure 2-4 Architecture of a dual-core processor

Threads

To better understand the implication of multi-core architecture, let us consider how programs are executed. A server will run a kernel (e.g., Linux^®, Windows^®) and multiple processes. Each process can be further subdivided into "threads". Threads are the minimum unit of work allocation to cores. A thread needs to execute on a single core, and it cannot be further partitioned among multiple cores (see Figure 2-5).

Figure 2-5 Processes and threads

Processes can be single-threaded or multi-threaded. A process that is single thread process can execute in only one core and is limited by the performance of that core. A multi-threaded process can execute on multiple cores at the same time, and therefore its performance can exceed the performance of a single core.

Since many applications are single-threaded, a multi-socket, multi-core architecture is typically convenient in an environment where multiple processes are present. This is always true in a virtualized environment, where a hypervisor allows consolidating multiple logical servers into a single physical server creating an environment with multiple processes and multiple threads.

Intel^® Hyper-Threading Technology

While a single thread cannot be split between two cores, some modern processors allow running two threads on the same core at the same time. Each core has multiple execution units capable of working in parallel, and it is rare that a single thread will keep all the resources busy.

Figure 2-6 shows how the Intel^® Hyper-Threading Technology works. Two threads execute at the same time on the same core, and they use different resources, thus increasing the throughput.

Figure 2-6 Intel Hyper-Threading Technology

Front-Side Bus

In the presence of multi-sockets and multi-cores, it is important to understand how the memory is accessed and how communication between two different cores work.

Figure 2-7 shows the architecture used in the past by many Intel^® processors, known as the Front-Side Bus (FSB). In the FSB architecture, all traffic is sent across a single, shared bidirectional bus. In modern processors, this is a 64-bit wide bus that is operated at 4X the bus clock speed. In certain products, the FSB is operated at an information transfer rate of up to 1.6 GT/s (Giga Transactions per second, i.e., 12.8 GB/s).

Figure 2-7 A server platform based on a front-side bus

The FSB is connected to all the processors and to the chipset called the North-bridge (aka MCH: Memory Controller Hub). The Northbridge connects the memory that is shared across all the processors.

One of the advantages of this architecture is that each processor has knowledge of all the memory accesses of all the other processors in the system. Each processor can implement a cache coherency algorithm to keep its internal caches in synch with the external memory and with the caches of all other processors.

Platforms designed in this manner have to contend with the shared nature of the bus. As signaling speeds on the bus increase, it becomes more difficult to implement and connect the desired number of devices. In addition, as processor and chipset performance increases, the traffic flowing on the FSB also increases. This results in increased congestion on the FSB, since the bus is a shared resource.

Dual Independent Buses

To further increase the bandwidth, the single shared bus evolved into the Dual Independent Buses (DIB) architecture depicted in Figure 2-8, which essentially doubles the available bandwidth.

Figure 2-8 A server platform based on Dual Independent Buses

However, with two buses, all the cache consistency traffic has to be broadcasted on both buses, thus reducing the overall effective bandwidth. To minimize this problem, "snoop filters" are employed in the chipset to reduce the bandwidth loading.

When a cache miss occurs, a snoop is put on the FSB of the originating processor. The snoop filter intercepts the snoop, and determines if it needs to pass along the snoop to the other FSB. If the read request is satisfied with the other processor on the same FSB, the snoop filter access is cancelled. If the other processor on the same FSB does not satisfy the read request, the snoop filter determines the next course of action. If the read request misses the snoop filter, data is returned directly from memory. If the snoop filter indicates that the target cache line of the request could exist on the other FSB, the snoop filter will reflect the snoop across to the other segment. If the other segment still has the cache line, it is routed to the requesting FSB. If the other segment no longer owns the target cache line, data is returned from memory. Because the protocol is write-invalidate, write requests must always be propagated to any FSB that has a copy of the cache line in question.

Dedicated High-Speed Interconnect

The next step of the evolution is the Dedicated High-Speed Interconnect (DHSI), as shown in Figure 2-9.

Figure 2-9 A server platform based on DHSI

DHSI-based platforms use four independent FSBs, one for each processor in the platform. Snoop filters are employed to achieve good bandwidth scaling.

The FSBs remains electrically the same, but are now used in a point-to-point configuration.

Platforms designed using this approach still must deal with the electrical signaling challenges of the fast FSB. DHSI drives up also the pin count on the chipset and require extensive PCB routing to establish all these connections using the wide FSBs.

Intel^® QuickPath Interconnect

With the introduction of the Intel Core i7 processor, a new system architecture has been adopted for many Intel products. This is known as the Intel^® QuickPath Interconnect (Intel^® QPI). This architecture utilizes multiple high-speed uni-directional links interconnecting the processors and the chipset. With this architecture, there is also the realization that:

• A common memory controller for multiple sockets and multiple cores is a bottleneck.
• Introducing multiple distributed memory controllers would best match the memory needs of multi-core processors.
• In most cases, having a memory controller integrated into the processor package would boost performance.
• Providing effective methods to deal with the coherency issues of multi-socket systems is vital to enabling larger-scale systems.

Figure 2-10 gives an example functional diagram of a processor with multiple cores, an integrated memory controller, and multiple Intel^® QPI links to other system resources.

Figure 2-10 Processor with Intel QPI and DDR3 memory channels

In this architecture, all cores inside a socket share IMCs (Integrated Memory Controllers) that may have multiple memory interfaces (i.e., memory buses).

IMCs may have different external connections:

• DDR3 memory channels: In this case, the DDR3 DIMMs (see the next section) are directly connected to the sockets, as shown in Figure 2-12. This architecture is used in Nehalem-EP (Xeon 5500) and Westmere-EP (Xeon 5600).
• High-speed serial memory channels, as shown in Figure 2-11. In this case, an external chip (SMB: Scalable Memory Buffer) creates DDR3 memory channels where the DDR3 DIMMs are connected. This architecture is used in Nehalem-EX (Xeon 7500).
  

  Figure 2-11 Processors with high-speed memory channels

IMCs and cores in different sockets talk to each other using Intel^® QPI.

Processors implementing Intel^® QPI also have full access to the memory of every other processor, while maintaining cache coherency. This architecture is also called "cache-coherent NUMA (Non-Uniform Memory Architecture)"—i.e., the memory interconnection system guarantees that the memory and all the potentially cached copies are always coherent.

Intel^® QPI is a point-to-point interconnection and messaging scheme that uses a point-to-point differential current-mode signaling. On current implementation, each link is composed of 20 lanes per direction capable of up to 25.6 GB/s or 6.4 GT/s (Giga Transactions/second); (see "Platform Architecture" in Chapter 2, page 46).

Intel^® QPI uses point-to-point links and therefore requires an internal crossbar router in the socket (see Figure 2-10) to provide global memory reachability. This route-through capability allows one to build systems without requiring a fully connected topology.

Figure 2-12 shows a configuration of four Intel^® Nehalem EX, each processor has four QPI and interconnects with the three other processors and with the Boxboro-EX chipsets (SMB components are present, but not shown).

Figure 2-12 Four-socket Nehalem EX