[Date Prev][Date Next]
[Chronological]
[Thread]
[Top]
Re: (ITS#3611) Index clustering patch for fast slapadd
Howard Chu wrote:
> Why are you explicitly storing the number of IDs in the IDL, when the
> BDB library already maintains that counter itself?
Never mind. I see that since an ID cluster is written as a single
element, the BDB counter does not apply.
The idea of grouping writes and buffering/caching the groups is good,
but this approach still creates a fair amount of transaction log
overhead. I.e., even if only a few elements of a cluster are changed,
the entire cluster must eventually be flushed back to the database, and
since the cluster is fairly large, it will result in large transaction
log records.
If instead, you retained the current use of individual IDs, and buffered
a list of deltas such that only individual IDs needed to be flushed back
to the database, the transaction log overhead would remain small.
Meanwhile, flushing a large group of writes to an individual index all
at once would still minimize seeks and thrashing of the BDB cache.
--
-- Howard Chu
Chief Architect, Symas Corp. Director, Highland Sun
http://www.symas.com http://highlandsun.com/hyc
Symas: Premier OpenSource Development and Support