High Performance Applications over the Cloud: Gains and Losses

(1)

High Performance Applications over the Cloud: Gains

and Losses

Dr. Leila Ismail

Faculty of Information Technology

United Arab Emirates University

[email protected] http://citweb.uaeu.ac.ae/citweb/profile/leila

Collaboration:

Nimrod-G Project

Ankabut Users’

Meeting 11 – 12

January 2012

(2)

2

Agenda



Introduction



What is cloud computing?



Benefits of cloud computing



What is HPC?



Cloud computing challenges



Go to market: experimentations with

Amazon cloud

(3)

Cloud Computing

Cloud computing is a model for enabling

ubiquitous, convenient, on-demand

network access to a shared pool of

configurable computing resources (e.g.,

networks, servers, storage, applications,

and services) that can be rapidly

provisioned and released with minimal

management effort or service provider

interaction. This cloud model is

composed of five essential

characteristics, three service models, and

four deployment models.

NIST – SP800-145

September 2011

(4)

4

Cloud Computing Characteristics

1. On-demand self-service

2. Broad network access

3. Resource pooling

4. Rapid elasticity

(5)

Cloud Computing Service Models

Physical Layer/hardware resources (server, storage and network)

Middleware (automatic scheduling, resource allocation, provisioning, virtualization, metering, billing, data

storage management, etc.) Software as a Service (SaaS)

(providers’ applications) Platform as a Service (PaaS)

(means of deploying users’ applications: programming languages, libraries, tools, etc.)

Infrastructure as a Service (IaaS)

(provisioning of processing, storage and network to deploy OSs and users’ applications)

(6)

6

Cloud Computing Types

1)

Private cloud



Exclusive use by a single company

2)

Community cloud



Exclusive use by a specific group of consumers/companies

of common interest

3)

Public cloud



_{Used by the public}

4)

Hybrid cloud

(7)

Cloud Computing Challenges

•

Fast allocation of instances



Performance of dynamic allocation,

provisioning, virtualization, etc.

• Scalable infrastructure



Monitoring



Dynamic construction of loosely coupled

architectures

• _{Usability and user trust in pricing}

• _{Security models}

(8)

Where is HPC used?

Today, HPC directly impacts our lives in ways we seldom

recognize. Product packaging, golf balls, and even diapers

are all better because of HPC.

8

(9)

Portfolio risk analysis and optimization, derivatives pricing, fraud detection....

FINANCE

Simulate mechanical and electronic systems before building costly prototypes

ENGINEERING

Weather and climate forecasting

WEATHER AND CLIMATE

Simulate the effects and prognosis of chemotherapy,

radiation, and surgery

RESEARCH/DEVELOPMENT

Render photo-realistic scenes, customer materials, product

illustrations and videos

ENTERTAINMENT OIL AND GAS

Analyze seismic data to find oil and gas and determine how to extract it in the most efficient way

Advancements in HPC

(10)

Core 4 Core 3 Core 2 Core 1 Cache

A serial problem.

10

(11)

A Semi-Parallel Execution

(12)

Parallel Task

12

(13)

Distributed Memory Shared Memory

Easy to Program (Shmem/OpenMP)

Easy to Administer Difficult to Scale_~64p

SMP Scales Well ~1024p Easy to Program (Shmem/OpenMP/MPI) Easy to Administer Proprietary/Expensive NUMA

Cluster Difficult to Program

(MPI)

Less Flexible to Administer

Price Competitive

Highly Scalable ~4096p

(14)

HPC in the Region



_{Oil and Gas is the largest segment that uses HPC}

in this region



Followed by Education



22.4% of TOP500 HPC systems are used by

Researchers



15% of TOP500 HPC systems are used by

Academia



57.6% of TOP500 HPC systems are used by

Industry

14 Ankabut Users' Meeting 11-12 January 2012

(15)

(16)

HPC Cloud Software

Components

Operating System

File System

Compilers Libraries _ManagementCluster

Runtime Tools Application Software Visualization In teg rati on and T es ting Sy st em Ma na gem en t

Backup and Archiving

16

Cloud API

(17)

Today’s Limitations of the HPC



Memory Bandwidth

 _{Commodity memory interfaces [SDRAM, RDRAM, DDRAM]}  Up to 85% CPU Cycles Waiting for Memory



Communications fabric/CPU/Memory Integration

 Limits bandwidth and latency and communication semantics



Node and system packaging density

 _{Commodity components and cooling technologies limit densities}



Applications

 _{Has little functionality of checkpointing}  Don’t survive multiple failure

 _{Are not “many-core” efficient}



… but departmental and single applications clusters are highly

successful

(18)

Cloud computing in 2012



_{$900 million investment in datacenters in USA and}

Ireland by Microsoft



_{$600 million datacenter in Oklahoma by Google}



$100 million in Ireland, Singapore, Taiwan and Hong

Kong by Google



$450 million datacenter in North Calorina by

facebook to support image sharing and SaaS cloud

computing

Article by Oleg Komissarov , VP of Entreprise Solutions, New York, 15 December 2011

(19)

Amazon Cloud Computing

 Amazon Elastic Compute Cloud (EC2)

 Use web service API to create, start, and terminate instances

 Choose instances type(s) and OS

 Standard (small, medium, large)  Micro

 High-memory (extra-large, double extra-large, quadruple extra-large)  High-CPU (medium, extra-large)

 Cluster compute (quadruple extra-large, eight extra-large)  Cluster GPU (quadruple extra-large)

 Computing Unit ≈ 1.0 – 1.2 GHz 2007 Opteron or 2007 Xeon or early-2006 1.7 GHz Xeon CPU

 Attach persistent storage to instances

(20)

Experimental Environment



Amazon Elastic Compute Cloud (EC2) (16 virtual cores)

(21)

Cost per Use

Platform Type Price/hour ($) per Node Number of Machines Uses