Commit graph

282 commits

Author SHA1 Message Date
Sears Russell
86fab6a22b Turn on paranoid slotted page consistency checks for long running tests. 2006-07-22 00:05:40 +00:00
Sears Russell
2af892709f Rewrote slottedCompact; added fsckSlottedPage 2006-07-22 00:01:42 +00:00
Sears Russell
0725af91be Fixed page leak in Talloc(). 2006-07-21 01:07:09 +00:00
Sears Russell
eb13531c0f Partial handling of incomplete log entries. (It detects when the log file does not end at the right byte, but doesn't checksum, or otherwise look for
sectors that didn't make it to disk.)
2006-07-21 01:06:35 +00:00
Sears Russell
1ff1d4c128 Re-enabled truncation (accidentally disabled it before...) 2006-07-20 01:30:58 +00:00
Sears Russell
ffd3bd960b switched to region allocator, but TpageAlloc currently grabs its own region... 2006-07-20 01:29:39 +00:00
Sears Russell
7bc60c3ac4 fixed the fixedPageAlloc hack. 2006-07-20 00:49:28 +00:00
Sears Russell
8ed909e79d Fixed the fixedPageAlloc hack. 2006-07-20 00:48:42 +00:00
Sears Russell
1ac3dc6779 More progress on boundary tags. (Still not tested/working yet...) 2006-07-18 23:59:00 +00:00
Sears Russell
bf98e32c73 fixed build 2006-06-22 23:31:42 +00:00
Sears Russell
93e7e96ea3 Started implementing the region allocator 2006-06-22 19:10:02 +00:00
Sears Russell
d26263262b Completely new blob implementation. Slower than the old one was supposed to be, but is correct, and extremeley
simple.  It simply writes the pre-image and post-image of each blob to the log using TpageSet.
2006-06-21 00:45:18 +00:00
Sears Russell
c8c9b7af19 the logger understands special record types. (At least partially...); shortened blob tests, since the blob
implementation is fundamentally broken, and slow
2006-06-20 23:02:27 +00:00
Sears Russell
4d79c0a466 Simplified allocation 2006-06-20 21:40:21 +00:00
Sears Russell
e14877821f Fixed blob bug. The reason it's so slow is that it's logging a preimage + postimage of each blob and then synchronously writing to a second copy each time the blob is
written to(!!!)
2006-06-20 00:03:38 +00:00
Sears Russell
3e4cac0fe3 Continued cleaning up allocation code. It seems really slow all of a sudden, so this commit probably introduced a performance bug. 2006-06-17 00:25:09 +00:00
Sears Russell
209a6916e5 Updated some comments. 2006-06-16 00:27:02 +00:00
Sears Russell
dc8a3e5d60 Eliminated slottedPreRalloc 2006-06-16 00:05:44 +00:00
Sears Russell
50515ffd23 More steps toward factoring page allocation out of record allocation. 2006-06-15 05:31:20 +00:00
Sears Russell
59bd80a2a8 half-hearted attempt to fix -Wextra warnings... 2006-06-13 23:58:04 +00:00
Sears Russell
a3678aad5c Now AM_CFLAGS can be set by modifying GLOBAL_CFLAGS in the root
Makefile.am
2006-06-13 22:48:50 +00:00
Sears Russell
f53a46e788 removed heisenbug 2006-06-13 22:47:38 +00:00
Sears Russell
1b3a08d434 added some todos 2006-06-13 22:46:11 +00:00
Sears Russell
1cf3119eb7 renamed __really_do_ralloc to really_do_ralloc 2006-06-13 22:45:30 +00:00
Sears Russell
66ae3b9763 Made it more obvious that a suspicious looking assert is commented out 2006-06-13 22:44:40 +00:00
Sears Russell
24a5da9648 Fixed deprecated volatile static declaration (-> static volatile) 2006-06-13 22:43:58 +00:00
Sears Russell
33af9805ad printf -> fprintf(stderr, and assert(0) -> abort() 2006-06-13 22:43:20 +00:00
Sears Russell
f77e8b240d It was improperly tracking the log's global offset. (Hopefully)
fixed offset tracking, made it optional, and disabled it.
2006-06-13 22:42:15 +00:00
Sears Russell
c5add15bd3 assert(0) -> abort 2006-06-13 22:40:53 +00:00
Sears Russell
eab6385ca5 Moved #defines below definition of Page_s. 2006-06-13 22:40:24 +00:00
Sears Russell
f89903d529 moved stdio.h include out of header file. 2006-06-13 22:38:25 +00:00
Sears Russell
e56c46b95c Moved blob_record_t into transactional.h (part of transition to using
slot type lookup tables.)
2006-06-13 22:37:29 +00:00
Sears Russell
a11fbdd9e6 Did away with slottedPreRallocFromPage. 2006-06-13 20:00:46 +00:00
Sears Russell
cef9e2c4ce Added settings for longer-running tests 2006-06-12 22:45:41 +00:00
Sears Russell
e0b34633ec Fixed broken min_rec_lsn() semantics; before, if there were not outstanding transactions / writes,
then these calls would return flushedLSN().  The problem was that flushedLSN wasn't guaranteed to
return a pointed to the beginning of a log entry.
2006-06-12 21:50:21 +00:00
Sears Russell
7bb0ce0d17 Fixed truncation bug (truncation is more aggressive now), and added more debugging info. 2006-06-12 20:55:49 +00:00
Sears Russell
27f7df6f09 Fixed races on flushedLSN and page->dirty. 2006-06-12 20:05:28 +00:00
Sears Russell
35a5e80871 removed bogus .deps files. 2006-06-01 22:18:50 +00:00
Sears Russell
fd19036843 Added more complete hashtable test; fixed incorrect return value in LH_ENTRY(find) 2006-05-31 00:37:56 +00:00
Sears Russell
0649f179b3 Got rid of the ~1 second delay on exit. 2006-05-30 22:59:22 +00:00
Sears Russell
68dcc2067e No more warnings with -O2. 2006-05-25 22:50:52 +00:00
Sears Russell
8ce9e6127f Fixed return value in removeFromLinkedList 2006-05-25 22:48:15 +00:00
Sears Russell
b5d578f4e0 Fixed remaining warnings, got rid of alTupdate 2006-05-25 20:08:12 +00:00
Sears Russell
fda416090b Fixed a bunch of 'easy' warnings 2006-05-25 00:02:46 +00:00
Sears Russell
cb3a4ae20d Bucket list extension working. 2006-05-24 23:24:31 +00:00
Sears Russell
8b54b0d265 No longer attempts to truncate when long transactions prevent it from
being useful...
2006-05-24 23:21:33 +00:00
Sears Russell
827e0d891a The linear hashtable implementation (to replace pbl's) 2006-05-24 02:23:58 +00:00
Sears Russell
2e06cafaa8 2006-05-24 02:19:04 +00:00
Sears Russell
4c168d8fb2 Fixed truncation bugs, and a minor test case bug. 2006-05-22 23:59:02 +00:00
Sears Russell
f06159b3cd forgot to add these... 2006-05-19 20:17:44 +00:00
Sears Russell
ef43d5e4ce Implemented log truncation. 2006-04-14 03:45:26 +00:00
Sears Russell
11c311bc91 Added support for raw pages, implemented posix_memalign workaround, and re-enabled O_DIRECT by default 2006-04-12 01:41:35 +00:00
Sears Russell
33cbbb0e2d In memory logging implementation 2006-04-12 01:40:47 +00:00
Sears Russell
ab541b7dc8 Created bufferPool, moved memory management for page structs to it 2006-04-11 20:35:21 +00:00
Sears Russell
c29b16a6ed Created new module, bufferPool, and moved memory management for page structs to it. 2006-04-11 20:34:52 +00:00
Sears Russell
c1d8906d28 removed memcpy() calls from inMemoryLog; added "const" qualifier to many LogEntry pointers. 2006-04-11 02:20:21 +00:00
Sears Russell
7decae7172 All tests that pass with LOG_TO_FILE now pass with LOG_TO_MEMORY, but LOG_TO_MEMORY is slow, and seems to be
killing malloc's performance for some reason.  (linearHashNTA spends many seconds in free() between unit
tests...)
2006-04-08 00:05:09 +00:00
Sears Russell
d3d7f2c788 Implemented "LOG_TO_MEMORY", which mostly works (linear hash NTA goes into an infinite loop....) 2006-04-07 03:40:40 +00:00
Sears Russell
1c979f3052 Factored logWriter.h out of the rest of the system. The only file that directly depends on it is logger2.c, which can
now dispatch requests to different (hardcoded) log implementations.
2006-04-06 03:22:57 +00:00
Sears Russell
10e7434d32 corrected slottedFrespace()'s declaration (it was returning an int, but needed to return a size_t 2006-04-05 02:54:14 +00:00
Sears Russell
d9f1df0476 Fixed behavior when __really_do_ralloc() is called, but the record's slot is greater than numslots_ptr().
Optimized slotted_compact and __really_do_ralloc() so that they try to put lower numbered slots earlier in the freelist.
2006-04-05 02:52:40 +00:00
Sears Russell
d896155b82 Fixed freespace measurement during recovery/abort by adding slottedFreespaceForSlot()
(This version is still broken, because __really_do_ralloc() does not properly increment numslots_ptr
2006-04-05 00:38:11 +00:00
Sears Russell
041bfce736 The bucket list entries for variable lenth hash tables were too long, causing memory corruption, and wasting space+log bandwidth 2006-04-04 22:34:35 +00:00
Sears Russell
959233566c changed type of lsn variables to "lsn_t" 2006-04-04 22:33:30 +00:00
Sears Russell
8e0bd2ecd2 Extra check for end of iterator. 2006-04-04 22:32:48 +00:00
Sears Russell
2250b7a6c7 More freespace troubles... Should be correct now. 2006-03-28 21:27:10 +00:00
Sears Russell
3c55daede9 The freespace calculations were double counting the new slot's header. 2006-03-28 21:21:47 +00:00
Sears Russell
ced022d0fe partial fixes 2006-03-28 20:15:31 +00:00
Sears Russell
bb37db0857 Partial port to amd64. 2006-03-20 23:11:46 +00:00
Colleen M. Lewis
5bfcbdb076 Bug fixed: TfixedPageAlloc calls fixedPageInitialize with the wrong arguments. 2005-10-31 23:21:18 +00:00
Colleen M. Lewis
345f0cde78 Added the base functionality and test case for the BTree... just a hello world method. but it has the h, c and check files. -colleen ;) 2005-05-05 21:34:12 +00:00
Sears Russell
6b95cefc62 Fixed logWriter abort performance bug, also cleaned up the file, simplified locking, etc. 2005-04-14 04:23:22 +00:00
Sears Russell
a053b19757 removed extraneous abort 2005-04-14 01:43:14 +00:00
Sears Russell
3844b641a3 final transclos benchmark 2005-03-25 16:33:36 +00:00
Sears Russell
08ceca11ce Enabled compactPage in TallocFromPage() 2005-03-24 11:04:06 +00:00
Sears Russell
4c1ca194d0 Major edits throughout the paper. 2005-03-24 10:00:08 +00:00
Sears Russell
cf58e1fb72 Transitive closure working! 2005-03-24 04:43:22 +00:00
Sears Russell
2f99ed4664 Generic consumer wrapper functions, and implementation registrations. 2005-03-23 08:10:30 +00:00
Sears Russell
ee9eb416ea Implemented fifo pool that allocates entries out of a pool (it should be easy to extend this to avoid deadlock, while allowing it to continue to provide backpressure) 2005-03-23 08:07:53 +00:00
Sears Russell
6d4216bc2a Preliminary (broken) transitive closure implementation. 2005-03-21 08:04:29 +00:00
Sears Russell
ef3c30b190 Benchmark updates. 2005-03-20 05:17:25 +00:00
Sears Russell
26d50f19e3 Cleaned up berkeley db tests / lladd logger optimizations + bug fix. 2005-03-19 20:28:30 +00:00
Sears Russell
d8eb6c9aa7 Major benchmark update; log force merging. 2005-03-17 04:38:55 +00:00
Sears Russell
bf4be359b3 Partially refactored logMemory and multiplexer into a fifo implementation. Added support to fifoPool for
asynchronous threads running round-robin through the dirty fifos consuming entries in idle ones.
2005-03-16 00:52:21 +00:00
Sears Russell
dffff067bb multiplexer and fifo are now working. (also added removed some timeouts from tests that were timing out) 2005-03-15 05:30:57 +00:00
Sears Russell
ed9b978bda Another commit on behalf of jimmy. (Fixing the build is hard!) 2005-03-15 01:57:43 +00:00
Sears Russell
602608b60b Adding new files for jimmy... 2005-03-15 01:29:04 +00:00
Jimmy Kittiyachavalit
80ed16c18e Initial implementation of logMemory.c. Haven't yet done any testing, but it seems to compile. However, as of now, the current source code in the cvs repository does not build, dying on check_<somehting (I think ringbuffer.x)>, due to an a reference to an undefined symbol 'mutex'. 2005-03-14 20:54:06 +00:00
Sears Russell
7971df14bf Bugfixes (compiles now) 2005-03-14 08:47:13 +00:00
Sears Russell
9be3c4edd3 Initial upload of (untested) multiplexer implementation. 2005-03-14 00:58:47 +00:00
Sears Russell
ed49949512 This commit breaks the build so that Jimmy and I can sync. Sorry. :( 2005-03-14 00:25:29 +00:00
Sears Russell
1bfb634103 Preliminary benchmark program uploads for SEDA-style response time percentiles vs requests / second
Also added a few cute little utilities:
  - truncate_log
  - run_recovery

Upped max # of concurrent transactions.  (Current throughput during benchmarking remains ~flat up to 32,000
threads, but past max concurrent transactions was 1,000...)
2005-03-12 22:11:18 +00:00
Sears Russell
18c772234d new benchmarks revealed that inserts for linkedList, pageOrientedList were taking O(n) time, yielding O(n^2)
time to build a list, since the lists could not contain duplicate keys.  These operations are now O(1), and
the lists can contain duplicates.
2005-03-10 20:10:49 +00:00
Sears Russell
f28df29736 Fixed unit tests, added iterator interface, and generic (almost) iterator test harness. 2005-03-10 03:19:04 +00:00
Sears Russell
935de3aeb5 minor changes 2005-03-09 03:24:36 +00:00
Sears Russell
e22d4b8e59 Partial implementation of ringbuffer for in-memory long, among other things. 2005-03-08 07:53:53 +00:00
Sears Russell
a3112ee81c Fixes for freebsd. 2005-03-02 05:47:38 +00:00
Sears Russell
7833177f05 Fixed the support for more than MAX_TRANSACTION total transactions. 2005-03-02 03:28:00 +00:00
Sears Russell
3e5f7aea8a Bugfixes for OASYS. 2005-03-01 07:32:02 +00:00
Sears Russell
3b88909c5f Fixed compensations test; FreeBSD fix (?) for makefile. 2005-02-28 00:00:17 +00:00
Sears Russell
890a7385d0 Fixeds cht test / finished error checking within LLADD, but haven't tested error handling yet. 2005-02-24 21:12:36 +00:00
Sears Russell
29dacbe7ed Error handling for linked list. 2005-02-22 23:11:03 +00:00
Sears Russell
41fb85eef0 More error handling... 2005-02-22 03:10:54 +00:00
Sears Russell
84a20a3c96 Started adding error handling code to LLADD. 2005-02-16 04:11:14 +00:00
Sears Russell
f4484e35b5 Update for perl script; added return type handling. 2005-02-16 02:13:54 +00:00
Sears Russell
fdf5344ec3 Implemented 'compensations' for C. They're not proper compensations, but they're more flexible than pthread's mechanisms, as they allow the stack to be partially rolled up. (Although calling functions need to either check for errors and return manually, or wrap calls to error-producing callees in a begin_action{ }end_action block. 2005-02-14 02:49:59 +00:00
Sears Russell
2ac1302062 Recovery now supports the lock manager. (And vice-versa ;) 2005-02-10 21:56:32 +00:00
Sears Russell
8bf6ea29ff Added @todo that details requirements for next version of alloc. 2005-02-10 21:55:35 +00:00
Sears Russell
b8a67cbfb2 Numerous lock manager fixes; more realistic test case for lock manager, added support for optional page-level locking.(Still need to report deadlocks back up to the caller.) 2005-02-10 03:51:09 +00:00
Sears Russell
6f022cc577 Some random code cleanups, and implemented a naive lock manager. (Still need to tie it into LLADD...) 2005-02-09 02:53:14 +00:00
Sears Russell
258dd5f4c4 Fixed up some return values; tests now pass with assert disabled. 2005-02-08 01:13:57 +00:00
Sears Russell
54ecc0d3c4 page->dirty is now set correctly (assuming LSN's are updated properly). pageFile.c makes use of this fact by ignoring write requests on clean pages. 2005-02-02 02:12:40 +00:00
Sears Russell
a4c13449b8 Bugfix for pageOrientedList. 2005-01-31 02:18:46 +00:00
Sears Russell
1260710a27 Got rid of another broken hash implementation... 2005-01-31 01:34:29 +00:00
Sears Russell
bb69197a65 Fixed blob implementation (it was impossible to distinguish between a blob and a record of length 12), removed jbhash, and ported cht to linearHashNTA. 2005-01-31 01:29:52 +00:00
Sears Russell
7cf5fdee6e Re-implemented pageOrientedListNTA from scratch. It's now more flexible, but relies on features that are not provided by Talloc/Tdealloc, so performance isn't as good as it could be. 2005-01-29 01:17:37 +00:00
Sears Russell
4d04155b0e An optimized, but still slow implementation of pageOrientedList. 2005-01-29 01:09:29 +00:00
Sears Russell
56d98adcbf Working, but slow and non-concurrent variable length key/value support for linear hashtable. 2005-01-28 21:28:23 +00:00
Sears Russell
17f76264aa Implemented pageOrientedList, also did some bug fixing, and added some support functions to alloc.c 2005-01-28 03:32:17 +00:00
Sears Russell
05934d296a Removed O_DIRECT since it breaks on 2.4 kernels. Need to figure out how to
detect this in the build script so that systems that support O_DIRECT can use it.

(Note:  RedHat kernels silently fail on O_DIRECT)
2005-01-24 19:58:09 +00:00
Sears Russell
b04e71360f Documentation update. 2005-01-20 21:19:47 +00:00
Sears Russell
54edde96c2 Possible fix for older 2.4 kernels. (Also should improve performance...) 2005-01-20 00:26:09 +00:00
Sears Russell
d7b846d370 CHT is starting to work. 2005-01-19 00:47:55 +00:00
Sears Russell
894e30085c libdfa ported to new hash table implementation. (Uses nested top actions.) 2005-01-15 23:55:49 +00:00
Sears Russell
ac0d413e58 - linkedListNTA is now re-entrant.
- linearHashNTA now uses nested top actions, is re-entrant, and supports a
  non-re-entrant iterator
- 'recursive' nested top actions allow nested top actions to be nested within each
  other.
2005-01-15 01:45:27 +00:00
Sears Russell
7a990886d2 The linked list implementation now uses nested top actions. (Also, many bugfixes regarding nested top actions were introduced.) 2005-01-14 10:08:10 +00:00
Sears Russell
360f0d15e2 Working, non-concurrent linear hash and linked list implementations.
The idea is to implement the non-concurrent versions, and then convert
them to concurrent versions using nested top actions.
2005-01-14 01:52:53 +00:00
Sears Russell
6198522971 Starting to rip the old jbhash implementation out, and replace it with a
linear hash implementation.  In particular, I've implemented (sort of) an
iterator for linear hash, and (sort of) ported libdfa to linearhash.c

I say 'sort of' since this functionality is broken with this commit.  On the other hand, CVS should build now, and the tests under lladd will pass.  libdfa's new tests don't pass at the moment.
2005-01-10 03:02:02 +00:00
Sears Russell
55b0ddf1b6 Fixed nasty bug that was causing the LLADD header to be lost (!!) 2004-12-06 01:20:48 +00:00
Sears Russell
72070acb67 Implemented TsetRange, fixed bug in linear hash. 2004-12-03 00:27:47 +00:00
Sears Russell
443a90ad7c Implemented TrecordType() 2004-12-01 01:26:25 +00:00
Sears Russell
ded5c92c9d Support for nested top actions, with simple unit test. (Haven't tested for thread safety yet.) 2004-11-29 21:28:13 +00:00
Sears Russell
646f9dfca9 Implemented recover_split(), which is needed by the linear hash to recover properly from crashes. It hasn't been tested at all, and is not currently called by anything. 2004-11-29 02:35:37 +00:00
Sears Russell
2c23fb8a43 Did some misc. cleanup work on linearHash 2004-11-28 01:16:48 +00:00
Sears Russell
164c7cfd69 Linear Hash now survives Tdeinit/Tinit cycle (but not crashes!) 2004-11-24 23:26:15 +00:00
Sears Russell
80d4af4c9b Added getRecordType() 2004-11-24 23:25:36 +00:00
Sears Russell
910649a387 Web-site update, a few minor code changes from before the Freenix deadline. 2004-10-27 01:40:09 +00:00
Sears Russell
aead6a5a83 New benchmark, fixed for PROFILE_LATCHES. 2004-10-20 21:55:00 +00:00
Sears Russell
af9e7adee0 Starting work on the actual graph-generation portion of the benchmarks. :) 2004-10-20 06:21:21 +00:00
Sears Russell
903ba4b97b Bugfix. 2004-10-19 22:58:56 +00:00
Sears Russell
ec99c0e258 Lots of optimizations; re-did on-disk layout for the linear hash table. 2004-10-19 21:16:37 +00:00
Sears Russell
6b265f28b0 Speedups, most notably in the logging subsystem. 2004-10-19 04:45:42 +00:00
Sears Russell
c850ce3a51 The logger now calls fwrite half as often as it used to. This could be improved significantly, but some latching details internal to the logger need to be worked out first. 2004-10-18 19:19:10 +00:00
Sears Russell
ec1276f26f Preliminary version of logical logging linear hash. (No latching yet, and there are some bugs re-opening a closed hash.) 2004-10-18 18:24:54 +00:00
Sears Russell
04cfc7da8e Optimized the 'naive' linear hash table, and moved to sourceforge. :) 2004-10-17 02:17:00 +00:00
Sears Russell
a078f25475 Implemented a linear hash from scratch. Concurrency hasn't been done, but it's probably correct in the single thread case. 2004-10-12 02:44:47 +00:00
Sears Russell
5064e3fac2 Implemented pages that store fixed sized records efficiently, and a java-style ArrayList data structure that efficiently supports (relatively) clustered, O(1) access time expandable arrays. (This will be used for the hash implementation...) 2004-10-06 06:08:09 +00:00
Sears Russell
96e7af92a4 Added test cases for Tprepare(), implemented some redo-only operations, and started to clean up record allocation/deallocation.
Also, numerous bugfixes.
2004-10-02 07:29:34 +00:00
Sears Russell
811bc5c710 Implemented a freespace manager that should safely allocate space, even in the face of crashes, and can reclaim unused space (unless an application opens more than one simultaneous transaction that performs allocations)
Fixed some blob bugs (by adding extra fdatasync() calls).

Began factoring out the page management code so that it is an extenstion, and a less integral part of lladd.
2004-08-21 00:03:30 +00:00
Sears Russell
b4d7883f66 Fixed regressions, major update of page / slotted interfaces in anticipation of moving LLADD towards a generic transactional page system. 2004-08-17 01:46:17 +00:00