Home Contents Search

Previous Status
Up ] Current Status ] [ Previous Status ]

 

LASSP: Condensed-Matter Physics Status Report - 4th Quarter 

Accomplishments
Next Quarter Plans
Contacts
Equipment Utilization
Feedback and Problems

Accomplishments

The Linux machines have taken over as the theory group's computational workhorse. Large-scale computations on exotic boron crystals, electronic structure calculations of metal-glass interfaces via electronic density functional theory, charge exchange near surfaces, and Barkhausen noise in magnetic hysteresis are currently underway within the group. The NT machines are becoming increasingly popular due to the Visual C++ programming environment, and many of the younger sophisticated graduate students use them extensively.

Next Quarter Plans

In the next quarter, we expect to re-install the operating systems of the NT machines, making them more stable and uniform.

Contacts

James P. Sethna, Professor, Department of Physics
Ralph B. Robinson, Programmer Analyst, Physics Department

Equipment Utilization

The two new Xeon machines are particularly popular, due both to their speed and (more importantly) due to the relatively large memories. We have not yet replaced our Silicon Graphics Onyx supercomputer with the Xeon machine, largely because of the latter's larger memory.

The configuration of the network and the other machines remain unchanged.

Feedback

We were discouraged to find that the planned upgrades of our existing machines to current processors was considered infeasible. We expect to shift the current machines to faculty offices when the new machines arrive in Q2 or Q3, and newer processors would generate more excitement...


Last modified on: 07/07/98

LASSP: Condensed-Matter Physics Status Report - 2nd Quarter

Accomplishments
Next Quarter Plans
Contacts
Equipment Utilization
Feedback and Problems

Accomplishments

In conjunction with the Cornell Center for Materials Research branch of the project, our machines are set up and running, both NT and Linux, using the AFS file system.

We have had performance difficulties with Developers Studio (our main tool for research computation) under NT, and many of the sophisticated NT programmers have remained using the physics service courses facility for their research work (see below under Equipment Utilization and Feedback). The NT machines have seen some substantial use in running working programs and in more routine applications. We are working hard on improving the performance.

The Linux machines are seeing heavy usage in studies of electromigration and void dynamics, in electronic structure calculations, in studying quantum eigenstates of chaotic systems, and in many other simulations.

Next Quarter Plans

Within the next quarter, we expect to see significant scientific results growing out of the use of this new equipment, which will be reported here. We also anticipate a prompt resolution of the performance problems under NT AFS.

Contacts

James P. Sethna, Professor, Department of Physics
Ralph B. Robinson, Programmer Analyst, Physics Department

Equipment Utilization

Five of the machines delivered at the beginning of this quarter are running dual-processor Linux in the main research facility; five are running NT over AFS (Andrew File System) there. One, running NT Server, is in use by a new faculty member in LASSP. All are networked via fast Ethernet via the 24-port Intel 510T switch, which also is connected to our AFS server.

AFS provides significant advantages over NFS or other shared file systems in administration, reliability, security, and network bandwidth. Mike Heisler, the head of our CCMR computing facility and the person in charge of the CCMR portion of the Intel project, provides the following list of crucial features:

Administration
Ease and transparency of moving volumes: Can move software or home files between disks or servers on-the-fly without disruptions and without users even knowing it.  
Replication: Performance improvement (least busy server responds), subnets across campus have "local" copy. 
Client specific access features: Binaries are client OS specific without user or system intervention i.e. typing cc on Linux or SGI gets you the appropriate compiler. 
Reliability
Backup facility:
"Yesterday" - nightly copy of your files for undelete
Replication - Redundancy of server info and files.
Security
Kerberos authentication - secure and robust
ACL (Access Control Lists) - Well beyond normal unix permissions. Similar to NT security
controls
Network Bandwidth
Local Cache keeps frequently used files on the client. (this is a 2-edged sword!)


We have extensive experience with it under various flavors of Unix. We are having significant difficulty configuring our NT machines under AFS. Microsoft software, particularly the help files, expects to write temporary files into system file directories. This seems bizarre to us. Naturally, AFS by default refuses to allow ordinary users to write in system directories. We don't understand how this is compatible with NT's security either.

There is a serious performance hit in running Developer's Studio on a remote server running AFS, that has discouraged users from shifting from the Intel teaching facility which they used last semester. We believe this is in part due to the large number of small system files that Windows incorporates as part of its user interface: a small lag time for each file translates into an enormous slowdown. We suspect that each file may be stored both in AFS cache and in NT cache; this is a flaw in the implementation of AFS, which Transarc should be encouraged to fix (see below). Our AFS server performance may also be part of the problem. Unfortunately, Transarc does not support AFS server on either Linux or NT: we suggest that this support might be of some importance to Intel (see feedback below). Our current plan is to shift to installing much of the software locally on each machine, which is a significant investment in disk space and a support headache.

Feedback

Currently we need to use a non-Intel machine as our file server, as Transarc does not support AFS servers under either Linux or NT. We strongly suggest supporting and encouraging Transarc to port their AFS server to Linux, and to improve their NT implementation of AFS client. AFS is the network file system of choice in the scientific and computational community, and runs at the various supercomputer centers: implemented properly, it could be a significant value for Intel's customers.

We still await a good method to run remote jobs under Windows NT.

Last modified on: 07/07/98

 

LASSP:  Condensed-Matter Physics Status Report - 1st Quarter

Accomplishments
Next Quarter Plans
Contacts
Equipment Utilization
Feedback and Problems

Accomplishments

Since the bulk of our equipment is yet to arrive, our focus has been on getting the software and hardware configurations set up properly so that the new machines will become immediately useful. In conjunction with the Cornell Center for Materials Research branch of the project, our machines now run on the same AFS server which runs our AIX and Linux systems.

Scientifically, the two machines in publicly available areas have seen heavy usage in electronic structure calculations, in studying quantum eigenstates of chaotic systems, and in many other simulations: they have whetted our appetite for the substantial influx to come! The one machine running Windows NT has been particularly popular with faculty who need access to Microsoft Office applications to interact with grant agencies.

Next Quarter Plans

In about one week, we expect a substantial influx of new machines. The existing machines will be removed from the computer rooms on the fifth and sixth floors, keeping only two RS6000's with large memories. The new machines will be installed, simultaneously shifting to fast switched Ethernet.

Over the past semester, we have been running a graduate physics course on Simulations under NT, running in the physics service courses facility. This course, which has been extremely successful and popular, has brought a large corps of talented graduate students and post-docs up to the leading edge of current simulations research, while introducing them to C++ and MFC GUI interfaces. We expect great things to happen in LASSP over the summer when these students come in to take over the computer room.

Contacts

James P. Sethna, Professor, Department of Physics
Ralph B. Robinson, Programmer Analyst, Physics Department

Equipment Utilization

Three dual-processor Pentium Pro servers have arrived so far. One is being used as a Linux compute server; one is running Windows NT Server and is running NTAFS: both are heavily in demand both for large-scale computations and for fast response. The third machine is being used by Barry Robinson for system development. All are networked with fast Ethernet via the 24-port Intel 510T switch, which also is connected to our AFS server. We currently are not using the eight-port Intel switch (see below).

Feedback

We still await a good method to run remote jobs under Windows NT. This will become crucial in the next quarter, when the bulk of the machines arrive. It may well be the primary bottleneck to shifting away from Linux.
Linux runs wonderfully well on this platform: Unix workstation vendors beware!
These three early machines came before the shift to Dell as the primary supplier; they had serious idiosyncrasies. When we installed third-party PCI video cards, the machines refused to boot. Installing a PCI ethernet card refused to work on the primary PCI bus, but would work on the second PCI bus. We aren't focusing on our experiences with these machines, because we expect that feedback about the mainstream supplier will be more useful both for Intel and for our compatriots at other institutions.
The fast Intel 510T Ethernet switches have worked extremely well: we had only relatively minor problems with the installation.
The earlier eight-port switch had an interesting feature: if set up for remote administration via Telnet, outsiders can log on with a backdoor password which could not be disabled. We initially disabled Telnet, but are currently running with the 24-port 510T switch which does not suffer from this security hole.
We learned a lot from setting up the Physics Service Courses facility: see that report for our major and minor experiences with cloning our workstation configurations with Drive Image Professional, and the search for XY plotting routines.
 

 

Back Home Up

Last modified on: 10/08/99