Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ②...
Transcript of Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ②...
![Page 1: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/1.jpg)
JIAXIN SHI, YOUYANG YAO RONG CHEN, HAIBO CHEN
Institute of Parallel and Distributed Systems (IPADS)Shanghai Jiao Tong University
Fast and Concurrent RDF Queries with RDMA-basedDistributed Graph Exploration
FEIFEI LI
School of ComputingUniversity of Utah
http://ipads.se.sjtu.edu.cn/projects/wukong
![Page 2: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/2.jpg)
2
Graphs are Everywhere
Online graph query plays a vital role for searching, mining and reasoning linked data
UnicornTAO
![Page 3: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/3.jpg)
3
Graph Analytics vs. Graph Query
Graph Analytics Graph QueryGraph Model Property Graph Semantic (RDF) Graph
Working Set A whole Graph A small frac. of Graph
Processing Batched & Iterative Concurrent
Metrics Latency Latency & Throughput
![Page 4: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/4.jpg)
4
RDF and SPARQLResource Description Framework (RDF)► Representing linked data on the Web► Public knowledge bases: DBpedia, PubChemRDF, Bio2RDF► Google’s knowledge graph
![Page 5: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/5.jpg)
5
RDF and SPARQLRDF is a graph composed by a set of ⟨Subject, Predicate, Object⟩ triples
Haibo mo IPADSHaibo to OSJiaxin ad RONGJiaxin tc OSRong mo IPADSRong to DSXingda ad Haibo. . .
mo: MemberOfad: ADvisorto: TeacherOftc: TakeCourse
to
Xingda
OS
ad
Rongto
Jiaxin
DS
adYouyang
tc
mo
mo
tc
tc
Haibo
IPADS
triple
Haibo
IPADS
![Page 6: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/6.jpg)
6
RDF and SPARQLSPARQL is standard query language for RDF
SELECT ?Y WHERE {?X mo IPADS .?X to ?Y .}
constant
variable triple pattern
to
Xingda
OS
ad
Rongto
Jiaxin
DS
adYouyang
tc
mo
mo
tc
tc
Haibo
IPADS
Courses (?Y) taught by Teachers (?X) from IPADS
![Page 7: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/7.jpg)
7
RDF and SPARQLSPARQL is standard query language for RDF
SELECT ?Y WHERE {?X mo IPADS .?X to ?Y .}
constant
variable triple pattern
Courses (?Y) taught by Teachers (?X) from IPADS
to
Xingda
OS
ad
Rongto
Jiaxin
DS
adYouyang
tc
mo
mo
tc
tc
Haibo
IPADS
toOS
Rongto
DS
mo
mo
Haibo
IPADS
?Y?X
![Page 8: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/8.jpg)
Triple Store
Nod
e1N
ode2
8
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
Existing SolutionsTriple Store and Triple Join► Store RDF data as a set of triples in RDBMS
Jiaxin ad RongRong mo IPADS.. .. ..
Triple Join?X ?Y ?Z
Haibo mo IPADSXingda ad Haibo.. .. ..
③
①②
③① ②
N1
N2
![Page 9: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/9.jpg)
Triple Store
Nod
e1N
ode2
9
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
Existing SolutionsTriple Store and Triple Join► Store RDF data as a set of triples in RDBMS
Jiaxin ad RongRong mo IPADS.. .. ..
Triple Join?X ?Y ?Z
Haibo mo IPADSXingda ad Haibo.. .. ..
③
①②
③① ②
N1
N2
► Costly distributed join► Large intermediate results
![Page 10: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/10.jpg)
10
Existing Solutions
Final Join
Graph Store and Graph Exploration► Store RDF data in a native graph model
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
Graph StoreNode1
Node2
③
①②
N1
N2
?X ?Y
①
?Z ?X
③
?Y ?Z
②
?X ?Y ?Z
One-step pruning
![Page 11: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/11.jpg)
11
Existing Solutions
Final Join
Graph Store and Graph Exploration► Store RDF data in a native graph model
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
Graph StoreNode1
Node2
③
①②
N1
N2
?X ?Y
①
?Z ?X
③
?Y ?Z
②
?X ?Y ?Z
One-step pruning
► Costly final join (90%)► Synchronized execution
![Page 12: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/12.jpg)
: A distributed in-memory RDF store
12
System OverviewWukong
RDMA
SPARQL queries
![Page 13: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/13.jpg)
: A distributed in-memory RDF store
13
System OverviewWukong
RDMA
SPARQL queries
![Page 14: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/14.jpg)
: A distributed in-memory RDF store
14
System OverviewWukong► RDMA-friendly graph model ► RDMA-based join-free graph exploration► Concurrent query processing► Results vs. state-of-the-art (TriAD/Trinity.RDF)
► Latency: 11.9X – 28.1X reduction► Throughput: 269K queries/sec (up to 740X improvement)
![Page 15: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/15.jpg)
Agenda
Graph-based Model & Store
Query Processing Engine
Evaluation
![Page 16: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/16.jpg)
16
Graph Model and Indexes
SELECT ?X ?Y WHERE {?X tc ?Y .?X type Student .
}Jiaxin
tc
DS
Xingda
tc
OS
tc
Student
Course
tcSELECT ?X WHERE {Jiaxin tc ?X .
}
Hard to query w/o indexing
Predicate index
Type index
Easy to start from a constant
![Page 17: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/17.jpg)
17
Differentiated Graph Partitioning
► Start from normal vertex► Exploit locality
Jiaxin
DS
Xingda
OS
tc
Student
Course
SELECT ?X WHERE {Jiaxin tc ?X .
}
SELECT ?X ?Y WHERE {?X tc ?Y .?X type Student .
}
► Start from index vertex► Exploit parallelism
![Page 18: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/18.jpg)
18
Differentiated Graph Partitioning
Jiaxin
tc
DS
Xingda
tc
OS
tc tc
StudentCourse
Student
Course
OS
tc
Jiaxin
DS
Xingda
OS
tc
Student
Course
Normal vertex : DistributedIndex vertex : Partitioned
► Start from normal vertex► Exploit locality
SELECT ?X WHERE {Jiaxin tc ?X .
}
SELECT ?X ?Y WHERE {?X tc ?Y .?X type Student .
}
► Start from index vertex► Exploit parallelism Inspired by PoweLyra [Eurosys’15]
![Page 19: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/19.jpg)
Predicate-based KV Store
19
Rongto
Jiaxin
DS
adIPADS
mo
SELECT ?X WHERE {Rong to ?X .
}
to…
□ Inefficient lookup□ Unnecessary data transfer
Rong in ad Jiaxin
Rong out mo IPADS to DS to . . .
Vertex (S/O)PredicateDirection
constant
![Page 20: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/20.jpg)
Predicate-based KV Store
20
Rongto
Jiaxin
DS
adIPADS
mo to…
□ Inefficient lookup□ Unnecessary data transfer
ad in JiaxinRong
mo out IPADSRong
to out DSRong . . .
Vertex (S/O)PredicateDirection
Rong in ad Jiaxin
Rong out mo IPADS to DS to . . .
Move predicate to key-side
Finer-grained vertex decomposition
SELECT ?X WHERE {Rong to ?X .
}constant
![Page 21: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/21.jpg)
Agenda
Graph-based Model & Store
Query Processing Engine
Evaluation
![Page 22: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/22.jpg)
22
Query Processing
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X .
}?Z
?X ?Y
tcad
to
The teacher advises the studentwho also takes a coursetaught by the teacher
mo: MemberOfad: ADvisorto: TeacherOftc: TakeCourse
to
Xingda
OS
ad
Rongto
Jiaxin
DS
adYouyang
tc
mo
mo
tc
tc
Haibo
IPADS
![Page 23: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/23.jpg)
23
Observation
Final Join
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }③
①②
N1
N2
?X ?Y
①
?Z ?X
③
?Y ?Z
②
► Costly final join (90%)► Synchronized execution
One-step pruning
![Page 24: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/24.jpg)
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
24
Observation
Graph exploration w/ full-history pruning
Final Join
N1
N2
?X ?Y
①
?Z ?X
③
?Y ?Z
②
One-step pruning
► Costly final join (90%)► Synchronized execution
the latency of RDMA is relatively insensitive to payload sizes (~2K)
e.g. 8B/1.56µs vs. 2KB/2.25µs
![Page 25: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/25.jpg)
25
Full-History Pruningto to
Query Query
Parallel execution on predicate index
Haiboto
Xingda
OSad
Rongto
Jiaxin
DS
adYouyangtc
IPADS mo
mo
tc
tc
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
![Page 26: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/26.jpg)
26
Full-History Pruningto
Rongto
DS
Youyang
tc
ad
H:Rong
H:Rong DS
H:Rong DS Youyang
toFull-history
Haiboto
Xingda
OSad
Rongto
Jiaxin
DS
adYouyangtc
IPADS mo
mo
tc
tc
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
![Page 27: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/27.jpg)
Jiaxin
27
Full-History Pruningto
Rongto
DS
Youyang
tc
ad
H:Rong
H:Rong DS
H:Rong DS Youyang
to
to
OS
Haibo
tc
ad
H:Haibo
H:Haibo OS
H:Haibo OS JiaxinHaibo OS Xingda
Haibo OS Xingda Haibo
Haibo
Xingda
Haiboto
Xingda
OSad
Rongto
Jiaxin
DS
adYouyangtc
IPADS mo
mo
tc
tc
Jiaxin
Rong
ad
Haibo OS Jiaxin Rong
H:Haibo OS Jiaxin
Full-history
SELECT ?X ?Y ?Z WHERE {?X to ?Y .?Z tc ?Y .?Z ad ?X . }
![Page 28: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/28.jpg)
28
Migrate Execution or Data
► Send sub-query by RDMA WRITE► Async exploration w/ full-History
Exploit parallelism
Fork-join(migrate exec)
![Page 29: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/29.jpg)
29
Migrate Execution or Data
► Fetch data by RDMA READ► Bypass remote CPU & OS
Exploit low latency
In-place (migrate data)
► Send sub-query by RDMA WRITE► Async exploration w/ full-History
Exploit parallelism
Fork-join(migrate exec)
Dynamic switch at runtime
![Page 30: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/30.jpg)
30
Other Designs of Wukong
► Logical task queues
► Multi-threading large-query
► Latency-centric work stealing
► Support Evolving graph
![Page 31: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/31.jpg)
Agenda
Graph-based Model & Store
Query Processing Engine
Evaluation
![Page 32: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/32.jpg)
EvaluationBaseline: state-of-the-art systems□ Centralized: RDF-3X, BitMat□ Distributed: TriAD, Trinity.RDF, SHARD
Platforms: a rack-scale 6-machine cluster□ Each: two 10-cores Intel Xeon, 64GB DRAM,
Mellanox 56Gbps InfiniBand NIC w/ RDMA1
Benchmarks□ Synthetic: LUBM, WSDTS□ Real-life: DBPSB, YAGO2
1 All machines run Ubuntu 14.04 with Mellanox OFED v3.0-2.0.1 stack. 32
10xCore 10xCore
56GBps IB NIC
40Gbps IB Switch
RDMA
![Page 33: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/33.jpg)
Single Query Latency (msec)
33
Outperform state-of-the-art systems(Geometric Mean)
► vs. Trinity.RDF: 28.1X► vs. TriAD: 11.9X
Group I (L1-3,7): large queries Group II (L4-6): small queries► Start from index vertex► Touch a large subset of graph► Speedup: 4.1X - 21.7X
► Start from normal vertex► Touch a small subset of graph► Speedup: 8.4X – 70.6X
![Page 34: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/34.jpg)
Factor Analysis of Improvement (msec)
34
BASE► Graph-exploration► One-step pruning► Comm. w/ TCP/IP
+RDMA► Comm. w/ RDMA
+FHP► Full-history pruning
+IDX► Index vertex► Diff. partitioning
+PBS► Predicate-base fine-
grained Store
+DYN► In-place execution► Dynamic switching
![Page 35: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/35.jpg)
Throughput of Mixed Workloads
351 The templates of 6 classes of queries are based on group (II) queries (L4, L5 and L6) and three additional queries from official website (A1, A2 and A3).
278~740X
50th: 0.80 ms99th: 5.90 ms269K
queries/sec
Mixed workload: 6 classes of small queries1
![Page 36: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/36.jpg)
Conclusion
: a distributed in-memory RDF store that leverages RDMA-based graph exploration to support fast and concurrent RDF queries
Achieving orders-of-magnitude lower latency & higher throughput than prior state-of-the-art systems
36
Wukong
New hardware technologies open opportunities
http://ipads.se.sjtu.edu.cn/projects/wukong
![Page 37: Fast and Concurrent RDF Queries with RDMA-based · Final Join N1 N2?X ?Y ①?Z ?X ③?Y ?Z ② One-step pruning Costly final join (90%) Synchronized execution the latency of RDMA](https://reader035.fdocuments.in/reader035/viewer/2022071020/5fd42bbd142f0945aa51d208/html5/thumbnails/37.jpg)
Questions
Thanks
http://ipads.se.sjtu.edu.cn/projects/wukong
Wukong, short for Sun Wukong, who is known asthe Monkey King and is a main character in theChinese classical novel “Journey to the West”.Since Wukong is known for his extremely fastspeed (21,675 kilometers in one somersault) andthe ability to fork himself to do massive multi-tasking, we term our system as Wukong.