1. Speedup over Ji et al.'s work
Motivations
Methods
Reduction-based MapReduce
Memory Hierarchy
MapReduce Programming Model
Emerged with the development of Data-Intensive Computing
GPUs
Cost-effective and Power-efficient
Suitable for implementing MapReduce
Effective Utilization of the Fast Shared Memory
Shared memory size is small
MapReduce generates a lot of key-value pairs
Reduction-intensive applications:
Reduction : associative and commutative
Lots of key-value pairs share the same key
map(input) {
(key, value) = process(input);
reductionobject->insert(key, value); }
reduce(value1, value2)
value1 = operation(value1, value2);
Stores reduction objects in both shared memory and device memory
Combines reduction results
Handles shared memory overflow
GPU
Reduction Object 0
Reduction Object 1
Reduction Object 0
Reduction Object 1
… …
… …
… …
Block 0’s Shared Memory
Block 0’s Shared Memory
Device Memory Reduction Object
Reduction Object
Result Array
Device Memory
… …
CPU
MapReduce
Host Memory
… …
KeyIdx[0]
ValIdx[0]
KeyIdx[1]
ValIdx[1]
Multi-group Scheme
Memory Allocator
Divides each thread block into sub-groups
Each sub-group owns a shared memory reduction object
Trades off memory overhead and locking overhead M M M M M M M M
Val Data
Key Size
Val Size
Key Data
Key Size
Val Size
Key Data
Val Data
K1:v k1:v k2:v
K1:v
K3:v k4:v
K4:v k5:v
K4:v
K1:v k3:v
Group by Key
Experimental Results
Conclusions
K1: v, v, v, v
K2:v
K3:v, v
K4:v, v, v
K5:v R R R R R
Presents methods of utilizing shared memory
for MapReduce on GPUs
Uses a reduction-based method to reduce memory overhead of the MapReduce, and performs reductions in shared memory
Designs an approach for storing reduction objects on the memory hierarchy of a GPU
Balances the memory overhead and locking overhead by developing a multi-group scheme.
Configuration
Multi-group Scheme
Speedup over Sequential
GPU: NVIDIA Quadro FX 5800
Device Memory: 4 GB
Shared Memory: 16 KB
GPUs
Processing component:
Memory component:
Host
Kernel 1
Kernel 2
Device
Grid 1
Block
(0, 0)
(1, 0)
(2, 0)
(0, 1)
(1, 1)
(2, 1)
Grid 2
Block (1, 1)
Thread
(3, 1)
(4, 1)
(0, 2)
(1, 2)
(2, 2)
(3, 2)
(4, 2)
(3, 0)
(4, 0)
(Device) Grid
Constant
Memory
Texture
Device
Block (0, 0)
Shared Memory
Local
Thread (0, 0)
Registers
Thread (1, 0)
Block (1, 0)
Host
Speedup over MapCG
(Reduction-intensive)
Speedup over MapCG
(Non-reduction-intensive)
Speedup over Ji et al.'s work