Sometimes no avoiding fighting it out with support *** sigh *** Sent from my iPhone > On May 22, 2014, at 11:22 AM, "Rich Jesse" <rjoralist3@xxxxxxxxxxxxxxxxxxxxx> > wrote: > > Mark replies: > >> It is unclear to me why you do not proceed with the 15 indentified tables >> non-parallel, if that works repeatably. > > This is the first time I've seen this. My concern is that this is both > unpredictable and apparently undocumented. I'll be running this process > again tomorrow, so I'll be able to see which, if any, PKs are affected. > >> With so many objects available to run in separate threads, it is unclear to >> me that parallel is an advantage for any of this process. > > The largest table takes 4 hours to create its PK single-threaded. All 4600+ > PKs take 45m total with DOP of 8. No matter how I divide the list, it will > take a minimum of 4h if I don't use parallel. > >> In fact, if you decide to go that way, I would expect a significant >> reduction in the overall elapsed time. For tables that will fit in the >> buffer cache, you might want to run a scanner to get nice clean blocks in >> there just before you unleash the index creates for that table, with a >> thread for each index create on a cached table. > > I'm dying to test this on our future production DB server as the buffer > cache could nearly fit every table at the time of this part of the > migration. I <3 new hardware. :) I'll also have 2x more cores to throw at > it, but I'm wondering if that could make this problem worse with increased > contention. > > Thanks! > > Rich > > -- > //www.freelists.org/webpage/oracle-l > > -- //www.freelists.org/webpage/oracle-l