PCI Express switching technology makes a remote system's memory available
to a local system for DMA operations and load/store CPU instructions. The
latency of data access, in this system, is low compared to the traditional system where an Ethernet frame passes
through Ethernet Host Bus Adapters and Switches.
Since a remote system’s memory is available to the local system, applications running on one system can read and write information to another application’s address space that is running on a different system directly, without incurring any overhead of copy from one protection domain to another. No socket or any form of communication library is necessary.
TCP/IP network has inherent robustness and reliability for data transfer from one system to other. Sending Ethernet frames over PCIe bus combines low latency data transfer with robustness of TCP/IP protocol. Though connection less user data gram protocol(UDP) works well as PCIe bus also ensures reliable data transfer.
This paper describes the anatomy of data transfer from one system to another over PCI Express bus and explains issues with sending Ethernet frames over a point-point PCI Express link. Network benchmarks such as netperf show that even in non-optimized programmed IO mode of data transfer throughput reaches 9.97 Gigabits per second with low CPU utilization over x4 Gen2 (5Gigabit/sec) PCI Express Bus. The latency numbers are close to 30k transactions/second which is as good as 10 Gigabit Ethernet cards that are available on the market.
Click here to download Powerpoint Presentation
Click here to download Presentation Content