Horst F. Wedde, Jon A. Lind und Guido Segbert
Achieving Internal Synchronization Accuracy of 30 ms Under Message Delays Varying More Than 3 msec
WRTP’99, 24th IFAC/IFIP Workshop on REAL TIME PROGRAMMING, 1999
In order to support real-time distributed experiments, in particular for safety-critical applications such as studied in our MELODY project, featuring task execution times of 20 – 50 msec, we could tolerate a clock synchronization accuracy of ca. 100 µsec between a distinguished (master) clock and any other system clock. Due to the lack of driver software in AIX for running GPS on IBM RS/6000 machines (models 570 and 580) – which would have resulted in a maximum deviation of 200 nsec - we investigated the message delay behavior in the Token Ring and Ethernet architectures in order to determine and compare clock drifts between different machines. Based on the insights gained we defined a novel two-step distributed clock synchronization protocol. First a clock at a master site initializes a synchronization procedure affecting the remainder (slave) clocks, and ensuingly only local updates are at times performed during a distributed experiment adding to making our internal synchronization non-intrusive. As a key feature a novel masking technique is employed through which messages with “undue“ delays have no effect. Thus, in spite of a variation of message delays ≥ 3 msec the protocol guarantees a maximum deviation of 30 µsec between the master and any slave clock in the system. The approach and experimental findings are discussed in detail.