Monitoring and Fault Tolerance

Helge Meinhard / CERN-IT

OpenLab workshop

08 July 2003

NodeConfiguration

SystemMonitoring

System

InstallationSystem

Fault MgmtSystem

Monitoring and Fault Tolerance: Context

History (1)

In the 1990s, “massive” deployments of Unix boxes required automated monitoring of system state

Answer: SURE Pure exception/alarm system No archiving of values, hence not useful for

performance monitoring Not scalable to O(1000) nodes

History (2)

PEM project at CERN (1999/2000) took fresh look at fabric mgmt, in particular monitoring

PEM tool survey: Commercial tools found not flexible enough and too expensive; free solutions not appropriate

Architecture, design and implementation from scratch

History (3)

2001 - 2003: European DataGrid project with work package on Fabric Management Subtasks: configuration, installation,

monitoring, fault tolerance, resource management, gridification

Profited from PEM work, developed ideas further

History (4)

In 2001, some doubts about ‘do-it-all-ourselves’ approach of EDG WP4

Parallel to EDG WP4, project launched to investigate whether commercial SCADA system could be used

Architecture deliberately kept similar to WP4

Monitoring and FT architecture (1)

Monitoring: Captures non-intrusively actual state of a system (supposed not to change its state)

Fault Tolerance: Reads and correlates data from monitoring system, triggers corrective actions (state-changing)

Monitoring and FT architecture (2)

MonitoringSensor

Agent (MSA)

SensorSensorSensor

Localcache

LocalconsumersLocalconsumersLocalconsumers

API

MR – Monitoring Repository

WP4: MR code with lower layer as flat file archive, or using Oracle

CCS: PVSS system

DBAPI

Monitoring and FT architecture (3)

MSA controls communication with Monitoring Repository, configures sensors, requests samples, listens to sensors

Sensors send metrics on request or spontaneously to MSA

Communication MSA – MR: UDP or TCP based

Monitoring and FT architecture (4)

FT system subscribing to metrics from monitoring subsystem

Rule-based correlation engine takes decisions on firing actuators

Actuators controlled by Actuator Agent, all actions logged by monitoring system

Deployment (1)

End 2001: Put early versions of MSA and sensors on big clusters (~800 Linux machines), sending data (~100 metrics per machine, 1/min…1/day) to a PVSS-based repository

At the same time, ~300 machines started sending performance metrics into flat file WP4 repository

Deployment (2)

Sensors more refined over time (metrics added according to operational needs)

Both exception and performance oriented sensors now deployed in parallel (some 150 metrics per node)

More special machines added, currently ~1500 machines being monitored

Test in May 2003: some 500 metric changes per second into the repository (~150 changes/s after “smoothing”)

Deployment (3)

Repository requirements: Repository API implementation Oracle based fully functional alarm display for operators

Currently using both an Oracle-MR based repository, and a PVSS based one

Operators using PVSS based alarm screen as alternative to Sure display

Deployment (4)

Interfaces: C API available, simple command line interface by end July, prototype Web access to time series of a metric available

Fault tolerance: Just starting to look at WP4 prototype

Configuration of monitoring: ad-hoc, to be migrated to CDB

Outlook

Near term: Production services for LCG-1 Add more machines (e.g. network), metrics Software and service monitoring

Medium term (end 2003): Monitoring for Solaris and Windows, …

2004 or 2005: Review of chosen solution for monitoring and FT Some of 1999 arguments no longer valid Will look at commercial and freeware solutions

Machine control

High level: interplay of State Management System, Configuration Management, Monitoring, Fault Tolerance, …

Low level: Past: CPU boxes didn’t have anything (5 rolling tables

with monitors and keyboards per 500…1000 machines), disk and tape servers with analog KVM switches

Future: Have investigated various options, benefit/cost analysis. Will go to serial consoles on all machines, 1 head node per 50…100 machines with serial multiplexers