Virtual Workspaces in the Grid Kate Keahey [email protected] Argonne National Laboratory Ian Foster, Tim Freeman, Xuehai Zhang, Daniel Galron

Virtual Workspaces in the Grid

Kate [email protected]

Argonne National Laboratory

Ian Foster, Tim Freeman, Xuehai Zhang, Daniel Galron

09/01/05 Kate Keahey, Europar 2005

The Grid Metaphor

How do we store energy?

How do we charge for energy?

How do we reliably deliver energy?

What happens if a power station fails?

How do we ensure quality of service?

What elements make for a safe and efficient power Grid?

How do we make sure that supply meets demand?

Grid Computing is much harder: heterogeneous and multi-dimensional


The Missing Link in Grid Computing We need to define mechanisms for and dynamic

deployment and management of remote environments

Ideal environment is deployed

magichappens

Requirements: Flexibly define an environment

The more we can customize it, the more useful it is

Deploy and manage such environments Can such environments be deployed securely? How fast/dynamic can this deployment be? How can I control resources allocated to such an environment?

Dream up an ideal environment


Virtual Workspaces

Virtual Workspaces: environments that can be made available dynamically the Grid with well-understood properties

Examples: A TeraGrid node with well-defined software environment

and adjustable access and sharing policies A physical cluster booted to a desired configuration

(e.g. Cluster on Demand) An “ATLAS node” dynamically configured using Pacman A virtual machine configured to represent a specific

environment whose resource consumption can be controlled


Virtual Machines as Workspaces

Virtual Machines Highly customizable software configuration Enforcement properties

Grid 2004 paper: “Dynamic environments in the Grid” F. Cappello & lab: Comparison of different hypervisors

Pausing, serialization, migration Performance:

L X V USPEC INT2000 (score)

L X V ULinux build time (s)

L X V UOSDB-OLTP (tup/s)

L X V USPEC WEB99 (score)

0.00.10.20.30.40.50.60.70.80.91.01.1

SOSP 2003 paper: “Xen and the Art of Virtualization”


Workspace Template Aspects Environment Aspect (workspace meta-data)

Generic information Name, time to live, etc.

Software partition information Software description: OS, “OSG configuration”, “application partition”,

etc. Software meta-data is bundled with the actual software and attested by

its issuer

Services: ssh, GRAM, pre-configured job Deployment independent

Resource allocation request (deployment time) Memory, disk, networking, etc.

See GGF JSDL standard

On deployment the actual resource allocation information becomes available


Atomic Workspaces and Virtual Clusters

Atomic workspace One or more homogeneous workspaces

The only differences are in names

Cluster/aggregate workspace A set of interdependent heterogeneous

workspaces Example: a headnode and a set of worker nodes

Interdependencies of metadata are expressed through tags and pointers


Deploying Workspaces in the Grid

Define workspace environment

Manage workspace

Negotiate workspace deployment characteristic

WorkspaceWizard

(VW Factory)

Workspace Management

Service(VW Repository)

Workspace Service

(VW Manager)

request a workspace

workspace meta-data

manage workspace environment

workspace metadata

Workspace

terminate workspace deployment

negotiate workspace deployment

manage/monitor/renegotiate workspace deployment

manage activities within the workspace


Current Implementation Current prototype using Globus Toolkit 4

Leveraging standard Grid Service features such as lifetime management

Workspace Wizard Returns workspace meta-data Very rudimentary implementation

Workspace Serivce Create: takes workspace meta-data and a deployment

descriptor Manage:

renegotiate resource allocation (moving towards a WS-Agreement model) Also traditional Grid Service management: TTL, etc.

Destroy Different options: pause, shutdown or destroy


How dynamic is the deployment?

Automatic Protocol-based Moving towards better articulation of migration Renegotiation of resource allocation

How fast is this deployment? Deployment of workspace for EMBOSS suite:

Manual: ~45 minutes Based on pre-configured Vmware VMs: ~6 minutes Based on pre-configured Xen VM: < 1 second

How much overhead does workspace deployment add over what we have today?


Workspace Service: Individual Workspaces

Using a paused VM allowed us to “save” on initiation time

8

8

8

0.7

0.7 1.7

0.8

0.8

0 2 4 6 8 10 12

a)

b)

c)

job startup scenario

time (in seconds)

VM setup

VM boot

job setup

GRAM job

a) GRAM job executionb) GRAM job execution in a paused Xen VMc) job execution in a booted Xen VM (pre-configured job)


Workspace Service: Virtual Clusters


Deploying Workspaces Across Technologies

Basic node configuration (+/-boot from image) Cluster on Demand, PXE, bcfg On the order of many minutes (~30 minutes)

Refining configuration, creating access Dynamic account with workspace service: < 1s

(mostly GT4 request processing time) Refining Installation: ~2 hours to configure an ATLAS

node using Pacman Virtual machines

Deploying images Xen: ~100 ms VMware Workstation: ~ several seconds


Nested Workspaces

Physical machineprocure hardware

program program program

…VM

Hypervisor/OSdeploy hypervisor/OS workspace

VM VMdeploy VM workspace (with hypervisor/OS)


Computational Grids

Hypervisor 1 Hypervisor 2 TeraGrid Configuration

Grid Power Station

Grid Protocols

Clients


Conclusions We need mechanisms for dynamically deploying and

managing environments in the Grid Workspaces are a fundamental building block of a Grid

environment Workspaces are implemented using wide variety of

technologies VMs are a highly promising one: a “computon” for the Grid

Workspace aspects Deployment-independent environment definition Deployment-time policy and enforcement negotiation

Many challenges remain Security and deployment issues Protocols, protocols, protocols Leveraging the opportunities

Documents

Virtual Workspaces in the Grid Kate Keahey [email protected] Argonne National Laboratory Ian Foster, Tim Freeman, Xuehai Zhang, Daniel Galron