From: Leyne, S. <Se...@br...> - 2007-01-04 23:40:27
|
Ann, > The bookkeeping is horrendous, You are doing a great job of teaching me this! ;-) > and the results of an error are a mess, > so I'd suggest we try some tests not worrying about precedence (and > if the test database goes corrupt in a crash, so what) and be sure > that group writes are a big performance gain before putting a lot of > thought into which pages can in fact be written as a group. Wouldn't a purely theoretical test app to measure the difference of Overlapped IO be the first step, rather than create a private build which incorporates the necessary changes to process the current writes in a batch? > I'd want > a big gain because we'll lose some efficiency by partitioning the > page writes into a safe order. Isn't the engine already performing this partitioning task to some degree? > > However, since the engine determines the type of page being created or > > the change being made, it is possible for the engine to set a > > "precedence-order" value for each page. >=20 > It keeps a precedence graph to recognize that page 100 depends > on page 102 which depends on 103 which depends on 100. The > precedence information is carried in the buffer descriptor for > the page. How does the engine currently resolve this circular relationship, when it comes to careful writes? Sean |