Method to synchronize EXA blade servers
Original Publication Date: 2004-Dec-13
Included in the Prior Art Database: 2004-Dec-13
In a typical scalable system the clock domains between systems that comprise an 8-way, 16-way or, 32-way are asynchronous. This creates a problem for software as timestamp information presented by different CEC's will not be consistent. This will lead to problems such as the following: 1)A task reads a timecounter on one system. Then, that task gets migrated to another CEC, reads the timecounter again and performs an action based on the time elapsed between the two reads. If the timecounter on the second system is "in the past," the task will error. 2)The system time of day clocks will drift because the input clocks will be at slightly different frequencies. The differing times have the potential to cause confusion as events may not be ordered correctly in system logs. This invention will prevent the problems described by forcing all blades to be in a common clock domain. This will allow all timestamp methods to be used in applications and will not require the intervention of an OS or Service Processor to properly synchronize clocks. The invention is primarily targeted to Blade server systems but could be used in other environments. The invention also has the benefit of increasing the reliability of the system by providing a redundant path for the system clock in all partition configurations.