[Hackathon2020 - Do not merge] Greedy-piecewise-linear-regression (PLR) to speed up int type key search for LSM-tree (bourbon idea) by liufuyang · Pull Request #211 · pingcap/badger

liufuyang · 2021-01-16T17:38:03Z

Disclaimer

Original implementation made by @spongedu at here https://github.com/spongedu/badger/tree/hackathon_go_plr_with_pointget_uint64key
I ported over and did some clean up work.

Idea from paper: From WiscKey to Bourbon: A Learned Index for Log-Structured Merge Trees, by Yifan Dai

Rust code source:

binary plr (which used in badger code) code here https://github.com/liufuyang/plr-test/blob/master/src/main.rs
plr library code here https://github.com/RyanMarcus/plr

Benchmark code here:

https://github.com/liufuyang/tbadger/blob/master/badger_test.go

With PLR, with pointGet

(fuyang-go-bench-badger *%) 🦀 👉  go test -bench=BenchmarkBadger  -benchtime=30s -run=^s
2021/01/16 18:36:38 plr opening file: data/00000004.sst
2021/01/16 18:36:38 data/00000004.sst
2021/01/16 18:36:38 plrSegment loaded: [{0 43371 0.002763512135689483 1.301027126230357} {43371 173143 0.001957717574102289 32.139598333991884} {173143 436947 0.0016452183457147617 82.19281535952956} {436947 833712 0.0014415577306141915 167.1788822889456} {833712 1.393807e+06 0.0013017392682151607 279.7244833388081} {1.393807e+06 2.144502e+06 0.0012110451657496644 402.0680961828125} {2.144502e+06 3.243884e+06 0.001120694858783678 591.6862955068373} {3.243884e+06 4.252725e+06 0.001064653793312162 769.4048384938992} {4.252725e+06 5.489191e+06 0.0010158619299836292 972.7984525022184} {5.489191e+06 7.260306e+06 0.0009696070011254119 1222.5316082781937} {7.260306e+06 8.683148e+06 0.000932753999857703 1485.8753435962026} {8.683148e+06 1.0541136e+07 0.0009026047455994166 1743.6066808302094} {1.0541136e+07 1.2621543e+07 0.0008782587667118885 1996.1946893039676} {1.2621543e+07 1.4617779e+07 0.0008541220294640251 2296.6850507190775} {1.4617779e+07 1.7976931348623157e+308 0.0008369829862892688 2543.212360918902}]
-------> Hit rate: 0.000000
goos: darwin
goarch: amd64
pkg: tbadger
BenchmarkBadger-12    	-------> Hit rate: 0.040000
-------> Hit rate: 0.042700
-------> Hit rate: 0.039912
-------> Hit rate: 0.039828
-------> Hit rate: 0.039901
40003507	       985 ns/op
PASS
ok  	tbadger	69.299s

WithoutPLR, with pointGet

(fuyang-go-bench-badger *%) 🦀 👉  go test -bench=BenchmarkBadger  -benchtime=30s -run=^s
2021/01/16 18:40:05 plr opening file: data/00000004.sst
2021/01/16 18:40:05 data/00000004.sst
2021/01/16 18:40:05 plrSegment loaded: [{0 43371 0.002763512135689483 1.301027126230357} {43371 173143 0.001957717574102289 32.139598333991884} {173143 436947 0.0016452183457147617 82.19281535952956} {436947 833712 0.0014415577306141915 167.1788822889456} {833712 1.393807e+06 0.0013017392682151607 279.7244833388081} {1.393807e+06 2.144502e+06 0.0012110451657496644 402.0680961828125} {2.144502e+06 3.243884e+06 0.001120694858783678 591.6862955068373} {3.243884e+06 4.252725e+06 0.001064653793312162 769.4048384938992} {4.252725e+06 5.489191e+06 0.0010158619299836292 972.7984525022184} {5.489191e+06 7.260306e+06 0.0009696070011254119 1222.5316082781937} {7.260306e+06 8.683148e+06 0.000932753999857703 1485.8753435962026} {8.683148e+06 1.0541136e+07 0.0009026047455994166 1743.6066808302094} {1.0541136e+07 1.2621543e+07 0.0008782587667118885 1996.1946893039676} {1.2621543e+07 1.4617779e+07 0.0008541220294640251 2296.6850507190775} {1.4617779e+07 1.7976931348623157e+308 0.0008369829862892688 2543.212360918902}]
-------> Hit rate: 0.000000
goos: darwin
goarch: amd64
pkg: tbadger
BenchmarkBadger-12    	-------> Hit rate: 0.060000
-------> Hit rate: 0.039500
-------> Hit rate: 0.039985
-------> Hit rate: 0.039894
-------> Hit rate: 0.039881
38633751	      1025 ns/op
PASS
ok  	tbadger	60.518s

WithPLR, without pointGet

(master %) 🦀 👉  go test -bench=BenchmarkBadger  -benchtime=30s -run=^s
2021/01/16 18:52:33 plr opening file: data/00000004.sst
2021/01/16 18:52:33 data/00000004.sst
2021/01/16 18:52:33 plrSegment loaded: [{0 43371 0.002763512135689483 1.301027126230357} {43371 173143 0.001957717574102289 32.139598333991884} {173143 436947 0.0016452183457147617 82.19281535952956} {436947 833712 0.0014415577306141915 167.1788822889456} {833712 1.393807e+06 0.0013017392682151607 279.7244833388081} {1.393807e+06 2.144502e+06 0.0012110451657496644 402.0680961828125} {2.144502e+06 3.243884e+06 0.001120694858783678 591.6862955068373} {3.243884e+06 4.252725e+06 0.001064653793312162 769.4048384938992} {4.252725e+06 5.489191e+06 0.0010158619299836292 972.7984525022184} {5.489191e+06 7.260306e+06 0.0009696070011254119 1222.5316082781937} {7.260306e+06 8.683148e+06 0.000932753999857703 1485.8753435962026} {8.683148e+06 1.0541136e+07 0.0009026047455994166 1743.6066808302094} {1.0541136e+07 1.2621543e+07 0.0008782587667118885 1996.1946893039676} {1.2621543e+07 1.4617779e+07 0.0008541220294640251 2296.6850507190775} {1.4617779e+07 1.7976931348623157e+308 0.0008369829862892688 2543.212360918902}]
-------> Hit rate: 0.000000
goos: darwin
goarch: amd64
pkg: tbadger
BenchmarkBadger-12    	-------> Hit rate: 0.030000
-------> Hit rate: 0.039600
-------> Hit rate: 0.039574
-------> Hit rate: 0.039863
-------> Hit rate: 0.039875
18224655	      1865 ns/op
PASS
ok  	tbadger	60.895s

WithoutPLR, without pointGet

(master %) 🦀 👉  go test -bench=BenchmarkBadger  -benchtime=30s -run=^s
2021/01/16 18:50:51 plr opening file: data/00000004.sst
2021/01/16 18:50:51 data/00000004.sst
2021/01/16 18:50:51 plrSegment loaded: [{0 43371 0.002763512135689483 1.301027126230357} {43371 173143 0.001957717574102289 32.139598333991884} {173143 436947 0.0016452183457147617 82.19281535952956} {436947 833712 0.0014415577306141915 167.1788822889456} {833712 1.393807e+06 0.0013017392682151607 279.7244833388081} {1.393807e+06 2.144502e+06 0.0012110451657496644 402.0680961828125} {2.144502e+06 3.243884e+06 0.001120694858783678 591.6862955068373} {3.243884e+06 4.252725e+06 0.001064653793312162 769.4048384938992} {4.252725e+06 5.489191e+06 0.0010158619299836292 972.7984525022184} {5.489191e+06 7.260306e+06 0.0009696070011254119 1222.5316082781937} {7.260306e+06 8.683148e+06 0.000932753999857703 1485.8753435962026} {8.683148e+06 1.0541136e+07 0.0009026047455994166 1743.6066808302094} {1.0541136e+07 1.2621543e+07 0.0008782587667118885 1996.1946893039676} {1.2621543e+07 1.4617779e+07 0.0008541220294640251 2296.6850507190775} {1.4617779e+07 1.7976931348623157e+308 0.0008369829862892688 2543.212360918902}]
-------> Hit rate: 0.000000
goos: darwin
goarch: amd64
pkg: tbadger
BenchmarkBadger-12    	-------> Hit rate: 0.030000
-------> Hit rate: 0.037500
-------> Hit rate: 0.039512
-------> Hit rate: 0.039923
14039182	      2240 ns/op
PASS
ok  	tbadger	34.574s

Summary:

With PLR, with pointGet 985 ns/op
W/ot PLR, with pointGet 1025 ns/op (Baseline)

Turn off badger index/pointGet to see the effect of PLR vs bare binary search

With PLR, w/ot pointGet 1865 ns/op
W/ot PLR, w/ot pointGet 2240 ns/op

liufuyang · 2021-01-16T17:41:35Z

 	itr.reset()

-	idx := itr.seekBlock(key)
+	idx := itr.seekBlockWithPlr(key)


One can change this single line to itr.seekBlock(key) to turn off PLR (bourbon) feature

liufuyang · 2021-01-16T18:00:45Z

+			// k := y.KeyWithTs([]byte(fmt.Sprintf("%016x", i)), 0)
+			bs := make([]byte, 4)
+			binary.BigEndian.PutUint32(bs, uint32(i))
+			k := y.KeyWithTs(bs, 0)


This test runs slow and doesn't work on Pingcap/badger's master branch. So we didn't use this to evaluate in the end.

liufuyang · 2021-01-16T18:01:23Z

+		log.Println(filename)
+		log.Printf("plrSegment loaded: %v", data)
+		plr = &plrSegments{inner: data, FName: filename}
+	}


This code above for loading the plr model from the .mod file

liufuyang · 2021-01-16T18:01:37Z

+	}
+	segment := s.inner[segmentIndex]
+	return segment.Slope*key + segment.Intercept, nil
+}


This code for model prediction

liufuyang · 2021-01-16T18:02:19Z

+		// only support key that can turn into uint
+		blockIndex := len(b.baseKeys.endOffs) - 1
+		b.mw.Write([]byte(fmt.Sprintf("%d,%d\n", binary.BigEndian.Uint32(firstKey), blockIndex)))
+	}


Here we dump the block's min key and block's index for training model plr

Update a benchmark test

a540f42

liufuyang commented Jan 16, 2021

View reviewed changes

liufuyang changed the title ~~[Final version] Plt (bourbon) change based on master~~ [Hackathon2020 - Do not merge] Plt (bourbon) change based on master Jan 16, 2021

liufuyang commented Jan 16, 2021

View reviewed changes

Add plr

6601369

liufuyang force-pushed the plt-based-on-master branch from f279d02 to 6601369 Compare January 16, 2021 18:05

liufuyang changed the title ~~[Hackathon2020 - Do not merge] Plt (bourbon) change based on master~~ [Hackathon2020 - Do not merge] Greedy-piecewise-linear-regression to speed up int type key search for LSM-tree (bourbon idea) Jan 16, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Hackathon2020 - Do not merge] Greedy-piecewise-linear-regression (PLR) to speed up int type key search for LSM-tree (bourbon idea) #211

[Hackathon2020 - Do not merge] Greedy-piecewise-linear-regression (PLR) to speed up int type key search for LSM-tree (bourbon idea) #211
liufuyang wants to merge 2 commits into
pingcap:masterfrom
liufuyang:plt-based-on-master

liufuyang commented Jan 16, 2021 •

edited

Loading

Uh oh!

liufuyang Jan 16, 2021 •

edited

Loading

Uh oh!

liufuyang Jan 16, 2021

Uh oh!

liufuyang Jan 16, 2021

Uh oh!

liufuyang Jan 16, 2021

Uh oh!

liufuyang Jan 16, 2021

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

liufuyang commented Jan 16, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Disclaimer

Rust code source:

Benchmark code here:

With PLR, with pointGet

WithoutPLR, with pointGet

WithPLR, without pointGet

WithoutPLR, without pointGet

Summary:

Uh oh!

liufuyang Jan 16, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

liufuyang Jan 16, 2021

Choose a reason for hiding this comment

Uh oh!

liufuyang Jan 16, 2021

Choose a reason for hiding this comment

Uh oh!

liufuyang Jan 16, 2021

Choose a reason for hiding this comment

Uh oh!

liufuyang Jan 16, 2021

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

liufuyang commented Jan 16, 2021 •

edited

Loading

liufuyang Jan 16, 2021 •

edited

Loading