Ccs Technical Press Workshop Drilldown

26
yril Faenov, Director of HPC icrosoft Corporation ttp://www.microsoft.com/hpc

Transcript of Ccs Technical Press Workshop Drilldown

Page 1: Ccs Technical Press Workshop Drilldown

Kyril Faenov, Director of HPCMicrosoft Corporationhttp://www.microsoft.com/hpc

Page 2: Ccs Technical Press Workshop Drilldown

Excel “12”Excel “12”

BrowserBrowser100% thin100% thin

View and Interact

CustomCustomapplicationsapplications

Web Services Access

Excel “12”Excel “12”clientclient

Open Spreadsheet/Snaps

hot

Author & Publish Spreadsheets

Excel Services

Page 3: Ccs Technical Press Workshop Drilldown

Scaling Excel

DesktopDesktop ServersServers ClustersClusters

Excel ServicesExcel Services

Excel Services on Windows Excel Services on Windows Compute Cluster Server 2003Compute Cluster Server 2003

Excel “12”Excel “12”

Page 4: Ccs Technical Press Workshop Drilldown

Excel & Windows CCSCustomer requirements

Faster spreadsheet calculationFree-up client machines from long-running calculationsTime/mission critical calculations that must runParallel iterations on models

Example scenariosSchedule overnight risk calculationsFarm out analytical library calculationsScale-out Monte Carlo iterations, parametric sweeps

Page 5: Ccs Technical Press Workshop Drilldown

Windows Compute Cluster Server 2003

Page 6: Ccs Technical Press Workshop Drilldown

CCS Key FeaturesNode Deployment and Administration

Task-based configuration for head and compute nodesUI and command line-based node managementMonitoring with Performance Monitor (Perfmon), Microsoft Operations Manager (MOM), Server Performance Advisor (SPA), & 3rd-party tools

Integration with existing Windows and management infrastructureIntegrates with Active Directory, Windows security technologies, management, and deployment tools

Extensible job scheduler3rd-party extensibility at job submission and/or job assignmentSubmit jobs from command line, UI, or directly from applicationsSimple job management, similar to print queue management

Secure MPIUser credentials secured in job scheduler and compute nodesStandardized MPI stackMicrosoft provided stack reduces application/MPI incompatibility issues

Integrated Development EnvironmentOpenMP Support in Visual Studio, Standard EditionParallel Debugger in Visual Studio, Professional Edition

Page 7: Ccs Technical Press Workshop Drilldown

MPIMPINetworkNetwork

PrivatePrivateNetworkNetwork

PublicPublicNetworkNetwork

Corporate IT InfrastructureCorporate IT Infrastructure

Compute NodeCompute NodeHead NodeHead Node

ADAD

DNSDNS

DHCPDHCP

Windows Windows UpdateUpdate

MonitoringMonitoring

• MOMMOM

• 33rdrd party party

SystemsSystemsManagementManagement

• SMSSMS

• 33rdrd party party

Compute Cluster

Typical Cluster Topology

Admin / User ConsAdmin / User Cons

RISRIS

Job SchedulerJob Scheduler

MPIMPI

ManagementManagement

NATNAT

Node ManagerNode Manager

MPIMPI

ManagementManagement

Compute NodeCompute Node

Node ManagerNode Manager

MPIMPI

ManagementManagement

Page 8: Ccs Technical Press Workshop Drilldown

PrivatePrivateNetworkNetwork

PublicPublicNetworkNetwork

Compute NodeCompute Node Head NodeHead Node

ADAD

Admin / User ConsAdmin / User Cons

ManagementManagementManagementManagement

Corp\JohnCorp\John

Corp\Sally Corp\Sally

AdministratorsAdministrators Local Local

UsersUsers Local Local

AdministratorsAdministrators Local Local

UsersUsers Local Local

PropagationPropagation

Single Point of User management

Page 9: Ccs Technical Press Workshop Drilldown

New Admin ConsoleNot building a new systems management paradigm

Leveraging MMC for simple management (MMC 3.0)

MMC used for cluster administration

Has 5 branches

Start Page

To Do List

Node Management

Remote Desktop

Performance

Page 10: Ccs Technical Press Workshop Drilldown

New Job Submission and Monitoring UIWin32 App

Centrally manage the entire job queue

Print queue manager like experience

Ability to save and submit jobs using templates

Automatically create multiple tasks

Page 11: Ccs Technical Press Workshop Drilldown

Job/Task Conceptual ModelSerial JobSerial Job

TaskTask

ProcProc

Parallel MPI JobParallel MPI Job

TaskTask

ProcProc ProcProcIPCIPC

Parameter Sweep JobParameter Sweep Job

TaskTask

ProcProc

TaskTask

ProcProc

TaskTask

ProcProc

Task Flow JobTask Flow Job

TaskTask

TaskTask

TaskTask

TaskTask

Page 12: Ccs Technical Press Workshop Drilldown

Job Life Cycle

Page 13: Ccs Technical Press Workshop Drilldown

Secure channelSecure channel

credenticredentialal

End-To-End Security

Node Node MgrMgr

Secure channelSecure channel

credenticredentialalClientClient SchedulSchedul

erer

MSDEMSDE

Data Protection Data Protection APIAPI

credenticredentialal

TaskTask

SpawnSpawn

Logon Logon tokentoken

Logon as Logon as useruser

DB/FS

ActiveActiveDirectoryDirectory

DataData

KerberosKerberos

KerberosKerberosLSALSA

Automatic Automatic Ticket renewalTicket renewal

Page 14: Ccs Technical Press Workshop Drilldown

Command Line InterfaceJob job new [job_terms] create a job container

job add jobID [task_terms] add tasks to a job

job submit /id:jobid submit job / tasks

job submit [job_terms][task_terms] submit a job

job cancel jobID cancel a job

job modify [options] modify a job

job requeue JobID requeue a job

job list list jobs in the cluster

job listtasks list tasks of a job

job view JobID view details of a job

Task task view view details of a task

task cancel cancel a task

task requeue requeue a task

Cluscfg cluscfg view view details of a cluster

cluscfg params/setparams view/set config params

cluscfg listenvs/setenv list / set cluster wide env

cluscfg delcreds/setcreds set /delete user cred

Node node list list nodes

node approve approve selected node

node pause pause selected node

node resume resume selected node

Page 15: Ccs Technical Press Workshop Drilldown

MS MPI Leverages Winsock Direct

Networking HardwareNetworking Hardware

Virtual Bus DriverVirtual Bus Driver

IPIP

TCPTCPUser Host User Host Channel Channel

Adapter DriverAdapter Driver

IB WinSock Provider IB WinSock Provider DLLDLL

MiniporMiniport t

(GigE)(GigE)

MiniporMiniport t

(IPoIB) (IPoIB) Kernel API (verbs Kernel API (verbs based)based)

User API (verbs User API (verbs based)based)

Host Channel Adapter Host Channel Adapter DriverDriver

GigE GigE RDMA RDMA

WinSock WinSock Provider Provider

DLLDLL

IHV-IHV-provided provided

componentcomponent

OS OS componentcomponent

NDISNDIS

WinSock DLLWinSock DLL

IBIBw/ RDMAw/ RDMA

GigE GigE

w/ RDMAw/ RDMA

EthernetEthernet

User ModeUser Mode

Kernel ModeKernel Mode

Switch trafficSwitch trafficbased on sub-netbased on sub-net

Manage Manage hardware hardware resources in resources in user space (eg., user space (eg., Send and Send and receive queues)receive queues)

HPC HPC ApplicationApplication

MPIMPI

Winsock SwitchWinsock Switch

Page 16: Ccs Technical Press Workshop Drilldown

Futures

Page 17: Ccs Technical Press Workshop Drilldown

The Future: Supercomputing On A Chip

IBM Cell processor256 Gflops today

4 node personal cluster => 1 Tflops

32 node personal cluster => Top100

Microsoft Xbox3 custom PowerPCs + ATI graphics processor

1 Tflops today

$300

8 node personal cluster => “Top100” for $2500 (ignoring all that you don’t get for $300)

Intel many-core chips“100’s of cores on a chip in 2015” (Justin Rattner, Intel)

“4 cores”/Tflop => 25 Tflops/chip

17

Page 18: Ccs Technical Press Workshop Drilldown

Evolution Of HPCEvolving Scenarios Key Factors

Batch computing on supercomputers

Compute cycles are scarce and require careful partitioning and allocationCluster systems administration major challengeApplications split into UI and compute parts

Interactive computing on departmental clusters

Compute cycles are cheapInteractive applications integrate UI/compute parts Emergence of turnkey personal clusters

Complex workflow spanning applications

Compute and data resources are diffused throughout the enterpriseDistributed application, systems and data management is the key source of complexityMultiple applications are organized into complex workflows and data pipelinesFocus on service orientation and web services

Interactive Interactive Computation Computation and and VisualizationVisualization

Manual, batchManual, batchexecutionexecution

IT IT MgrMgr

SQL

Page 19: Ccs Technical Press Workshop Drilldown

“Grid Computing”

A catch-all marketing termDesktop cycle-stealing

Managed HPC clusters

Internet access to giant, distributed repositories

Virtualization of data center IT resources

Out-sourcing to “utility data centers”

“Software as a service”

Parallel databases

Page 20: Ccs Technical Press Workshop Drilldown

HPC Grids And Web Services

Compute gridForest of clusters

Coordinated scheduling of resources

Data gridDistributed storage facilities

Coordinated management of data

Web ServicesGlue for heterogeneous platforms/applications/systems

Cross- and intra-organization integration

Standards-based distributed computing

Interoperability and composability

Page 21: Ccs Technical Press Workshop Drilldown

Cluster-based HPC

Page 22: Ccs Technical Press Workshop Drilldown

Intra-organization HPC

Page 23: Ccs Technical Press Workshop Drilldown

Virtual Organizations

Page 24: Ccs Technical Press Workshop Drilldown

Example - SC05 Bill Gates Keynote Demo

Page 25: Ccs Technical Press Workshop Drilldown

Resources

Microsoft HPC web sitehttp://www.microsoft.com/hpc/

Microsoft Windows Compute Cluster Server 2003 community sitehttp://www.windowshpc.net/

Microsoft Windows Compute Cluster Server 2003 Betahttp://www.microsoft.com/windowsserver2003/ccs/beta.mspx

Windows Server x64 informationhttp://www.microsoft.com/64bit/

http://www.microsoft.com/x64/

Windows Server System informationhttp://www.microsoft.com/wss/

Page 26: Ccs Technical Press Workshop Drilldown

© 2003-2005 Microsoft Corporation. All rights reserved.This presentation is for informational purposes only. Microsoft makes no warranties, express or implied, in this summary.