- Feb 04, 2014
-
-
Alex Gorrod authored
-
- Feb 03, 2014
-
-
Alex Gorrod authored
-
Susan LoVerso authored
-
Susan LoVerso authored
-
Susan LoVerso authored
-
Susan LoVerso authored
-
Michael Cahill authored
-
Alex Gorrod authored
-
- Feb 02, 2014
-
-
Alex Gorrod authored
-
Michael Cahill authored
zlib fixes, including changes to the raw compression interface so that the callback can give up gracefully when compression doesn't work.
-
Michael Cahill authored
Only retry compress_raw if it returns EAGAIN. Otherwise, if no rows were compressed, assume that compression failed and fall back to standard compression.
-
Michael Cahill authored
Make the first block sent to compress_raw fill a disk allocation block. If it doesn't compress, there is now a way for the callback to indicate that.
-
Keith Bostic authored
-
- Feb 01, 2014
-
-
Keith Bostic authored
value (new in Helium 2.2), for the store instead of maintaining our own statistics (which implies updating a shared cache line). This is only lightly tested, it looks to me like we're triggering a deadlock inside of the Helium library. Don't update the cache-inuse flag if it's already set, avoid writing a shared cache line to no purpose.
-
Keith Bostic authored
any changes from the cache to the primary, that pass would be holding locks and it's not likely to help things much.
-
Keith Bostic authored
-
Keith Bostic authored
-
Keith Bostic authored
just insert.
-
Keith Bostic authored
-
Keith Bostic authored
any compression style is configured, turn on Helium compression. Yeah, it's lazy, but that's how I roll.
-
- Jan 31, 2014
-
-
Keith Bostic authored
in raw-test mode, it's a lot simpler to load two compressors, the normal one and a raw-test version.
-
Keith Bostic authored
configuration to bzip-raw. Leave bzip raw in place, it deliberately tries to hit some of the edge cases around raw compression, which can't hurt.
-
Keith Bostic authored
could be compressed" isn't the same as "don't try again".
-
Keith Bostic authored
don't in adjacent paragraphs, talking about the same variables.
-
Keith Bostic authored
performance reasons.
-
Keith Bostic authored
-
Michael Cahill authored
* always try to compress at least one item (WiredTiger makes the first "row" artificially large); * loop on all calls since the zlib API allows calls to work incrementally * reserve more bytes to finish up the last block, and fall back to "normal" compression if the reserved space is insufficient.
-
Michael Cahill authored
* when shuffling entries around, ranges can overlap (found by valgrind). Use memmove rather than memcpy. * use the allocation size to figure out which keys can't be split: with 512 byte leaf pages, the first offset was always near 1KB so compression almost always failed. * provide a way for raw compression to indicate that compression failed so that WiredTiger immediately falls back to compressing the entire page, rather than calling raw compression again with more records.
-
Michael Cahill authored
-
Alex Gorrod authored
-
- Jan 30, 2014
-
-
Alex Gorrod authored
-
Alex Gorrod authored
Improve in-memory merge efficiency when there is a lot of forced eviction
-
Susan LoVerso authored
-
Keith Bostic authored
just calculate us.
-
Keith Bostic authored
-
Keith Bostic authored
configuration argument depends on where WiredTiger is installed, not where the compression library is installed. Don't try and specify the directory from which configure is being run. List snappy first, bzip2 is a less likely choice. Remove mention of the test suite to confirm compression is working.
-
Michael Cahill authored
Remove WT_MERGE_MAX_REFS -- it was always arbitrary, and now when we create an internal page with more than 100 children, it won't be marked split merge, so we won't merge it again.
-
Michael Cahill authored
Set the memory_page_max of LSM chunks to double the chunk size.
-
Michael Cahill authored
-
Michael Cahill authored
Set the memory_page_max of LSM chunks to double the chunk size, so that application threads don't try forced eviction as soon as the "no eviction" flag is cleared by the LSM worker thread. refs #859 (discovered during testing)
-