Each processor has some amount of memory, connected directly to the processor. To access
memory connected to another processor, each processor uses QPI links through another
processors. This design creates a non-uniform memory access (NUMA) system. Similarly, I/O
can be local to a processor or remote through another processor.
For QPI use, Intel modified the MESI cache coherence protocol to include a forwarding state.
Therefore, when a processor asks to copy a shared cache line, only one other processor
responds.
For more information about QPI, visit the following web page:
http://www.intel.com/technology/quickpath
Intel Data Direct I/O
For I/O, Intel no longer has a separate I/O hub. It now integrates PCI Express 3.0 I/O into the
processor itself. Data Direct I/O helps to optimize data transfer between local CPU and PCIe
devices. The combination of Data Direct I/O and PCI 3.0 provides a significantly higher I/O
performance with lower latencies and reduced power consumption.
For more information about Data Direct I/O, visit the following web page:
http://www.intel.com/content/www/us/en/io/direct-data-i-o.html
RAS features
The Intel Xeon processor E7 family of processors has additional reliability, availability, and
serviceability (RAS) features on their interconnect links (SMI and QPI):
Cyclic redundancy checking (CRC) on the QPI links:
The data on the QPI link is checked for errors.
QPI packet retry:
If a data packet on the QPI link has errors or cannot be read, the receiving processor can
request that the sending processor try resending the packet.
QPI clock failover:
If there is a clock failure on a coherent QPI link, the processor on the other end of the link
can become the clock. This action is not required on the QPI links from processors to I/O
hubs, as these links are asynchronous.
QPI self-healing:
If there are persistent errors detected on a QPI link, the link width can be reduced
dynamically to allow the system to run in a degraded mode until repair can be performed.
QPI link can reduce its width to a half width or a quarter width, and slowdown its speed.
Scalable memory interconnect (SMI) packet retry:
If a memory packet has errors or cannot be read, the processor can request that the
packet be resent from the memory buffer.
27
Chapter 2. Technology