A Map-Reduce System with an Alternate API for Multi-Core Environments

1
A Map-Reduce System with an Alternate API for
Multi-Core Environments

• Presented by Wei Jiang

May 17, 2014
1
May 17, 2014
1
2
Outline

• Background
• MapReduce
• Generalized Reduction
• System Design and Implementation
• Experiments
• Related Work
• Conclusions

May 17, 2014
2
May 17, 2014
2
3
Background

• We have evaluated FREERIDE and Hadoop MapReduce
  based on a set of applications
• Phoenix is one of the implementations of
  MapReduce for shared-memory systems, written in
  C, of small code size
• We also want to make FREERIDE smaller and
  light-weighted

May 17, 2014
3
May 17, 2014
3
4
Googles MapReduce Engine
May 17, 2014
4
May 17, 2014
4
5
Phoenix implementation

• It is based on the same principles but targets
  shared-memory systems
• Consists of a simple API that is visible to
  application programmers
• An efficient runtime that handles
  parallelization, resource management, and fault
  recovery

May 17, 2014
5
6
Phoenix runtime
May 17, 2014
6
7
Generalized Reduction

• Processing structures

May 17, 2014
7
8
A Case Study Apriori
May 17, 2014
8
9
A Case Study Apriori
10
System Design and Implementation

• Basic dataflow of MATE (MapReduce with AlternaTE
  API)
• Data structures to communicate between the user
  code and the runtime
• Three sets of functions in MATE
• Example, how to write a user application

May 17, 2014
10
11
MATE runtime dataflow

• Basic one-stage dataflow

May 17, 2014
11
12
Data structures-(1)

• scheduler_args_t Basic fields

Field Description
Input_data Input data pointer
Data_size Input dataset size
Data_type Input data type
Stage_num Computation-Stage number
Splitter Pointer to Splitter function
Reduction Pointer to Reduction function
Finalize Pointer to Finalize function
May 17, 2014
12
13
Data structures-(2)

• scheduler_args_t Optional fields for performance
  tuning

Field Description
Unit_size of bytes for one element
L1_cache_size of bytes for L1 data cache size
Model Shared-memory parallelization model
Num_reduction_workers Max of threads for reduction workers(threads)
Num_procs Max of processor cores used
May 17, 2014
13
14
Functions-(1)

• Transparent to users

Function Description R/O
static inline void schedule_tasks(thread_wrapper_arg_t ) R
static void combination_worker(void ) R
static int array_splitter(void , int, reduction_args_t ) R
void clone_reduction_object(int num) R
static inline int isCpuAvailable(unsigned long, int) R
May 17, 2014
14
15
Functions-(2)

• APIs provided by the runtime

Function Description R/O
int mate_init(scheudler_args_t args) R
int mate_scheduler(void args) R
int mate_finalize(void args) O
void reduction_object_pre_init() R
int reduction_object_alloc(int size)return the object id R
void reduction_object_post_init() R
void accumulate/maximal/minimal(int id, int offset, void value) O
void reuse_reduction_object() O
void get_intermediate_result(int iter, int id, int offset) O
May 17, 2014
15
16
Functions-(3)

• APIs defined by the user

Function Decription R/O
int (splitter_t)(void , int, reduction_args_t ) O
void (reduction_t)(reduction_args_t ) R
Void (combination_t)(void) O
void (finalize_t)(void ) O
May 17, 2014
16
17
Implementation Considerations

• Data partitioning dynamically assigns splits to
  worker threads
• Buffer management two temporary buffers, one for
  reduction objects, the other for combination
  results
• Fault tolerance re-executes failed tasks
  checkingpoint may be a better solution

18
What is in the user code ?

• Implements necessary functions such as reduction,
  splitter, finalize, and etc.
• Generates the input dataset
• Setups the fields in scheduler_args_t
• Initializes the middleware and declare reduction
  object(s)
• Executes reduction tasks by calling
  mate_scheduler(one or more passes)
• Maybe does some finalizing work

May 17, 2014
18
19
K-means user code

• int main (int argc, char argv)
• parse_args()
• generate_points()
• generate_means()
• mate_init(scheduler_args_t)
• reduction_object_pre_init()
• while(needed) reduction_object_alloc(size)
  
• reduction_object_post_init()
• while(not finished)
• mate_scheduler()
• update_means()
• reuse_reduction_object()
• process_next_iteration()
• mate_finalize()

May 17, 2014
19
20
Experiments K-means

• K-means 400MB, 3-dim points, k 100 on one WCI
  node with 8 cores

May 17, 2014
20
21
Experiments K-means

• K-means 400MB, 3-dim points, k 100 on one AMD
  node with 16 cores

May 17, 2014
21
22
Experiments PCA

• PCA 8000 1024 matrix, on one WCI node with 8
  cores

May 17, 2014
22
23
Experiments PCA

• PCA 8000 1024 matrix, on one AMD node with 16
  cores

May 17, 2014
23
24
Experiments Apriori

• Apriori 1,000,000 transactions, 3 support, on
  one WCI node with 8 cores

May 17, 2014
24
25
Experiments Apriori

• Apriori 1,000,000 transactions, 3 support, on
  one AMD node with 16 cores

May 17, 2014
25
26
Related Work

• Improves MapReduces API or implementations
• Evaluates MapReduce across different platforms
  and application domains
• Acadamia CGL-MapReduce, Mars, MITHRA, Phoenix,
  Disco
• Industry Facebook (Hive), Yahoo! (Pig Latin,
  Map-Reduce-Merge), Google (Sawzall), Microsoft
  (Dryad)

May 17, 2014
26
27
Conclusions

• MapReduce is simple and robust in expressing
  parallelism
• Two-stage computation style may cause
• performance losses for some subclasses of
  applications in data-intensive computing
• MATE provides an alternate API that is based on
  generalized reduction
• This variation can reduce overheads of data
  management and communication between Map and
  Reduce

May 17, 2014
27
28
Questions?
May 17, 2014
28
May 17, 2014
28