An Operational Comparison Between CODINE And Monsanto-NQS

Abstract

This paper provides a detailed review of CODINE, a detailed comparison of CODINE against Monsanto-NQS in light of the stated needs of UK HE [1], and a look at what each product hopes to deliver in its next revision.

This paper concludes that currently neither product is superior on features, and that Monsanto-NQS is currently superior on cost.

This paper is based upon an earlier paper by Professor R Hynds of Imperial College, London, presented at the JISC NTI Clusters Workshop, 12th-13th December 1994. We gratefully acknowledge the assistance of Professor Hynds and his staff during the preparation of this paper.

Contents

• Introduction
• CODINE
• Comparison Of CODINE and Monsanto-NQS
• Future Directions
• Pricing
• Conclusions
• References

Introduction

The Joint Information Services Committee (JISC), through its New Technologies Initiative (NTI), has looked at the two very different approaches to this technology - Imperial College has looked at CODINE, a product based on the DQS approach, while the University of Sheffield has looked at Monsanto-NQS, a product based on the NQS approach.

Imperial College presented a paper in December 1994 which looked at CODINE, and subsequently the NTI committee has decided to fund support for CODINE.

The purpose of this paper is to provide an alternative review of CODINE, to provide a detailed comparison of CODINE against Monsanto-NQS, and to show what both products intend to provide in their next revisions.

Relevant Literature

• A Comparison of Queueing, Cluster and Distributed Computing Systems, Joseph A. Kaplan, Michael L. Nelson, NASA Langley Research Center June 1994.

  This paper looks at a large number of batch systems, including CODINE, and several versions of NQS. All systems are compared against a prepared checklist, and this is followed by discussion of major points about each product.

  Much of the evaluation is based solely on supplied documentation and discussion with the vendors - only two of the twelve systems were actually installed and tested.

  No conclusions about the superiority of one product over another are drawn.

• Systems Analysis - Batch Processing Systems, Stuart Herbert, Academic Computing Services, University of Sheffield, October 1994.

  This paper looks at a small number of batch processing systems, with a view to recommending a product to be further developed to meet the requirements of UK Higher Educational sites [1]. The selection of products was determined by knowledge of the availability of the systems. The features of the different products are compared, and the concerns about each product are documented.

  Much of the evaluation is based on the attempted installation of each product, with secondary evaluation based on documentation where installation attempts were unsuccessful.

  This paper concludes that Monsanto-NQS is the most suitable product to develop further to meet the stated needs of UK Higher Education.

• CODINE And LoadLeveller - Packages for Running Serial Batch Work in a Workstation Cluster Environment, Professor R.J.Hynds, Centre for Computing Services, Imperial College, December 1994.

  This paper looks at the suitability of CODINE and LoadLeveller for batch processing work over a period of several months. The features of each product are discussed, along with their implications for practical application, and a limited comparison with Monsanto-NQS is included.

  The evaluation is the result of a JISC NTI project to look at these packages.

  This paper concludes that the use of CODINE and LoadLeveller is of significant advantage, that both products are superior to Monsanto-NQS, and that checkpointing is a feature which users are loathe to use.

These issues were then raised with GENIAS, authors of CODINE, and their comments were included into a revision of Professor Hynds' original paper.

The Issues

• CODINE has features which simplify administration when compared to Monsanto-NQS.

  However, batch processing systems are intended to be installed, configured, and then left unaltered - both CODINE and NQS are well suited to this task.

• CODINE and NQS cannot interact - sites must realistically choose one or the other.

• While elements of the Graphical User Interface (GUI) of CODINE are of value, using the GUI places heavy demands on computing hardware.

• CODINE does not provide specific support for the extra features of the Silicon Graphics IRIX operating system. A number of UK sites require support of these features.

• CODINE cannot be used to share resources between different institutions (over SuperJANET), because CODINE requires that user and group id values are the same across all CODINE nodes.

• CODINE's support of interactive use is limited to starting an `xterm' process on a remote machine, something which Monsanto-NQS can be configured to do.

Such software should :

• protect the owner's normal usage

• permit scheduling and load balancing of jobs

• control the number of jobs

• provide job checkpointing

• provide job migration

• Main computing resources distributed to departments

• Large research income - many workstations

• Substantial batch work requirement

• Four main vendors - DEC, IBM, SGI, SUN

• Wide variety of types of batch work

• NQS already in use in many departments

• CONDOR, developed at the University of Wisconsin-Madison by Miron Levy and co-workers, is used to provide the checkpointing facilities.

• DQS, developed at Florida State University by Tom Green and co-workers, is used to provide the queueing and scheduling facilities.

CODINE was tested on each client-server pair (HOMOGENEOUS CONFIGURATION) then in combination between client-server pairs (HETEROGENEOUS CONFIGURATION). The DEC pair were tested under the OSF operating system.

Generally speaking the package performed according to its specifications.

General Features Of CODINE And LoadLeveller

• Graphical User Interface

• Support for batch jobs

• Support for job checkpointing

• Support for migration of checkpointed jobs

• Intelligent load balancing

• Definable job limits for specific users

• Definable priority levels for particular users

• Configurable queues with specific resource limits

• Suspension and enabling of queues

• Definable queue owners

• Definable 'managers' and 'users'

• Supports centralised job accounting

• Usable in both homogeneous and heterogeneous clusters

• (Processes NQS scripts)*

• (Support parallel job environments e.g. PVM)

>  --------------------------------
>  No. Of Students	Cost (DM)
>
>     1	  - 20,000	6200
>  20,001 - 35,000	8600
>  35,001+		14800
>  --------------------------------