SLIDE 1
Efficient Join Processing across Heterogeneous Processors
Henning Funke, Sebastian Breß, Stefan Noll, Jens Teubner December 15, 2015
1 / 14
Efficient Join Processing across Heterogeneous Processors Henning - - PowerPoint PPT Presentation
Efficient Join Processing across Heterogeneous Processors Henning - - PowerPoint PPT Presentation
Efficient Join Processing across Heterogeneous Processors Henning Funke, Sebastian Bre, Stefan Noll, Jens Teubner December 15, 2015 1 / 14 GPUs IN DATABASES ARE LIKE A MUSCLE CAR IN A TRAFFIC JAM 3 / 14 Bottleneck 3 / 14 really? 3 / 14
SLIDE 2
SLIDE 3
3 / 14
SLIDE 4
3 / 14
Bottleneck
SLIDE 5
3 / 14
really?
SLIDE 6
GPU – Hashjoin Algorithm
Cuckoo Hashing Implementation1 → Strict limit on number probes per query key
◮ Pipeline join probe and result compaction in shared memory
1Based on: Alcantara, Dan Anthony Feliciano. Efficient hash tables on the
- GPU. University of California at Davis, 2011.
4 / 14
SLIDE 7
GPU – Hashjoin Algorithm
Cuckoo Hashing Implementation1 → Strict limit on number probes per query key
◮ Pipeline join probe and result compaction in shared memory
Performance: Build (GTX970) 0.5 0.6 0.7 0.8 0.9 5 10 PCIe Fill factor Build Throughput GB/s
1Based on: Alcantara, Dan Anthony Feliciano. Efficient hash tables on the
- GPU. University of California at Davis, 2011.
4 / 14
SLIDE 8
Performance: GPU Join Probe (GTX970)
100 101 102 103 104 105 5 10 15 20 PCIe Build Size KB Probe Throughput GB/s
5 / 14
SLIDE 9
Performance: GPU Join Probe (GTX970)
100 101 102 103 104 105 5 10 15 20 PCIe > 100K elements Real world data Build Size KB Probe Throughput GB/s
6 / 14
SLIDE 10
Joins on Multiple Heterogeneous Processors
Challenges
◮ Scalability to large data ◮ Communication ◮ Local and remote resources
Related work
Figure : P. Frey, R. Goncalves, M. L. Kersten, and J. Teubner. Spinning
Relations: High-Speed Networks for Distributed Join Processing. In DaMoN, 2009.
7 / 14
SLIDE 11
Star Join on Heterogeneous Processors
Processing Strategy
→ Allocate smaller tables on all devices → Asynchronous data transfers at computation speed → Merge results into continuous arrays
8 / 14
SLIDE 12
Performance: Join Probe Across Heterogeneous Processors
Intel Xeon E5-1607 v2 and NVIDIA Geforce GTX970 2 4 6 8 1 2 3 4 5 CPU worker threads Probe throughput GB/s CPU alone CPU + GPU
9 / 14
SLIDE 13
Coprocessor Control Thread Scheduling
Figure : Profiling coprocessor kernel invocations
→ Steer control flow from coprocessor itself → Increase block size
10 / 14
SLIDE 14
Hardware Schema – Memory Bandwidth Utilization
MC CPU GPU
149 GB/s
MEM MEM
Per core (4) scan 15.8 GB/s gather 4 GB/s even share 7.8 GB/s Local prefix scan 84 GB/s gather 33.2 GB/s
PCIe
12 GB/s 31 GB/s
11 / 14
SLIDE 15
Hardware Schema – Memory Bandwidth Utilization
MC CPU GPU
149 GB/s
MEM MEM
Per core (4) scan 15.8 GB/s gather 4 GB/s even share 7.8 GB/s Local prefix scan 84 GB/s gather 33.2 GB/s
PCIe
12 GB/s 31 GB/s
12 / 14
SLIDE 16
Hardware Schema – Memory Bandwidth Utilization
MC CPU GPU
149 GB/s
MEM MEM
Per core (4) scan 15.8 GB/s gather 4 GB/s even share 7.8 GB/s Local prefix scan 84 GB/s gather 33.2 GB/s
PCIe
12 GB/s 31 GB/s
◮ PCI express bus and main memory can become a bottleneck ◮ Take bandwidth footprint and throughput into account
→ Instead of bulk processing, apply dataflow perspective
12 / 14
SLIDE 17
Improving Resource Utilization
Cache awareness
◮ Materialize tuples in hash table
→ Useful payload in cacheline
◮ Order probe data by hash function
GPU Optimizations
◮ Pipeline data between GPU kernels ◮ Concurrent kernel execution
13 / 14
SLIDE 18
Key Insights
◮ PCIe is not the dominating
bottleneck for GPU joins
◮ Dataflow oriented processing
→ Join arbitrary outer relation sizes
◮ Move part of probes to coprocessor
→ Save memory bandwidth → Gain throughput
Future Work
◮ Join processing → query processing ◮ Compile pipelined operator sequences ◮ Stream arbitrary columns
Figure : Operator
- pipelines. From Leis et al.
Morsel-driven parallelism SIGMOD 2014
14 / 14
SLIDE 19
Key Insights
◮ PCIe is not the dominating
bottleneck for GPU joins
◮ Dataflow oriented processing
→ Join arbitrary outer relation sizes
◮ Move part of probes to coprocessor
→ Save memory bandwidth → Gain throughput
Future Work
◮ Join processing → query processing ◮ Compile pipelined operator sequences ◮ Stream arbitrary columns
Thank you!
Figure : Operator
- pipelines. From Leis et al.
Morsel-driven parallelism SIGMOD 2014
14 / 14