Details
-
Type:
Bug
-
Status: Resolved
-
Priority:
Major - P3
-
Resolution: Fixed
-
Affects Version/s: None
-
Fix Version/s: WT2.7.0
-
Labels:None
Description
The current direct write path in logging, where the read/write lock was removed, is causing issues. I'm going to redo how unbuffered log records are accounted in order to eliminate the direct write path.
Issue Links
- is depended on by
-
SERVER-20060
WiredTiger changes for MongoDB 3.1.8
-
- Closed
-
- is duplicated by
-
WT-2082
Logging illegal seek failure
-
- Resolved
-
Activity
Author:
{u'username': u'sueloverso', u'name': u'Susan LoVerso', u'email': u'sue@wiredtiger.com'}Message: WT-2094 Needed to rebuild the stat files.
Branch: develop
https://github.com/wiredtiger/wiredtiger/commit/5bc599052dfa234415a14740d233ead2245a50fe
Author:
{u'username': u'sueloverso', u'name': u'Susan LoVerso', u'email': u'sue@wiredtiger.com'}Message: WT-2094 Force unbuffered path more often.
Branch: develop
https://github.com/wiredtiger/wiredtiger/commit/0f76bd0e1aac73c9904a3efc4e1193384504cdff
Author:
{u'username': u'sueloverso', u'name': u'Susan LoVerso', u'email': u'sue@wiredtiger.com'}Message: WT-2094 Add unbuffered statistic.
Branch: develop
https://github.com/wiredtiger/wiredtiger/commit/2d58c0a76bce34a2636a238fc90f53b8704f4e37
Author:
{u'username': u'michaelcahill', u'name': u'Michael Cahill', u'email': u'michael.cahill@mongodb.com'}Message: Merge pull request #2180 from wiredtiger/log-unbuffered
WT-2094 Eliminate direct write and record unbuffered log records that
Branch: develop
https://github.com/wiredtiger/wiredtiger/commit/f883d27375b2ef8966a3ec82c63ce9a63c9ad432
Sue LoVerso, I've merged the branch, thanks!
I take it the bug yesterday was not tracking the size of unbuffered writes correctly?
I ran the log-append.wtperf workload with varying value sizes. In the case of using 130000 value sizes, where every record will exceed our buffer, the new tree is slower by about 40% compared against an old-develop (changeset 12044d2 - before any logging changes at all). In the case of moderate sized records (24000) the two runs are within 1-2% of each other. The current log-append.wtperf using default value sizes (100 bytes) runs about 40% faster in this branch. These were run on my Mac.
The improvements for small log writes look great! With large writes, do you understand why they are slower? We already had fixed-sized log buffers in revision 12044d2, right?
Issues
WT-2082,WT-2084,WT-2087seem to all be related to direct writes and race conditions after removing the read/write lock. I believe with a few changes, the current unbuffered record path can accomplish the same thing.