 andrew
|
68bd39e130
|
Put skip list in its own shared lib
Now we can right benchmarks against one api, and just change the library
path
|
2024-02-27 12:02:01 -08:00 |
|
|
andrew
|
25a1226667
|
Fix setOldestVersion bug
We should update the oldest version even if we don't have enough
keyUpdates to do
|
2024-02-27 10:59:59 -08:00 |
|
|
andrew
|
d4d2dbcbda
|
Strip off leading {P,L}<tab>
|
2024-02-27 10:29:59 -08:00 |
|
|
andrew
|
6d9b35396f
|
Rename to real_data_bench
|
2024-02-26 19:37:15 -08:00 |
|
|
andrew
|
0c177fb40f
|
Measure throughput in RealDataBench
|
2024-02-26 19:34:28 -08:00 |
|
|
andrew
|
f38198a39d
|
Amortize cost of pausing gc in setOldestVersion
|
2024-02-26 19:23:29 -08:00 |
|
|
andrew
|
5f7789128e
|
Add RealDataBench.cpp
|
2024-02-26 17:17:07 -08:00 |
|
|
andrew
|
3f45535f89
|
Fix setOldestVersion
We were not accumulating keyUpdates strictly faster than we were
spending them for a monotonically increasing workload.
|
2024-02-26 13:51:03 -08:00 |
|
|
andrew
|
2b0dbabb5c
|
Add public showMemory symbol (when SHOW_MEMORY=1)
|
2024-02-26 13:48:20 -08:00 |
|
|
andrew
|
be7f643f14
|
Add Node1 and SHOW_MEMORY
|
2024-02-26 12:25:06 -08:00 |
|
|
andrew
|
6a08bdd40e
|
Add bitSet asserts for Node256 paths
|
2024-02-26 10:58:16 -08:00 |
|
|
andrew
|
333ac74d91
|
Use forEachInRange for overfill for Node48
|
2024-02-26 10:55:50 -08:00 |
|
|
andrew
|
17ac9b38fb
|
Use forEachInRange in setChildrenParents
|
2024-02-26 10:37:28 -08:00 |
|
|
andrew
|
172dd40648
|
Check for full words on boundaries
|
2024-02-26 10:36:45 -08:00 |
|
|
andrew
|
c97c7eee8e
|
Speed up sparse queries
|
2024-02-25 23:08:01 -08:00 |
|
|
andrew
|
9fcfc44dc3
|
Add forEachInRange
|
2024-02-25 21:04:39 -08:00 |
|
|
andrew
|
0d3475e229
|
Use uint64_t array and fewer branches in BitSet
|
2024-02-24 18:20:00 -08:00 |
|
|
andrew
|
70e3377eac
|
Update benchmark in README.md
Don't try to make sense of the diff - the benchmark itself changed too
|
2024-02-23 17:31:48 -08:00 |
|
|
andrew
|
adaa652d0d
|
Remove Node::maxVersion
|
2024-02-23 17:27:19 -08:00 |
|
|
andrew
|
921005edb3
|
Add childMaxVersion
|
2024-02-23 17:02:46 -08:00 |
|
|
andrew
|
d43a8a5907
|
Run benchmark in jenkins
|
2024-02-23 16:17:28 -08:00 |
|
|
andrew
|
588e8eb87f
|
Add monotonic increasing point writes workload
|
2024-02-23 16:13:53 -08:00 |
|
|
andrew
|
df8d092a84
|
Avoid some unnecessary work
|
2024-02-23 14:49:42 -08:00 |
|
|
andrew
|
438a2e2abc
|
Make sure setOldestVersion has work to do in bench
|
2024-02-23 14:41:05 -08:00 |
|
|
andrew
|
305c218888
|
Skip checking for partial key match if partial key len == 0
This saves instructions according to cachegrind
|
2024-02-23 14:27:20 -08:00 |
|
|
andrew
|
db60782c48
|
Use **self directly in insert
|
2024-02-23 14:09:04 -08:00 |
|
|
andrew
|
4f32ecc26e
|
Make "begin" a template parameter to insert
cachegrind says this saves instructions
|
2024-02-23 14:00:55 -08:00 |
|
|
andrew
|
f84aa88202
|
Copy old node to new more efficiently
Presumably it's generating good code for the memcpy with a static size,
and hopefully also eliminating redundant stores
|
2024-02-23 13:31:05 -08:00 |
|
|
andrew
|
921da1cb3f
|
Remove redundant code
|
2024-02-23 13:13:48 -08:00 |
|
|
andrew
|
14de4ee297
|
Specialize setChildrenParents for each node type
cachegrind says this increases the instruction count somehow. I'm not
sure what's going on there.
|
2024-02-23 13:09:42 -08:00 |
|
|
andrew
|
7136b5a450
|
Save more instructions in getOrCreateChild
Take advantage of the property that Node4 is a prefix of Node16
|
2024-02-23 12:39:24 -08:00 |
|
|
andrew
|
116c79d3de
|
Optimize getOrCreateChild
Add fast path for if it exists
Avoid some unnecessary branches
Try to simplify some
|
2024-02-23 12:34:28 -08:00 |
|
|
andrew
|
aaf0283f66
|
Prefer getChild to getChildGeq in happy path
Saves instructions according to cachegrind
|
2024-02-23 12:12:01 -08:00 |
|
|
andrew
|
259f47664a
|
Save some instructions in getOrCreateChild
|
2024-02-22 22:46:08 -08:00 |
|
|
andrew
|
9a47f2d03a
|
Fixed missed usages of longestCommonPrefixPartialKey
|
2024-02-22 16:24:26 -08:00 |
|
|
andrew
|
b5d5085fd5
|
Add precondition back to firstNeqStride
|
2024-02-22 16:16:42 -08:00 |
|
|
andrew
|
40dca168ba
|
Add precondition to longestCommonPrefixPartialKey
|
2024-02-22 16:07:20 -08:00 |
|
|
andrew
|
7a8233ac61
|
More unnesting
|
2024-02-22 15:50:54 -08:00 |
|
|
andrew
|
3c93b9a3ce
|
Unnest some in checkPointRead
|
2024-02-22 15:43:13 -08:00 |
|
|
andrew
|
c16feda9f8
|
Use longestCommonPrefixPartialKey in SearchStepWise
|
2024-02-22 15:41:39 -08:00 |
|
|
andrew
|
4c3e7aef30
|
Save a few instructions in SearchStepWise
|
2024-02-22 15:30:24 -08:00 |
|
|
andrew
|
7f5598af2b
|
Save a few instructions in addWriteRange
|
2024-02-22 15:25:50 -08:00 |
|
|
andrew
|
8016d44c04
|
Use simple loop for longestCommonPrefixPartialKey
The benchmark has this as faster
|
2024-02-22 14:49:04 -08:00 |
|
|
andrew
|
505c060a28
|
Use longestCommonPrefixPartialKey in insert
|
2024-02-22 14:46:52 -08:00 |
|
|
andrew
|
608b4fb6c7
|
Remove some branches
|
2024-02-22 14:29:24 -08:00 |
|
|
andrew
|
0fbc8b0190
|
Use longestCommonPrefix on partial keys
|
2024-02-22 13:57:00 -08:00 |
|
|
andrew
|
125ce88268
|
Suppress gcc warning
Accept that our usage is non-portable: https://stackoverflow.com/questions/3129916/what-is-wrong-with-this-use-of-offsetof
|
2024-02-22 12:37:14 -08:00 |
|
|
andrew
|
3a5db2d2ac
|
Share some Node4/16 and Node48/256 implementations
This cuts down on the number of instructions (confirmed with
cachegrind). Also avoid initializing some memory unnecessarily.
|
2024-02-22 12:31:10 -08:00 |
|
|
andrew
|
bd5d0259d9
|
Update corpus
|
2024-02-21 22:13:57 -08:00 |
|
|
andrew
|
09782ba833
|
Fix O(k^2) behavior when begin is a prefix of end
|
2024-02-21 20:11:37 -08:00 |
|