23
Development and use of MonALISA high level monitoring services for Meta-Schedulers Stratos Efstathiadis a , Levente Hajdu a , Jerome Lauret a , Iosif Legrand b a Brookhaven National Laboratory, b California Institute of Technology CHEP’04 27 th September – 1 st October, 2004 Interlaken, Switzerland

Development and use of MonALISA high level monitoring services for Meta-Schedulers

Embed Size (px)

DESCRIPTION

Development and use of MonALISA high level monitoring services for Meta-Schedulers. CHEP’04 27 th September – 1 st October, 2004 Interlaken, Switzerland. Stratos Efstathiadis a , Levente Hajdu a , Jerome Lauret a , Iosif Legrand b - PowerPoint PPT Presentation

Citation preview

Page 1: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Development and use of MonALISA high level monitoring

services for Meta-Schedulers

Stratos Efstathiadisa, Levente Hajdua, Jerome Laureta, Iosif Legrandb

a Brookhaven National Laboratory, b California Institute of Technology

CHEP’0427th September – 1st October, 2004

Interlaken, Switzerland

Page 2: Development and use of MonALISA high level monitoring services for Meta-Schedulers

OUTLINE

a. Queue Monitoring.

b. Description of the mechanism providing monitoring data to SUMS using the MonALISA monitoring Framework.

c. First tests using queue monitoring in queue selection mechanisms.

d. Conclusions & Plans

Page 3: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Motivations

Possible phases in Grid Scheduling: Resource Discovery

Information Service, JINI, …

Status of Available Resources Schedulers look into improving application turnaround time

by using:• Current and past Resource status (Load, Memory, Space,

Number of running/pending jobs, bandwidth etc...)• Predictions of the status of the Resources over a specific

time interval or point in the future [NWS …]

Job Execution

Queue Monitoring, in particular, is part of the second phase

Page 4: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Queue Monitoring

Queue Monitoring provides aggregate status of a queuing system. It is not job monitoring. It does not provide status of individual jobs but rather the overall status of the queuing system.

Why is Queue Monitoring important ? Resource Brokers need info about the state of Local Resource

Management Systems (LRMS) in order to make decisions.

Resource Brokers do not have control over LRMS (they cannot decideon job priorities, on which particular host a job will run … ), they cannot change local decision making mechanisms or local Usage Policies (UP) which are controlled by the resource owners.

Global policies may need to adjust depending on local ones.

Page 5: Development and use of MonALISA high level monitoring services for Meta-Schedulers

As a starting point, we reused work done by the GGF/GLUE Schema.

The Computing Element (CE) represents the entry point into a Queuing System. One CE per Queue.

Attributes in the CE State Object (per queue):RunningJobs: Number of currently running jobs.

TotalJobs:  Number of jobs in the CE (RunningJobs+WaitingJobs)

Status: States a queue can be in (Queuing, Production, Closed, etc) WaitingJobs: Number of jobs that are in a state different than running.

WorstResponseTime: Worst time between job submission till when job starts its execution in sec

EstimatedResponseTime: Estimated time between job submission till when job starts its execution in sec

FreeCPUs: Number of free CPUs available to a scheduler.

Page 6: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The MonALISA Service System

Queue Monitoring Module

Page 7: Development and use of MonALISA high level monitoring services for Meta-Schedulers

One monitoring module collects information for all queues/pools. Clients requesting data specify parameters as Farm/Cluster/Node/ParameterName

Page 8: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Initial Implementation

Queue Monitoring using the MonALISA framework:

We are developing a ML Monitoring Module that provides the values of the Attributes of the CE Status Objects to be installed at each site of the group. The ML Monitoring Module will provide values for the same attributes for the most popular LRMS.

Monitoring data from each site are available as a Web Service.

Integrated Web Service Client into SUMS.

Solution is not scalable, time consuming and bypasses many of the ML framework features.

Page 9: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The MonALISA Web Repository

More details in Iosif’s presentation.

Page 10: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The same WS methods can retrieve monitoring data from either a monitoring site or a Web Repository.

Page 11: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The Web Repository Solution

Scalable solution: ML services are automatically discovered (LUSs)

The ML Web Repository provides aggregate monitoring data (real time and historical values) for several sites in a group.

The web service client in SUMS did not need to change.

Issues Retrieving monitoring data from web Service: 1) Averaged (mediated) data2) Slightly delayed data.3) Single point of failure (Web Repository may become

unavailable)

Page 12: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Local Pseudo-Client Solution

The pseudo-client was provided by the ML developers. Modified to spawn another thread for each connection. Provides latest, un-averaged data Fast monitoring data retrieval Easy deployment. Several local Pseudo-client deployments are possible. The Web Repository solution is still available as a fail-

over when pseudo-clients are unavailable.

No local DataBase available, so no historical data available

The ML developers have provided additional WS methods to access real-time, unmediated values: getLastValues() and getFilteredLastValues(Farm,Cluster,Node,Parameter)

Page 13: Development and use of MonALISA high level monitoring services for Meta-Schedulers
Page 14: Development and use of MonALISA high level monitoring services for Meta-Schedulers
Page 15: Development and use of MonALISA high level monitoring services for Meta-Schedulers

First tests using the queue monitoring information for queue selection in SUMS

Two policies are implemented in the STAR Unified Meta-Scheduler:

one is based on submitting jobs to alternating queues (Passive Policy)

the other based on selecting queues using monitoring data: the queue monitoring attributes (Monitoring Policy).

For the testing we used local resources (LSF Queues).

For each jobs we recorded submitTime, startTime, endTime, queue, etc to calculate the actual job Pending, Running times.

Page 16: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The STAR Unified Meta-Scheduler

Page 17: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The Monitoring Policy in SUMS

The Monitoring Policy uses the Response Time (RT) of each queue to decide where jobs will be submitted.

This is calculated for every job that is Pending or Running or finished within the last hour.

EstimatedResponseTime = (startTimei – submitTimei)/NJobs

Page 18: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Passive Policy

Monitoring Policy

Page 19: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Typical case of clear choice between queues based on Response Time, but ….jobs submitted to the selected queue ended up pending longer than jobs submitted to the queue with worse Response Time.

Page 20: Development and use of MonALISA high level monitoring services for Meta-Schedulers

The number of running jobs for the “chosen” queue is small, but still thesubmitted jobs remained pending for a longer than expected period.

Page 21: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Both clusters are saturated. One of the two, though, with production jobs we do not haveany information about. These “external” jobs changed the profile drastically.The Response Time, being an average over time, did not “react” quickly.

Page 22: Development and use of MonALISA high level monitoring services for Meta-Schedulers

First testing results In a balanced cluster, our tests were successful, an indication that queue information could provide a good selection mechanism.

In saturated cases, our approach did not lead to satisfactory results.

The turnaround rate of Pending to Running jobs over a time interval could provide a better (next) approach.

Using the MonALISA monitoring framework we were able to establish a reliable mechanism that provided monitoring data to the STAR Unified Meta-Scheduler.

Page 23: Development and use of MonALISA high level monitoring services for Meta-Schedulers

Plans:

1) Continue studying and testing of monitoring policy.

2) Grid testing

3) How to handle special situations (when monitoring data are unavailable, network problems …)

4) Can we provide information per user (fairshare policy…) ?

http://www.star.bnl.gov/STAR/comp/Grid/Monitoring/MonaLisa/