Speculative Defragmentation - 
             A Technique to Improve the Communication
             Software Efficiency for Gigabit Ethernet 

  Christian Kurmann, Michel Muller, Felix Rauch, Thomas Stricker
            Swiss Federal Institute of Technology (ETH)
                   CH-8092 Zurich, Switzerland
                {kurmann,rauch,tomstr}@inf.ethz.ch

Clusters of commodity PCs (CoPs) offer good compute performance at 
low costs, but so far the platform can not utilize the potential of
Gigabit/s networking running with existing network adapters like
Ethernet NICs or standard protocols like TCP/IP. The speed of Ethernet
has grown from 10 Mbit/s to 1 Gbit/s over the last decade, but the
functionality and the architectural support in the network interfaces
remained almost the same, leaving the memory system as the primary
bottleneck. Therefore for good communication speeds in applications a
strict ``zero-copy'' communication architecture is required.

All widely used protocol stacks use at least one copy for the 
(de)fragmentation of all network packets transferred, since the
standard Ethernet packets are smaller than a memory page. Defragmenting 
packets of various communication protocols in hardware remains an 
extremely complex task. We therefore consider a speculative 
defragmentation technique, that can eliminate the last defragmenting
copy operation in regular TCP/IP stacks on existing Ethernet hardware.

For the evaluation of our ideas, we integrated a network interface
driver with speculative defragmentation into an existing TCP/IP 
protocol stack and added some well known zero-copy strategies like page 
remapping and fast buffers. With all these ideas together, we improved 
the performance for Gigabit Ethernet over a standard Linux 2.2 TCP/IP 
stack by a factor of 1.5-2 for uninterrupted burst transfers in 
conventional TCP/IP streams. Furthermore, our study investigates the 
rate of speculation success in application codes and predicts quite 
favorable success rates for database and scientific codes.