In releases prior to Ingres II 2.5, a fixed number of Write Behind threads were configured in each server in an installation and initiated when the server started. These threads served all caches and were awakened when the number of modified pages in any cache exceeded a predefined threshold. In a shared cache environment, all Write Behind threads in all servers were simultaneously activated when this threshold was reached. This led to a “thundering herd” phenomenon in which n Write Behind threads concurrently pounded through the caches, competing for modified pages to flush.

The optimum number of Write Behind threads is the minimum number required to:

The optimum number of Write Behind threads varies with the instantaneous demand for free pages in a particular cache; it always begins at one when the threshold is first breached and a Write Behind event signaled. In Ingres II 2.5, if the single Write Behind thread is unable to keep up with the demand, additional Write Behind threads are created until either equilibrium is achieved or the upper limit on thread numbers is reached. If better than equilibrium is achieved (more modified pages are being freed than are in demand), the excess Write Behind threads terminate one-by-one, while the remaining threads continue to monitor the free buffer demand to achieve the write-behind end threshold.

Ingres II 2.5 introduces cache-specific Write Behind threads, resulting in increased concurrency and eliminating the chance of free page starvation.