ANL 2014 - Chicago Elastic and Efficient Execution of Data- Intensive Applications on Hybrid Cloud Tekin Bicer Computer Science and Engineering Ohio State.

ANL 2014 - Chicago

Elastic and Efficient Execution of Data-Intensive Applications on Hybrid Cloud

Tekin Bicer

Computer Science and Engineering

Ohio State University

ANL 2014 - Chicago

Introduction

• Scientific simulations and instruments– X-ray Photon Correlation Spectroscopy

• CCD Detector: 120MB/s now; 44GB/s by 2015

– Global Cloud Resolving Model• 1PB for 4km grid-cell

• Performed on local clusters– Not sufficient

• Problems– Data Analysis, Storage, I/O performance

• Cloud Technologies– Elasticity– Pay-as-you-go Model

2

ANL 2014 - Chicago

Hybrid Cloud Motivation

• Cloud technologies– Typically associated with

computational resources• Massive data generation

– Exhaust local storage• Hybrid Cloud

– Local Resources: Base– Cloud Resources: Additional

• Cloud– Compute and storage

resources

3

Local Resources

ANL 2014 - Chicago

Cloud Storage

Usage of Hybrid Cloud

4

Local Storage

Data Source

Local Nodes

Cloud Compute Nodes

ANL 2014 - Chicago

Challenges

• Data-Intensive Processing– Transparent Data Access and

Analysis

– Programmability of Large-Scale Applications

• Meeting User Constraints– Enabling Cloud Bursting

• Minimizing Storage and I/O Cost– Domain Specific Compression– In-Situ and In-Transit Data

Analysis

5

MATE-HC: Map-reduce with AlternaTE APIover Hybrid Cloud

Dynamic Resource Allocation Framework

for Hybrid Cloud

Compression Methodology and

System for Large-Scale App.

ANL 2014 - Chicago

Programmability of Large-Scale Applications on Hybrid Cloud

• Geographically distributed resources• Ease of programmability

– Reduction-based programming structures• MATE-HC

– A middleware for transparent data access and processing

– Selective job assignment– Multi-threaded data retrieval

6

ANL 2014 - Chicago

Middleware for Hybrid Cloud

...

...Data

Slaves

MasterLocal Cluster

LocalReduction

Job Assignment

...

...Data

Slaves

Master

Cloud Environment

Job Assignment

LocalReduction

Index

7

Remote DataAnalysis

Job Assignment

Job Assignment

GlobalReduction

GlobalReduction

ANL 2014 - Chicago

MATE vs. Map-Reduce Processing Structure

8

• Reduction Object represents the intermediate state of the execution

• Reduce func. is commutative and associative• Sorting, grouping.. overheads are eliminated with red. func/obj.

ANL 2014 - Chicago

Simple Example

3 5 8 4 1 3 5 2 6 7 9 4 2 4 8

9

Our large Dataset

Our Compute NodesRobj[1]= Robj[1]= Robj[1]=

Local Reduction (+) Local Reduction(+)Local Reduction(+)

8 15 1421 23 27

Result= 71 Global Reduction(+)

ANL 2014 - Chicago

Experiments

• 2 geographically distributed clusters– Cloud: EC2 instances running on Virginia

– 22 nodes x 8 cores– Local: Campus cluster (Columbus, OH)

– 150 nodes x 8 cores

• 3 applications with 120GB of data– KMeans: k=1000; KNN: k=1000; – PageRank: 50x10 links w/ 9.2x10 edges

• Goals:

– Evaluating the system overhead with different job distributions– Evaluating the scalability of the system

10

ANL 2014 - Chicago

System Overhead: K-Means

11

Env-* Global Reduction

Idle Time Total Slowdown

Stolen # Jobs (960)local EC2

50/50 0.067 0 93.871 20.430 (0.5%) 0

33/67 0.066 0 31.232 142.403 (5.9%) 128

17/83 0.066 0 25.101 243.31 (10.4%) 240

ANL 2014 - Chicago

Scalability: K-Means

12

ANL 2014 - Chicago

Summary

• MATE-HC is a data-intensive middleware developed for Hybrid Cloud

• Our results show that – Low inter-cluster comm. overhead– Job distribution is important– Overall slowdown is modest – Proposed system is scalable

13

ANL 2014 - Chicago

Outline

• Data-Intensive Processing– Programmability of Large-Scale

Applications– Transparent Data Access and

Analysis

• Meeting User Constraints– Enabling Cloud Bursting

• Minimizing Storage and I/O Cost– Domain Specific Compression– In-Situ and In-Transit Data

Analysis

14

MATE-HC: Map-reduce with AlternaTE APIover Hybrid Cloud

Dynamic Resource Allocation Framework

for Cloud Bursting

Compression Methodology and

System for Large-Scale App.

ANL 2014 - Chicago

Dynamic Resource Allocation for Cloud Bursting

• Performance of cloud resources and workload vary– Problems:

• Extended execution times• Unable to meet user constraints

– Cloud resources can dynamically scale• Cloud Bursting

– In-house resources: Base workload– Cloud resources: Adopt performance requirements

• Dynamic Resource Allocation Framework– A model for capturing “Time” and “Cost” constraints

with cloud bursting

15

ANL 2014 - Chicago

System Components

• Local cluster and Cloud • MATE-HC processing structure• Pull-based job distribution• Head Node

– Coarse grained job assignment– Consideration of locality

• Master node– Fine grained job assignment

• Job Stealing– Remote data processing

16

ANL 2014 - Chicago

Resource Allocation FrameworkEstimate required time for local cluster processing

Estimate required time for cloud cluster processing

All variables can be profiled during execution, except estimated # stolen jobs

Estimate remaining # jobs after local jobs are consumed

Ratio of local computational throughput in system

17

ANL 2014 - Chicago

Execution of Resource Allocation Framework

• Head Node– Evaluates profiled info.– Estimates # cloud inst.

• Before each job assign.

– Informs Master nodes• Master Node

– Each cluster has one– Collects profile info.

• During job req. time– (De)allocates resources

• Slave Nodes– Request and consume jobs

18

ANL 2014 - Chicago

Experimental Setup

• Two Applications– KMeans (520GB): Local=104GB; Cloud=416GB– PageRank (520GB): Local=104GB; Cloud=416GB

• Local cluster: Max. 16 nodes x 8 cores = 128 cores

• Cloud resources: Max. 16 nodes x 8 cores = 128 cores

• Evaluation of model– Local nodes are dropped during execution– Observed how system is adopted

19

ANL 2014 - Chicago

KMeans – Time Constraint

# Local Inst.: 16 (fixed)# Cloud Inst.: Max 16 (Varies)Local: 104GB, Cloud:416GB

• System is not able to meet the time constraint because max. # of cloud instances is reached• All other configurations meet the time constraint with <1.5% error rate

20

ANL 2014 - Chicago

KMeans – Cloud Bursting

• 4 local nodes are dropped …• After 25% and 50% of time constraints are elapsed, error rate <1.9%• After 75% of time constraint is elapsed, error rate <3.6%

• Reason of higher error rate: Shorter time to profile new environment

# Local Inst.: 16 (fixed)# Cloud Inst.: Max 16 (Varies)Local: 104GB, Cloud:416GB

21

ANL 2014 - Chicago

Summary

• MATE-HC: MapReduce type of processing– Federated resources

• Developed a resource allocation model– Based on feedback mechanism– Time and cost constraints

• Two data-intensive applications (KMeans, PR)– Error rate for time < 3.6%– Error rate for cost < 1.2%

22

ANL 2014 - Chicago

Outline

• Data-Intensive Processing– Programmability of Large-Scale

Applications– Transparent Data Access and

Analysis

• Meeting User Constraints– Enabling Cloud Bursting

• Minimizing Storage and I/O Cost– Domain Specific Compression– In-Situ and In-Transit Data

Analysis

23

MATE-HC: Map-reduce with AlternaTE API

over HC

Dynamic Resource Allocation Framework

for Cloud Bursting

Compression Methodology and

System for Large-Scale App.

ANL 2014 - Chicago

Data Management using Compression

• Generic compression algorithms– Good for low entropy sequence of bytes– Scientific dataset are hard to compress

• Floating point numbers: Exponent and mantissa• Mantissa can be highly entropic

• Using compression is challenging– Suitable compression algorithms– Utilization of available resources– Integration of compression algorithms

24

ANL 2014 - Chicago

Compression Methodology• Common properties of scientific datasets

– Multidimensional arrays– Consist of floating point numbers– Relationship between neighboring values

• Domain specific solutions can help• Approach:

– Prediction-based differential compression• Predict the values of neighboring cells• Store the difference

25

ANL 2014 - Chicago

Example: GCRM Temperature Variable Compression

• E.g.: Temperature record• The values of neighboring cells

are highly related• X’ table (after prediction):

• X’’ compressed values– 5bits for prediction +

difference• Lossless and lossy comp.• Fast and good compression

ratios

26

ANL 2014 - Chicago

Compression Framework

• Improve end-to-end application performance• Minimize the application I/O time

– Pipelining I/O and (de)compression operations• Hide computational overhead

– Overlapping application computation with compression framework

• Easy implementation of compression algorithms• Easy integration with applications

– Similar API to POSIX I/O

27

ANL 2014 - Chicago

A Compression Framework for Data Intensive Applications

Chunk Resource Allocation (CRA) Layer• Initialization of the system• Generate chunk requests, enqueue processing • Converting original offset and data size requests to

compressed

28

Parallel Compression Engine (PCE)• Applies encode(), decode() functions to chunks• Manages in-memory cache with informed prefetching• Creates I/O requests

Parallel I/O Layer (PIOL)• Creates parallel chunk requests to storage medium• Each chunk request is handled by a group of threads• Provides abstraction for different data transfer

protocols

28

ANL 2014 - Chicago

Integration with a Data-Intensive Computing System

• Remote data processing– Sensitive to I/O bandwidth

• Processes data in…– local cluster– cloud– or both (Hybrid Cloud)

29

ANL 2014 - Chicago

Experimental Setup

• Two datasets:– GCRM: 375GB (L:270 + R:105)– NPB: 237GB (L:166 + R:71)

• 16x8 cores (Intel Xeon 2.53GHz)• Storage of datasets

– Lustre FS (14 storage nodes)– Amazon S3 (Northern Virginia)

• Compression algorithms– CC, FPC, LZO, bzip, gzip, lzma

• Applications: AT, MMAT, KMeans

30

ANL 2014 - Chicago

Performance of MMAT

31

Local Remote Hybrid0

200

400

600

800

1000

1200

1400

1600

1800OriginalLZOCC

Exec

ution

Tim

e (s

ecs)

Original CC Original CCLocal Remote

0

200

400

600

800

1000 ReadDecomp.Reduction

Exe

cuti

on

Tim

e (s

ec)

Speedups

Local Remote Hybrid

CC 1.63 1.90 1.85

LZO 1.04 1.24 1.14

Compression Ratios

CC 51.68% (186GB)

LZO 20.40% (299GB)

Breakdown of Performance• Overhead (Local): 15.41%• Read Speedup: 1.96

I/O Throughput (128np)

GB/sec Orig. CC

Local 1.62 3.21

Remote 0.1 0.19

ANL 2014 - Chicago

Lossy Compression (MMAT)

32

Local Remote Hybrid0

100

200

300

400

500

600

700

800

900

1000CCLossy 2eLossy 4e

Exec

ution

Tim

e (s

ec)

Speedups

Local Remote Hybrid

2e vs CC 1.07 1.18 1.09

4e vs CC 1.13 1.43 1.18

4e vs orig. 1.76 2.41 2.18

Compression Ratios

Lossless 51.68%

2e 56.88% (162GB)

4e 62.93% (139GB)

Lossy• #e: # dropped bits• Error bound: 5x(1/10^5)

ANL 2014 - Chicago

Summary• Management and analysis of scientific datasets are

challenging– Generic compression algorithms are inefficient for

scientific datasets• Proposed a compression framework and methodology

– Domain specific compression algorithms are fast and space efficient• 51.68% compression ratio• 53.27% improvement in exec. time

– Easy plug-and-play of compression– Integration of the proposed framework and methodology

with a data analysis middleware

33

ANL 2014 - Chicago

Outline

• Data-Intensive Processing– Programmability of Large-Scale

Applications– Transparent Data Access and

Analysis

• Meeting User Constraints– Enabling Cloud Bursting

• Minimizing Storage and I/O Cost– Domain Specific Compression– In-Situ and In-Transit Data

Analysis

34

MATE-HC: Map-reduce with AlternaTE APIover Hybrid Cloud

Dynamic Resource Allocation Framework

for Cloud Bursting

Compression Methodology and

System for Large-Scale App.

ANL 2014 - Chicago

In-Situ and In-Transit Analysis

• Compression can ease data management– But may not always be sufficient

• In-situ data analysis– Co-locate data source and analysis code– Data analysis during data generation

• In-transit data analysis– Remote resources are utilized– Forward generated data to “staging nodes”

35

ANL 2014 - Chicago

In-Situ and In-Transit Data Analysis

• Significant reduction in generated dataset size– Noise elimination, data filtering, stream mining…– Timely insights

• Parallel data analysis– MATE-Stream

• Dynamic resource allocation and load balancing– Hybrid data analysis– Both in-situ and in-transit

36

ANL 2014 - Chicago

Robj[...]

Robj[...]

Robj[...]

Robj[...]

LR

LR

LR

LR

Parallel In-Situ Data Analysis

37

DataSource

Disp LRobj[...]

Local Combination• Intermediate results• Timely insights• Continuous global red.

Local Reduction• Filtering, stream mining• Data reduction• Continuous local red.

Data Generation• Scientific instruments,

simulations, etc.• (Un)bounded data

ANL 2014 - Chicago

Robj[...]

Robj[...]

Robj[...]

Robj[...]

LR

LR

LR

LR

Robj[...]

Robj[...]

Robj[...]

Robj[...]

LR

LR

LR

LR

Elastic In-Situ Data Analysis

38

DataSource

Disp

LRobj[...]

Insufficient resource utilization• Dynamically extend resources• New local reduction proc.

ANL 2014 - Chicago

Robj[...]

Robj[...]

Robj[...]

Robj[...]

LR

LR

LR

LR

Robj[...]

Robj[...]

Robj[...]

Robj[...]

LR

LR

LR

LR

Elastic In-Situ and In-Transit Data Analysis

39

DataSource

Disp LRobj[...]

Disp LRobj[...]

GRobj[...]

Staging node is set• Forward dataReduction process:1. Local comb.2. Global comb.

N0

N1

ANL 2014 - Chicago

Future Directions

• Scientific applications are difficult to modify– Integration with existing data sources– GridFTP, (P)NetCDF and HDF5 etc.

• Data transfer is expensive (especially for in-transit)– Utilization of advanced network technologies– Software-Defined Networking (SDN)

• Long running nature of large-scale app.– Failures are inevitable– Exploit features of processing structure

40

ANL 2014 - Chicago

Conclusions

• Data-intensive applications and instruments can easily exhaust local resources

• Hybrid cloud can provide additional resources • Challenges: Transparent data access and processing; meeting user

constraints; minimizing I/O and storage cost• MATE-HC: Transparent and efficient data processing on Hybrid

Cloud• Developed a “dynamic resource allocation framework” and

integrated with MATE-HC– Time and cost sensitive data processing

• Proposed a “compression methodology and a system” to minimize storage cost and I/O bottleneck

• Design of “in-situ and in-transit data analysis” (on going work)

41

ANL 2014 - Chicago 42

Thanks for your attention!

ANL 2014 - Chicago

MATE-EC2 Design

• Data organization– Three levels: Buckets, Chunks and Units– Metadata information

• Chunk Retrieval– Threaded Data Retrieval– Selective Job Assignment

• Load Balancing and handling heterogeneity– Pooling mechanism

43

ANL 2014 - Chicago

MATE-EC2 vs. EMR

44

• KMeans• Speedups

• vs. combine3.54 – 4.58

• PageRank• Speedups

• vs. combine4.08 – 7.54

ANL 2014 - Chicago

Different Chunk Sizes

• KMeans• 1 retrieval threads• Performance

increase– 128KB vs. >8M– 2.07 to 2.49

45

ANL 2014 - Chicago

K-Means (Data Retrieval)

• Fig 1: 16 Retrieval Threads– 8M vs. others speedup: 1.13-1.30

• Fig. 2: 128M Chunk Size– 1 Thread vs. others speedup: 1.37-1.90

46

Fig. 1 Fig. 2

• Dataset: 8.2GB

ANL 2014 - Chicago

Job Assignment

47

• KMeans: – 1.01 (8M) and 1.10-1.14 (for others)

• PCA (2 iterations):– Speedups : 1.19-1.68

ANL 2014 - Chicago

Heterogeneous Conf.

48

Overheads– KMeans: 1%– PCA: 1.1%, 7.4%, 11.7%

ANL 2014 - Chicago

Kmeans – Cost Constraint

• System meets the cost constraints with <1.1% error rate• Maximum # cloud instances is allocated error rate is again <1.1%

• System tries to minimize the execution time with provided cost constraint

49

ANL 2014 - Chicago

Prefetching and In-Memory Cache

• Overlapping application layer computation with I/O

• Reusability of already accessed data is small

• Prefetching and caching the prospective chunks– Default is LRU– User can analyze history and

provide prospective chunk list

• Cache uses row-based locking scheme for efficient consecutive chunk requests

50

Informed Prefetching

prefetch(…)

ANL 2014 - Chicago 51

Local Remote Hybrid0

100

200

300

400

500

600

700

800

900

1000FPC2P – 4IO4P – 8IO

Exec

ution

Tim

e (s

ec)

Local Remote Hybrid0

200

400

600

800

1000

1200OriginalFPC

Exec

ution

Tim

e (s

ec)

Performance of KMeans

Speedups

Local Remote Hybrid

FPC 0.75 1.30 1.12

Speedups w/ multithreading

Local Remote Hybrid

2P - 4IO 1.25 1.17 1.19

4P - 8IO 1.37 1.16 1.21

4P – 8IO vs Orig.

1.03 1.51 1.36

• NPB dataset• Comp ratio: 24.01% (180GB)• More computation

– More opportunity to fetch and decompression