Current Status

[ Up ] [ Current Status ] [ Previous Status ]

Cornell Center for Materials Research&Intel Technology for Education 2000 Progress Report - Q2 1999 Status Report

	Windows NT
	Linux

Next Quarter Plans

Contacts

Equipment Utilization

Accomplishments

Windows NT

Installation and Maintenance Strategies

"We even accept the fact that we occasionally need to completely re-install the desktop OS and all of our applications, and restore all personal files from the latest backup." Mark Hall and Ralph Barker "The Death of the Desktop", Performance Computing, August 1998

Our first strategy is to make the installation and re-installation of WinNT and application software from scratch to be as simple and rapid a process as possible. We are using Ghost to save and load disk images on our AFS (Andrew File System, Transarc) fileserver via Samba, a public domain SMB server running under UNIX.

We have successfully deployed 8 of these machines and can re-install to an identical configuration quickly. However, this plan has run into a fatal flaw. New machines arriving have slightly different hardware configurations. A simple revision update in the ethernet controller rendered the boot disk inoperative. Once that was fixed a similar revision change in the SCSI controller made the previous images useless.

We are now working on a strategy that combines an installation "recipe", remote-distribution and an image. With this we will come up with a well defined minimal set of instructions that can be used to completely reinstall the OS. From this minimal installation we will run remote-distribution software (PcRdist) that will use a pull-mechanism to download various packages and configure the system and registry for printers and other resources. Finally we will make an image of this install and save it for later reinstallation as needed.

SMS

Systems Management Server, has been acquired from Microsoft. We have installed 2 servers and two clients and have started configuration and operation testing. This software is expected to primarily be used for system inventories, what software and hardware is available on various systems, and for software package distribution.

Facility Integration

CCMR Computing Facility is based on UNIX (AIX, SunOS, IRIX, Linux) servers and clients. Work is progressing on developing tools and techniques to integrate NT into this structure. We are using Transarc Corporation's AFS (Andrew File System) client for WinNT to leverage our existing AFS based file service, backup, authentication, and administration infrastructure for WinNT. We have purchased and installed NT-AFS for this purpose.

File Service

In contrast to NFS (Network File System) which is stateless and relies on local system caching, AFS uses it's own caching mechanism to keep frequently used files local to the client. This has the potential to greatly reduce network traffic and improve performance. Only the software that is used on that workstation is actually local to that station. However, additional memory overhead is significant and the type of files being accessed, many small or a few large files, can affect performance. Also, the initial access on a file takes longer as the client retrieves the file from the server and is placed into the client's local AFS caching system.

We are looking into the performance of this versus NT Server, Samba and local file access. We have seen some very irradic performance while exploring disks on the server. We are concerned about the interaction between the NT file cache and the AFS file cache. Both are in memory, both are essentially doing the same thing but with different mechanisms.

Authentication

For authentication we are experimenting with a GINA (Graphical Identification aNd Authentication?) replacement developed at Notre Dame. It is not desirable to maintain separate authentication systems for NT and UNIX for 500 users. The ND_GINA enables WinNT to use AFS Kerberos authentication for login. Roaming profiles are stored in each user's AFS home directory. By executing several scripts at boot time and login/logout time, the ND_GINA also provides an infrastructure for workstation administration (see printing below for example). For other GINA projects, see the MIT NT GINA Information Website.

So far, we have modified, built, and installed the ND_GINA. We have demonstrated that authentication, fileserver access, and roaming profiles work. Three workstations are available for users, an other is being used by the CCMR Ion Beam Facility. Full-scale deployment is awaiting completion of development and implementation of our WinNT installation strategy mentioned above.

Inconsistancy of the AFS home directory access and subsequent absence of the user's profile seems to have been fixed with NT-AFS 3.4a patch 10. However, a Windows DLL periodicly fails to load which renders AFS inaccessible. Restarting fixes the problem.

Printing

Printing within CCMR is accounted. Therefore login and authentication are an issue on each system. For NT were are using lpr configured print services. This is being managed by the GINA modifications and PcRdist so that as printers are removed or installed the user's view, Printer Folder, is kept current.

Software

We are using Hummingbird's Exceed Software for Xwindow client access to our UNIX systems. We have also installed Microsoft Suite provided to the Intel grant equipment by a grant from Microsoft. Currently we have not installed any specific scientific software on these systems. We have purchased IDL, a visualization package and have licenses for Abaqus, finite element modelling, and the Portland Group FORTRAN compiler which is essential for the migration of some software.

Terminal Server

We have acquired Microsoft NT Terminal Server and will be installing and evalutating it next quarter. The expectation is that this will provide access to software that is licensed on only a few machines and access to Windows NT applications from UNIX X-Window based systems.

Utilization

Lam/Ast - Micro Electro Mechanical/Optical System simulations. This kind
of simulation uses finite element or finite difference software. We have written a FORTRAN program to do finite difference time domain simulations. A 2-D simulation takes up 128M RAM. We hope to do 3-D simulations soon. Recently our group started using SUPREME and PISCES simulators. Both of the simulation packages require at least 64M RAM to only to start running. To use its graphical display at the same time an analysis is running requires significant CPU and video capabilities and was not possible on existing PC equipment.

Revesz - Development of remote access capability to the CCMR Ion Beam facility.

Facility User Room - 3 NT machines are available for users in the computing facilities general user rooms.

IRG2 / Ast Install and maintain a datebase on glass. A commercial package containing 38,000 data has been located.

Linux

Installation and Maintenance Strategies

We had a small installed base of Linux systems running on Intel systems. With the increase in numbers we have developed a more complete installation strategy for Linux on Intel architecture workstations. Like WinNT this involves an installation recipe that can be easily followed. We had developed a system configuration strategy for other UNIX variants and we modified that to handle the addition of Linux. Unlike NT where most system settings are in the registry, these modifications are spread across many files. However, these files are text based, easily modified with standard UNIX utilities and we have years of expertise in handling this. UNIX is inherently multi-user so the system needs no special software for handling setup upon login for individuals. Most software is located on the AFS fileserver so updates appear via the network without a specific distribution mechanism.

We are finding that recent Linux packages like that coming from RedHat are including more software pre-installed which conflicts with some of our fileserver versions. This will require us to reevaluate where this software should reside and how it will be kept current.

Facility Integration

File Service

We are using the MIT AFS client for Linux . This is public domain and unsupported by Transarc. Transarc has recenlty announced support of an AFS client for Linux 2.2. This client is still in beta test and we are currently using Linux 2.0. We plan on shifting to this in the 4th quarter of 1999. See the section of File Service under Windows NT for a description of AFS.

One problem that we've experienced is that the AFS cache system isn't always signally a need to refresh from system to system. For example if a file is changed from machine A and machine B tries to read the file sometime later, machine B will get the old copy of the file. Running 'fs flush' on the file fixes the problem but this shouldn't be necessary. It is especially annoying for updating web pages where the user is editting the files on the local machine and the web server doesn't see the changes.

Authentication

Printing

Because Linux uses BSD lpr print services and configuration it integrates quickly and seemlessly into our current environment.

Computation

Six of the Intel Grant workstations are available as compute servers using the Distributed Queuing System from Florida State University which we also use for our IBM/AIX and SGI/IRIX systems . These compute servers are available to our entire 500 user faculty and graduate student research community. This allows the user to send long jobs to several machines that will run unattended and may take from hours to weeks for completion.

A 4 system Extreme Linux Beowulf cluster for batch processing has been installed. We were unable to get our benchmark to work using the LAM MPI software. We then proceeded to try MPICH. This was successful. The HINT benchmark tests show that the cluster processes about three times as many net QUIPS (QUality Improvement Per Second) as each individual machine in the cluster. This is especially useful for large processes such as simulations. A Beowulf cluster is not only fast but it is also expandable. If more processing power is needed, more nodes can be added at any time. Our cluster produced the following results running HINT with the MPICH implementation of MPI. In general the MPI code causes the benchmark to start more slowly as seen by the differences in the single processor graphs. The benchmark code for MPI and multiple nodes sets parameters to start further along the "time scale" which explains the difference for the initial points for the cluster. The curves drop off as memory and paging occur.

No actual production work has been done on this cluster yet.

Software

We have also acquired, built, and installed a large selection of data analysis and graphics software for the Linux/Intel platform. In addition to our standard UNIX supported software we have purchased IDL, a visualization package and have licenses for the Portland Group FORTRAN compiler which is essential for the migration of some software.

Utilization

Schwarz - Numerical computation and graphics using Matlab

Sachse - Preparation of papers and presentations for scientific publication using TeX

Johnson - Investigations into solid hydrogen using VASP (Vienna Ab-Initio Simulation Package). It does first-principles quantum mechanics calculations for solids and liquids. The ab-initio calculation is a very standard computational tool in physics these days. Much of this work has been done using AIX/IBM-RS6000 and IRIX/SGI-PowerOnyx.

Pomeroy / IRG3 IRG3 uses scanning and scattering probes to characterize thin films, producing real space and momentum space (respectively) images and lineshapes. Extraction of quantitative information from these images often requires intense image analysis, including lineshape fitting of data Careful tuning and iteration of our fitting routines (a streamlined Mathematica notebook) allows us to quantitatively corroborate data from real and momentum space (STM and CHESS), as well as have the power and versatility to make on-the-fly corrections to experimental procedure based on the information extracted. We propose to use this computer to enhance our ability and efficiency to perform this analysis.

Clancy / Engstrom / IRG2 Will be performing experimental and simulation work in support of work related to IRG3 and IRG2. I intend to use the machine to run Kinetic Monte Carlo simulations of thin film growth using a code that was written in house (by Steve Levine- an IRG-supported student who graduated in August). Steve performed the first simulation of polycrystalline thin film growth by VPE in the literature, which allowed us to assess the merit of the existing theories (none were any good so we are developing our own) and showed that we can predict the RMS roughness of a semiconductor surface with close to quantitative accuracy as compared to the experimental AFM results. So this is a vital project with considerable visibility.

Facility User Room - 3 Linux machines are available for users in the computing facilities general user rooms.

Networking

The 100 MB switched network now includes our general user rooms in both Clark and Thurston Halls. Our current bottleneck is in the infrastructure wiring to the offices, labs and the building. Cornell Information Technologies is planning to offer 100 MB connections in the Summer of 1999.

100 MB switched Ethernet has been installed in our central computer rooms at both the Clark Hall and the Thurston Hall sites. This is helping to relieve the network congestion that having many fast Intel computers can create, and to improve throughput. Further testing needs to be done.

Transfer	Speed	KB/sec	Pkt/sec	Notes
	(MB)		(sustained)
NT-NT R/W	100	1,080	800	(SMB)
NT-AFS Read	10/100	771	740
NT-AFS Write	100/10	771	740
NT-AFS Read	10	415	740
AIX-AFS Read	10	402	400
AIX-AFS Write	10	519	740
AIX-AIX R/W	10	771	740	(rcp)

Performance

We have only begun to compare performance between Intel and other systems running as well as NT versus Linux. We expect to do this on standard applications that are used under normal operating conditions, i.e. network and normal services running. We are currently compiling this information and will be updating the results shortly.

Test Results
Test	CPU	OS	CPU Time	Real Time	Date	Notes
latex1	Intel PII/266	NT 4.0		3.2	7/9/98
		Linux 2.0.33	0.89	2.2	7/10/98
	IBM Pwr2/59	AIX 4.1	5.62	7.9
tex1	Intel PII/266	NT 4.0		1.20	7/9/98
		Linux 2.0.33	0.23	0.48	7/10/98
	IBM Pwr2/59	AIX 4.1	0.68	0.74
EB	Intel PII/300	Linux 2.033	533		2/5/99
	Intel Xeon/400	Linux 2.0.33	370		"
	IBM Power2SC/120	AIX 4.2	633		"
Neaton	Intel PII/266	Linux 2.0.33	981	16:34	2/10/99
	Intel PII Xeon/400x512	"	629	10:34	"
	SGI Power Onyx R10K	IRIX 6.2	694	11:49	2/12/99
	Intel Dual PII/333	Linux 2.0.33	813	13:48
Base SPECfp95	CPU	---	Unixware 2.0	Windows NT 4.0
	Intel PII/450		11.3	11.8
	Intel PII/400 Xeon			12.4
	Intel PII/400		10.8	11.4
	Intel PII/350		9.64	9.91
	Intel PII/266		6.97	7.39
HINT	CPU	UNIX Version	UNIX Result	Windows NT 4.0
	IBM 7012-530	AIX	1.563		Apr 8 1998
	IBM 7012-370	AIX	3.72
	IBM 7012-380	AIX	4.212		Jul 12 1997
	IBM 7013-550	AIX	2.865
	IBM 7013-590	AIX	5.582		Jul 12 1997
	IBM 7025-f40	AIX	6.551		Jul 25 1997
	Intel P5-90	Linux 2.0.29	1.951		Jul 21 1997
	Intel P5-120	Linux 2.0.29	2.725		Jul 11 1997
	Intel P5-133	Linux 2.0.28	3.174		Jul 11 1997
	Intel P5-166	Linux 2.0.28	3.717
	Intel P5-200mmx	Linux 2.0.33	4.779		Apr 8 1998
	Intel P6-200			4.297		errors!
	Intel P6II-233			5.378
	Intel P6-266	Linux 2.0.33	9.568		Apr 8 1998
	Intel P6-333	Linux 2.0.33	11.52		May 20 1998
	Intel P6II-450	Linux 2.0.36	18.58		Mar 22 1999	PGC
			15.23			PGC MPI code single processor
			46.5			PGC MPI cluster 4 processors
	Intel P6 Xeon-400	Linux 2.0.36	14.53		Dec 17 1998
	SGI R8k-90	IRIX 6.2	6.22
			4.485			Shared Memory 1 processor
			28.37			Shared Memory 8 processors
	SGI R10k-195	IRIX 6.2	15.88		Apr 8 1998
			11.63			Shared memory 1 processor
			66.64			Shared memory 8 processors
	SGI IP30-175	IRIX 6.4	13.52		Dec 3 1997

EB: Eberhard Bodenshatz simulation program
Neaton: Jeff Neaton, graduate student working with Neil Ashcroft, The times are from one iteration of a density-functional calculation of the electronic structure of solid lithium at high pressures using the VASP code (Vienna Ab-initio Simulations Package).
Base Spec95fp: Specmark 95 Base Floating Point values from SPECfp95
HINT: Hint Home Page

Next Quarter Plans

Development and implement a file system/registry lock-down strategy for NT.

Using MS SMS examine methods needed to upgrade or add applications to configured NT systems.

Install and test MS Terminal Server

Continue to test and benchmark UNIX workstation applications under Windows and Linux on Intel platform. (TeX, IDL,Hint,Spec)

Contacts

Faculty

	Eberhard Bodenschatz - 618 Clark Hall, 255-0794
	Alan Zehnder - 317 Kimball Hall, 255-9181

CCMR Research staff

	Mike Heisler - 302 Thurston Hall, 255-7344
	Barry Robinson - E20 Clark Hall, 255-7223
	Daniel Blakeley - E20 Clark Hall, 255-6064
	Tom Baldwin - 302 Thurston Hall, 255-0637

Graduate Students

CCMR Intel Grant Status - Local CCMR page will be most up-to-date information on this project

Equipment Utilization

Name	Type	Received	Site	OS	Utilization
	266Mhz Pentium II 128MB	11/97	Clark	WinNT	Relocating to CCMR Interdisciplinary Research Group
	266Mhz Pentium II 128MB	11/97	Clark	Linux	Clancy/Engstrom/IRG2-3
catalina	266Mhz Pentium II 128MB	11/97	Thurston	WinNT	Ast/IRG2
kaa	266Mhz Pentium II 128MB	11/97	Thurston	Linux	Pomeroy/IRG3
watson	300Mhz Pentium II 128MB	1/98	Clark	WinNT	SMS testing
bob	300Mhz Pentium II 128MB	1/98	Clark	WinNT	CCMR_GINA user room, replicated
maine	300Mhz Pentium II 128MB	1/98	Thurston	WinNT	Revesz
hood	300Mhz Pentium II 128MB	1/98	Thurston	WinNT	Lam
dinosaur	510T 24-port 10/100MHz switching hub	2/98	Clark	N/A	Network Hub
ccmrs01	510T 24-port 10/100MHz switching hub	4/98	Clark	N/A	Network Hub
bkts00	510T 24-port 10/100MHz switching hub	2/98	Thurston	N/A	Network Hub
bkts01	510T 24-port 10/100MHz switching hub	4/98	Thurston	N/A	Network Hub
utah	333Mhz Dual Pentium II 256MB w/tape	4/98	Thurston	Linux	Computer server
ohio	333Mhz Dual Pentium II 256MB	4/98	Thurston	Linux	Compute server Johnson
bumble	333Mhz Dual Pentium II 256MB	4/98	Clark	Linux	CCMR user room, Schwarz
bee	333Mhz Dual Pentium II 256MB w/tape	4/98	Clark	Linux	Computer server
seven	400Mhz Pentium II 128MB	8/98	Clark	WinNT	CCMR_GINA development
sugar	400Mhz Pentium II 128MB	8/98	Clark	Win98	Reseach Experience for Teachers Program user room
texas	400Mhz Pentium II 128MB	8/98	Thurston	WinNT	CCMR_GINA user room, replicated
iowa	400Mhz Pentium II 128MB	8/98	Thurston	WinNT	CCMR_GINA user room, replicated
wanda	400Mhz Pentium Xeon 256MB	12/98	Clark	Linux	CCMR user room
malcolm	400Mhz Pentium Xeon 256MB	12/98	Clark	Linux	CCMR staff test
neelix	400Mhz Pentium Xeon 256MB	12/98	Thurston	Linux	CCMR user room
kurn	400Mhz Pentium Xeon 256MB	12/98	Thurston	Linux	Compute server Sachse
Beowulf	450Mhz Pentium II 128MB (4 systems)	12/98	Clark	Linux	Extreme Linux MP cluster

Last modified on: 10/12/99

	Development and implement a file system/registry lock-down strategy for NT.
	Using MS SMS examine methods needed to upgrade or add applications to configured NT systems.
	Install and test MS Terminal Server
	Continue to test and benchmark UNIX workstation applications under Windows and Linux on Intel platform. (TeX, IDL,Hint,Spec)