[progress Communities] [progress Openedge Abl] Forum Post: Re: Setting The -pica Parameter...

[George Potemkin]

BIW is running on the target. AI on the target is not enabled. The customer's case: a rather high transaction activity (~ 0.7 MB/sec of AI/BI writes) caused the persistent increase of the used message entries in the pica queue on source db. When the pica queue was totally filled we got 100% busy MTX and AIB latches. It means that almost every snapshot in promon shown an owner of these latches. MTX latch was owned by ABL users (each time a different one). AIB latch was always owned by AIW process. Just for a record: the competition for AIB latch between ABL users are not possible because they always lock AIB (and BIB) latch under cover of MTX latch: they lock MTX latch first and only then they're requesting AIB/BIB lock. I guess there are only a few special processes that can compete for AIB latch directly but mainly it's AIW. The number of the AIB latch locks was rather low for a busy latch: 700-2000 locks per sec. Under normal conditions the lock duration of AIB latch is less than 100 nanoseconds. Hence to make the latch to be 100% busy we need 10 millions locks per sec - it's impossible to get such high numbers of AIB locks in a real database. > The duration of a lock of the pica queue is less than a millisecond. It's a huge time compared to the duration of the "spin" latches like AIB. And it also means that we can create only about a thousand locks per sec (that matches the numbers above). New explanation of the customer's case: I guess the AIW process is requesting the pica queue lock while it holds AIB latch. When AIW is forced to wait for the pica queue lock then the duration of AIB latch is increasing by 10 thousands times - from 100 ns to 1 ms. ABL process that already owns the MTX latch is forced to wait for AIB latch. That is why the MTX latch became 100% busy at the same time with the AIB latch. The number of MTX/AIB/BIB latch locks are closed to each other - the difference between MTX and AIB latch locks used to be plus/minus 1%, the difference between MTX and BIB latches used to be plus/minus 10% (or more during the idle time). If this my "theory" is correct then the only solution is to have a large pica queue to avoid its total filling during the periods with highest transaction activity.

Continue reading...