SciDAC SSS Infrastructure and BCM Update 10/2002

1
SciDAC SSS Infrastructure and BCM Update 10/2002

• Narayan Desai
• desai_at_mcs.anl.gov
• Argonne National Laboratory

2
Overview

• Infrastructure
• Service Directory
• Event Manager
• SSSlib
• Bindings
• Wire Protocol Modules
• Build and Configuration Management
• Cluster Hardware Infrastructure
• Build System
• Node State Manager
• Abstraction

3
Service Directory Status

• Static Schemas
• Complete Error Handling
• Deployed
• Reliable at Moderate Scale

4
Event Manager

• Usage Model
• Stable Schemas
• Relatively Complete Error Handling
• Completely Rewritten
• Received Moderate Testing
• Potential Subscription Type Differentiation

5
SSSlib

• Robust
• New C implementation
• Bindings
• C
• Java
• Python
• Perl
• Wire protocol modules
• Basic
• Challenge
• http-rm (in development)
• http (in development)

6
BCM Abstraction

• 2nd try Abstraction
• Components
• Node Manager
• Cluster Control
• Didnt Work
• 3rd is the charm
• Components
• Cluster Hardware Infrastructure
• Build System
• Node State Manager
• Why we think it will work

7
Cluster Hardware Infrastructure

• Handles new node integration
• Abstracts cluster hardware infrastructure
• Power Controllers
• Serial Consoles
• Bios Issues
• Node Inventory
• Node Identification

8
Node State Manager

• Node State Tracking
• Basic state monitoring
• Node Administrative State
• Online/Offline
• Corrective Action Facility
• Pull the plug on bad nodes
• Unknown criterion

9
Build System

• Disk Setup
• Software Configuration
• Software Deployment

10
Example Node Introduction

• New node added
• CHI identifies node
• CHI hands off control of node to BS
• BS builds node into proper configuration
• BS hands off control of node to NSM
• NSM can set node administrative state
• In case of errors, node can be rebuilt or other
  actions can be taken

11
Soon

• Start to standardize events
• Enhance event data format?
• Implement more wire protocols
• Complete hot-swap tests for BCM components
• Logic for the node state manager
• Implement a modular cluster hardware
  infrastructure

12
Issues

• Abstraction problems
• BCM model unknown
• Implementation feedback
• Multiple implementations help