I don't think it really matters at all and that it is simply not something to be concerned about. Did you notice the "ela=" values for the "direct path read" wait-events? They were quite a bit faster than those for the "db file scattered read" events. Instead of investigating the buffering of "direct path read" into the PGA, perhaps it would be more profitable to worry about the elapsed time per "db file scattered read" into the SGA? I'll bet that there is nothing that you can do about it, as the more heavily mutexed SGA is bound to be "slower" to access and modify than the non mutexed PGA, but it might still be something worth investigating... on 5/17/05 4:02 PM, jaromir nemec at jaromir@xxxxxxxxxxxx wrote: > Hi Tim, > >> I would guess that there is only room in those >> read buffers in each P00x process's PGA for sixty-one 16-Kb buffers > > > Exactly, I set the dfmrc to 64, so there are two reads per extent, 61 blocks > and 2 blocks. > > Lowering dfmrc to 10 there are 6 x 10 + 1 reads. > > I varied pga_aggregate_target between 10M and 50M with no effect on this > behavior. > > I'll examine the influence of parallel_execution_message_size but this will > take some preparation. > > It is not really a big issue , but it is *not nice*. -- //www.freelists.org/webpage/oracle-l