Copy Emulation in Checksummed, Multiple-Packet Communication

Data copying can be a bottleneck in end-to-end communication over high-speed networks. Emulated copy is an alternative I/O data passing scheme that preserves the API and integrity guarantees of copying but avoids the latter using VM manipulations – transient output copy-on-write (TCOW), input alignment, and page swapping. We characterize and evaluate the support necessary in network adapters for emulated copy in checksummed, multiple-packet communication. Our experiments on an ATM network show that: (1) Emulated copy gives performance better than that of copying even without hardware checksumming support; (2) TCOW improves multiple-packet output performance without any hardware support or changes in applications; (3) Page swapping provides additional similar improvements on multiple-packet input if there is input alignment, which requires either hardware support (earlydemultiplexed/system-aligned buffering) or changes in applications (pooled/application-aligned buffering); and (4) The performance of application-aligned buffering is largely unaffected by header/data splitting, a common optimization. We propose a new optimization, buffer snap-off, that extends system-aligned buffering to the general case of arbitrary, unmatched data transfer and application input buffer lengths.