DTJ503PF

This document details the development and implementation of an optimized transport interface for Digital's PATHWORKS for VMS version 4.0 server, aimed at significantly improving its file and print service performance for personal computers on LANs.

The core problem addressed was the high CPU overhead and poor performance, especially for small data transfers, inherent in the standard buffered I/O interface and variable transaction response times. To achieve optimization, the team selected the Local Area System Transport (LAST) protocol over DECnet, due to its client memory efficiency and Digital's control over its modifications.

The development process involved evaluating various transport interface models. They moved from a baseline standard VMS Queued I/O (QIO) interface to optimized variations like QIO/WAKE (optimized completion notification) and CMKRNL/WAKE (optimized initiation and completion). Key technical optimizations included the use of double-mapped buffers and shared queues to minimize data copying and system boundary crossings, along with a "transceive" operation for simultaneous read/write requests.

Challenges encountered included:

• The Microsoft LAN Manager redirector's two-phase read requests (header then data), which was resolved by a Buffer Management Protocol (BMP) layer that communicated client buffering capacity and managed data until the full request was made.
• The LAST protocol's lack of message acknowledgments, which was addressed by the BMP layer caching responses for retransmissions and using transaction identifiers to differentiate new requests from retransmissions.
• Incompatibility with standard NetBIOS applications (which don't follow a request/response model), solved by developing a common NetBIOS interface that routed redirector requests over LAST and other NetBIOS requests over DECnet.

Performance measurements confirmed the success of the optimization. Initial tests showed a significant reduction in transaction response times (e.g., from 2.2 milliseconds with standard QIO to 1.6-1.7 milliseconds with optimized models). Final benchmarks demonstrated that the optimized LAST protocol consistently achieved reduced elapsed times and comparable or lower CPU utilization for various workloads (including raw I/O and transaction processing) compared to the standard DECnet/QIO interface. The optimized interface also proved more resilient to concurrent I/O activity. While less impactful for server-I/O-bound tasks like directory copies, the optimizations significantly improved overall server performance for typical network file service scenarios.