[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: transaction backoff problems (ITS#2741)

> While running test008 with a large number of iterations (-l 2000) I see
> operation throughput decrease over time. There are long idle periods when
> slapd threads are asleep, and then a short burst of activity accompanied
> DEADLOCk messages, and then another long idle period. At least one thread
> to be active at all times, but it appears that all the writers have gone
> sleep, for approximately equal amounts of time.

Looks it is related to the transaction backoff patch.
In fact, the paramters of the transaction backoff themselves need to
be self adjusted. For example, with a parateter set empirically optimized
for back-bdb does not work well with back-hdb, because they have
different intensity of contention.
Maybe higher randomness or throttling is needed, but I guess that this is
the case
in which the contention is too high, that even backoff doesn't help
(I'm not familiar with -I option, though).
Without backoff, we might see even more contention.

- Jong