|M.Sc Student||Bilic' Hrvoye|
|Subject||TCP/IP Pipelining - Improving Performance of Large|
Transfers over Standard Gb/s Ethernet with
Unmodified Operating Systems and
|Department||Department of Electrical and Computer Engineering||Supervisor||ASSOCIATE PROFESSOR Yitzhak Birk|
Over the past few years, the speed of networking hardware has doubled or even tripled, while the efforts to reduce the standard protocol stack overhead have yielded only moderate increases in speed. The communication bottleneck is moving gradually from hardware to software and nowadays the TCP/IP stack has become a communication bottleneck for Gigabit networks.
Many previous studies have attempted to overcome the TCP/IP stack bottleneck. Most of them are based on TCP/IP protocol splitting between the host CPU and network adapter, offloading some or all of the TCP/IP functions from the host CPU to the network adapter. However, these research efforts and proposals generally do not take into consideration the implications of required changes to existing applications and operating systems or the required development efforts for modifying network adapters.
This thesis proposes a different approach towards improving TCP/IP communication performance over Gigabit Ethernet networks. We propose a TCP/IP pipelining approach; offloading only a small subset of the OS-independent TCP/IP functionality to the network adapter, while still pushing TCP/IP communication bandwidth over Gigabit Ethernet networks to the wire speed. The general idea is to permit large TCP frames through the protocol stack and subsequently segment them into small packets that can be carried by the network. This approach requires only minor network adapter development efforts and does not require any changes to existing applications or OS protocol stacks; it is suitable for immediate implementation.