The I/O Subsystem
The I/O subsystem is in charge of moving data from the server memory to the external world and vice versa. Historically this has been accomplished by providing in the server motherboards I/O buses compatible with the PCI (Peripheral Component Interconnect) standard. PCI was developed to interconnect peripheral devices to a computer system, it has been around for many years [1] and its current incarnation is called PCI-Express.
The Peripheral Component Interconnect Special Interest Group (PCI-SIG) is in charge of the development and enhancement of the PCI standard.
PCI Express®
PCI Express (PCIe®) [2] is a computer expansion card interface format designed to replace PCI, PCI-X, and AGP.
It removes one of the limitations that have plagued all the I/O consolidation attempts—i.e., the lack of I/O bandwidth in the server buses. It is supported by all current operating systems.
The previous bus-based topology of PCI and PCI-X is replaced by point-to-point connectivity. The resultant topology is a tree structure with a single root complex. The root complex is responsible for system configuration, enumeration of PCIe resources, and manages interrupts and errors for the PCIe tree. A root complex and its endpoints share a single address space and communicate through memory reads and writes, and interrupts.
PCIe connects two components with a point-to-point link. Links are composed of N lanes (a by-N link is composed of N lanes). Each lane contains two pairs of wires: one pair for transmission and one pair for reception.
Multiple PCIe lanes are normally provided by the SouthBridge (aka ICH: I/O Controller Hub) that implements the functionality of "root complex".
Each lane connects to a PCI Express endpoint, to a PCI Express Switch or to a PCIe to PCI Bridge, as in Figure 2-21.
)
Figure 2-21 PCI-Express root complex
Different connectors are used according to the number of lanes. Figure 2-22 shows four different connectors and indicates the speeds achievable with PCIe 1.1.
)
Figure 2-22 PCI Express connectors
In PCIe 1.1, the lanes run at 2.5 Gbps (2 Gbps at the datalink) and 16 lanes can be deployed in parallel (see Figure 2-23). This supports speeds from 2 Gbps (1x) to 32 Gbps (16x). Due to protocol overhead, 8x is required to support a 10 GE interface.
)
Figure 2-23 PCI Express lanes
PCIe 2.0 (aka PCIe Gen 2) doubles the bandwidth per lane from 2 Gbit/s to 4 Gbit/s and extends the maximum number of lanes to 32x. It is shipping at the time of writing. A PCIe 4x is sufficient to support 10 GE.
PCIe 3.0 will approximately double the bandwidth again. The final PCIe 3.0 specifications, including form factor specification updates, may be available by mid 2010, and could be seen in products starting in 2011 and beyond [3]. PCIe 3.0 will be required to effectively support 40 GE (Gigabit Ethernet), the next step in the evolution of Ethernet.
All the current deployments of PCI Express are Single Root (SR), i.e., a single I/O Controller Hub (ICH) controlling multiple endpoints.
Multi Root (MR) has been under development for a while, but it has not seen the light yet, and many question if it ever will, due to the lack of components and interest.
SR-IOV (Single Root I/O Virtualization) is another extremely relevant standard developed by PCI-SIG to be used in conjunction with Virtual Machines and Hypervisors. It is discussed in "DCBX: Data Center Bridging eXchange" in Chapter 3, page 73.