National Laboratory for Applied Network Research

Summary status report
1 May 1995 to 31 December 1995

(also first quarterly report)

Cooperative Agreement No. NCR-9415666 with the National Science Foundation


Participating organizations:

Cornell Theory Center (CTC)
National Center for Atmospheric Research (NCAR)
National Center for Supercomputing Applications (NCSA)
Pittsburgh Supercomputing Center (PSC)
San Diego Supercomputer Center (SDSC)


Over the last quarter, NLANR sites have worked on developing the vBNS as a national research resource, including technical/engineering support, outreach, coordination and management, user education, and contributing to supporting the overall research agenda.

The technical and engineering support was related to establishing and extending vBNS connectivity, performance testing, and troubleshooting. MCI's 2nd quarter 1995 (April to June) report on the vBNS provides generally a more detailed description of the first quarter vBNS deployment.

In the outreach, coordination and management role, NLANR focused on developing and refining procedures to facilitate access to vBNS resources by the research community. NLANR has also established several working groups to focus on dual-natured research areas, in that they involve research themselves but also will provide infrastructural support for other research on the vBNS. These areas include multicast, multimedia conferencing tools, collaborative environments, advanced routing functionality, security, middleware, TCP performance analysis, statistics monitoring. We also highlight current vBNS applications based at NLANR sites.


vBNS connectivity

All NLANR sites participated in establishing and testing vBNS connectivity, both on the production vBNS and MCI's testnet between Reston, PSC, and Dallas. At each site, NLANR personnel coordinated and managed the installation of the MCI/ICG OC3 in late March and early April 1995, including installing, terminating, and testing fiber from its demarcation point to the equipment location, and coordinating the installation testing of the vBNS hardware in late February 1995. They also assisted in troubleshooting the MCI/ICG OC3 and vBNS equipment as needed throughout the quarter.

As part of the initial testing of vBNS connectivity, sites set up IP network routes to exchange traffic between the shared FDDI rings at their site and the other vBNS sites. Initial testing showed all network components in the path (Cisco routers, Lightstream 20/20 ATM switches and GDC ATM switches) performing as expected and operating reliably.

The NetStar GigaRouters continue to exhibit unstable behavior, with the HIPPI interfaces frequently locking up. We were able, though, to establish a HIPPI to HIPPI path over the vBNS, and exchanged TCP/IP traffic between HIPPI attached machines at SDSC and PSC. We provided test results and feedback to MCI and to NetStar to help them fix problems. Steve Cunningham (PSC) assisted MCI with site planning for the installation of the vBNS and the testnet, and assisted vBNS users in the deployment of vBNS attached hardware at the Westinghouse Energy Center.

vBNS performance testing

The sites worked extensively on early debugging and performance evaluation of the vBNS network. PSC, in collaboration with SDSC and NCAR, tested the links between them to provide early performance results. SDSC and CTC demonstrated with FDDI to FDDI tests that the vBNS could sustain full bandwidth communication (100 Mbps) across the backbone network. We were able to sustain approximately 91 Mbps using TCP between 2 Digital Alpha workstations at SDSC and Cornell. The packet delay across this path was approximately 60 msec, for a delay-bandwidth product of 750 kilobytes.

HIPPI-to-HIPPI tests indicated that performance was severely limited by the NetStar GigaRouters. Using large packets (approximately 8 kbyte) allowed 100 Mbps throughput between two Cray computers (between SDSC and PSC), but the NetStar HIPPI interface does not perform well for large packets-per-second traffic loads, so smaller packets limit end-to-end performance of applications that use them. We continue to test both configurations and performance in the presence of resource contention.

NCSA ran a test in June between supercomputers at NCSA and SDSC, measuring performance as below.
Performance between NCSA Power Challenge SGIs
and SDSC C90 (13 June 1995)
WRITE Results:
Buffer size Avg Sys TimeAvg Usr TimeAvg Mb/sPeak Mb/s
1K 48.76% 1.84% 25.4 40.9
2K 26.96% 1.16% 31.4 51.1
4K 19.84% 0.82% 42.0 51.3
7K 13.36% 0.44% 37.7 51.6
8K 8.20% 0.42% 40.1 51.6
15K 5.96% 0.28% 32.8 48.0
16K 6.68% 0.30% 36.5 51.6
31K 4.48% 0.12% 26.3 51.6
32K 8.80% 0.20% 51.0 51.7
48K 8.34% 0.10% 50.8 51.8
57K 6.84% 0.10% 41.3 51.8
60K 7.44% 0.10% 46.0 51.8
63K 6.80% 0.10% 42.8 47.5
64K 5.16% 0.08% 32.7 51.2
128K 8.26% 0.10% 51.7 51.8
READ Results:
Buffer sizeAvg Sys TimeAvg Usr TimeAvg Mb/sPeak Mb/s
1K 13.04% 1.22% 16.2 30.2
2K 17.54% 1.14% 26.6 50.2
4K 21.62% 1.12% 39.1 49.7
7K 17.30% 0.82% 33.3 50.5
8K 18.54% 0.76% 36.6 50.3
15K 19.60% 0.78% 34.7 47.3
16K 17.62% 0.66% 30.3 49.4
31K 16.82% 0.60% 29.5 47.8
32K 27.28% 0.92% 48.3 49.7
48K 28.32% 0.88% 47.7 49.4
57K 23.66% 0.72% 41.5 50.1
60K 25.24% 0.76% 43.7 49.6
63K 24.36% 0.78% 41.9 49.9
64K 23.64% 0.70% 38.9 49.3
128K 28.58% 0.80% 49.9 50.3

nettest log: Tue Jun 13 11:30:27 CDT 1995
nettest command: nettest -l -b 768k
total data: 62914560
repetitions: 5
socket buffer size: 768k
data file: sif.to.c90-HIPPI.sdsc.edu.130695.113027
dedicated: no
tcp mss: 8948
average load on local host: 0.72
average load on remote host: -1.00

Cornell has connected and configured IP over HIPPI to the SP2, and tested performance with Matt Mathis to the Cray at PSC. NCSA and Cornell have also used HIPPI-attached Power Challenges at each site for the Seidel gravitational field calculation application, described later, and are beginning testing for that application between Cornell's SP2 and the SGI Power Challenge Array at NCSA.

Cornell connected their campus FORE-based network to the vBNS Lightstream with an upgrade of the FORE ASX-200 switch software to version 3.4. They also connected their IPOP machine via 155 Mbps ATM interface to this FORE switch and successfully tested IP connectivity across the vBNS, as well as testing PVC-based IP connectivity from NYNET through this FORE switch to the vBNS.

Direct ATM attachment to vBNS

As part of the infrastructure required for Supercomputing'95 conference, several of the sites (SDSC, NCSA, Cornell, PSC) purchased Sun Sparc's (Sparc 20 and a Sparc 5 at PSC) for direct ATM attachment to the vBNS (or directly attached other ATM-capable hosts). The workstation connected to the Lightstream ATM switch, sometimes via an intermediate Fore ATM switch (SDSC, NCSA, Cornell). Based on MCI's Virtual ATM LAN addressing scheme, sites also configured the ATM PVCs and IP addresses to enable direct ATM connections to other vBNS devices. They were successful in setting up ATM connections to vBNS Cisco routers at the other sites and exchanging IP traffic with these routers.

Facilitating research

The NSF is still evolving the processes and procedures by which the vBNS will be made available for use within the U.S. R&E community. However the NSF does not wish to leave the vBNS dormant either while these issues are being resolved or later if less than complete utilization is authorized for activities that have been approved. The NSF has delegated to the vBNS Technical Coordinating Committee (vTCC) the responsibility to coordinate and authorize AUP compliant uses of the vBNS among and between the vBNS sites (currently the five SCCs) that:
  1. do not require additional resources from NSF
  2. do not require major expenditures, effort or reconfigurations of the network by either NLANR or MCI personnel and
  3. do not involve transit of traffic from sites or networks not currently connected to the vBNS

NLANR maintains a public record of all such uses (In the event of a disagreement between members of the vTCC as to whether or not a proposed use falls within the parameters of the delegation, the issue will be escalated to the NSF Program Official). NSF expects this delegation will facilitate vBNS use for performance testing of both the network itself and for applications and technologies not specifically reviewed for merit. vBNS utilization authorized under this delegation may decline as NSF-approved connections and applications on the vBNS increase and will not impede the use of the vBNS by such NSF authorized connections and applications.

K. Claffy of SDSC serves as NLANR Research Coordinator to facilitate collaboration and outreach. Claffy established a publicly web-accessible database of applications that are either active on the vBNS or under review by the vBNS Technical Coordination Committee based on information in their allocation request form.

Jamshid Mahdavi of PSC serves as co-chair (with Rick Wilder of MCI) of the vBNS Technical Coordination Committee, and provides web-accessible versions of the pre-planning meeting notes for the bi-monthly vTCC tele-meetings. The use of the web for maintaining the most recent pre-meeting notes has enabled the vTCC to focus more time on technical discussions and planning, and led to significantly higher productivity for the group. Mahdavi also worked with the NSF and other members of the vTCC to develop the current vBNS allocation request form for use by anyone desiring access to vBNS resources.

The Principal Investigators at each site (Douglas S. Carlson, Marla Meehl, Randal L. Butler, Matthew Mathis, and Bilal A. Chinoy at CTC, NCAR, NCSA, PSC, and SDSC, respectively), and a second representative from each site (Paul Hyder, Phil Pishioneri, Paul Zawada, K. Claffy, Jamshid Mahdavi), have acted both as members of the vTCC and as engineering support for new applications as they come online.

These site contacts provide support for establishing application-specific routing and configuring both local site resources and the vBNS backbone as required to support new applications. In particular, technical staff have spent considerable time helping vBNS `friendly users' with vBNS/I-WAY projects for the Supercomputing'95 conference, including tasks such as I-POP scheduling, route setup, and testing over the vBNS.


Working groups for supporting NLANR research agenda

The NLANR network research agenda is focusing on mechanisms that will allow greater control over network integrity and performance. Key areas in need of progress include: routing, accounting, multiple qualities of service, multicast, collaborative environments, and security. In each case we recognize a two-pronged problem: changing the technology, from which the NSF is `three hops' away (NSF, MCI, vendor/ietf, equipment/algorithm), and changing the community mindset, from which the NSF is `two hops' away (NSF, policy on vBNS, which is role model to rest of community).

A useful example is the efficient support for multiple qualities of service, which vendors are not commonly building into their equipment. Understandably so, since their customers, ISPs, are not asking for it (or certainly not very loudly), perhaps partly due to an established mindset that ``the Internet does not support it.'' The vBNS provides an opportunity to push the vBNS-supplying vendors, which others will no doubt follow, into offering this kind of support, since it will prevent larger scale development efforts for service guarantees from being in the critical path of moderate scale improvements to the infrastructure. Indeed, we view the commodity Internet infrastructure as rapidly approaching a dangerous time period, where congestion will overwhelm many service providers before equipment vendors have full-scale quality of service and admission control mechanisms in place.

We believe the vBNS can play a critical role, not only with its lead in technology, but also in its unique position within the Internet community as a neutral and multi-site testing ground for proposed changes, whether in the form of completely new algorithms or modifications to existing algorithms. For example, the NSF can help by encouraging the vBNS to demonstrate a prototype system using existing functionality, e.g,. the IP precedence field, to implement a small range of service qualities on the vBNS.

Multicast support (leads: kc@nlanr.net and kthomp@mci.net )

Each NLANR site has set up a multicast router, in most cases in addition to the existing multicast router at the site, in order to support high-bandwidth multicast experiments on the vBNS without interfering with operational Mbone traffic. Most sites were initially using SGI Indigos purchased for a previous multi-center project. MCI has since decided to deploy the most recent version of mrouted (3.6) on the Digital alphas in each vBNS node. (PSC did not set up the Indigo but preferred to wait for this Alpha support.) Some sites, at least SDSC and Cornell, will continue to use the Indigo additionally in order to develop and maintain CU-SeeMe reflectors to nv/vat client gateways. David Mitchell at NCSA provided operational Mbone support at NCSA via the SGI probe system running mroutedv 3.4 for several months, then converted to a Sun OS Sparc5 running mrouted v3.6. Once a Solaris version is available and tested, NCSA will convert again, in parallel with MCI's Mbone support on the Alphas. This redundancy will be useful at times, including for multi-platform interoperability. NSF has approved the vBNS for preemptible use for tunnels to connect to the standard Mbone, and NLANR will begin coordinating with the Mbone community in deploying those tunnels.

Network based multimedia collaboration tools (lead: jeffwink@sdsc.edu)

Jeff Winkler (SDSC) has taken a lead in the coordination of a cross-platform, cross-site interoperability effort for multimedia collaboration tools, including coordinating an infrastructure that can provide the functionality of the circuit-based videoteleconferencing room facilities that the sites were using until 1 December 1995 when they were dismantled. NLANR is coordinating resources to replace the facilities to meet intercenter virtual meeting needs.

EMMI multimedia interface (NCSA/AT&T)

Mike Haberman and John Lockwood (NCSA) and Jeff Winkler (SDSC) have tested AT&T EMMI video equipment between NCSA and SDSC. The AT&T EMMI, an ATM-based video/audio codec, provides a multimedia interface for a local host to an ATM network. The video stream uses JPEG compression to yield five levels of quantization to match the available bandwidth using ATM Adaptation Layer 5, for bandwidth usage between 8 and 80 Mbps. The audio stream uses AAL 1 with a bandwidth of 1.85 Mbps. The data stream, which supports a White Board, runs over TCP/IP and on top of AAL 5, but we have not evaluated this feature yet. We currently have the EMMI's set up at NCSA and SDSC through the vBNS. The EMMI interfaces are available to us on a temporary basis, to provide feedback on their VTC usability.

Perhaps the most difficult aspect of getting the EMMI's configured has been getting the drivers to install properly. AT&T seems to think the problem lies in our using the SUN IPC machines rather than a Sparc station. We currently are using a SUN IPC.

So far we have tested the EMMI in the following configurations: a local loop back, local back-to-back, loop back through our ASX-200 switch, and across the vBNS to SDSC using 3 VCs. In each case, we have sent an audio and video feed. We have not had sufficient equipment to test the video conferencing software yet.

Despite not having any quantitative measurements, we can make a qualitative assessment of the EMMI's performance. With both EMMI's set at a Q factor of 25 (approximately 14 Mbps), the images received on both ends are smooth and free of jitter. Audio is also smooth and free of distortion. However, there seem to be some difficulties with the EMMI asking for more bandwidth than we can deliver, especially as we increase the Q-factor.

Currently, there is a lot of hands-on maintenance and babysitting in order to keep the EMMI's functioning properly. System log files fill quickly if there is a communication failure. The user interface frequently wedges the machine, and the video conferencing software is not very tolerant. It seems that if the EMMI is on a Sparc station, overall management is much easier. We have worked closely with the technical staff at AT&T to identify software bugs and expect a new release that resolves these problems soon.

Currently we are running tests between NCSA and SDSC. We have been offered a third unit and are talking to PSC as a possible site to host it. This would allow us a more realistic test of an actual conference system. In addition we plan to bring up the whiteboard capabilities in early December. We expect the project to be complete by the end of 1995. Butler (NCSA) also expects to submit a proposal soon in collaboration with an Industrial Partner that is interested in using MPEG transfers for live video clips on the air. They expect these live feeds to come from sites all over the country and would like to use the vBNS as a testbed to research real-time transfer of MPEG streams over a wide area ATM network.

Collaborative environments (lead: bbj1@cornell.edu )

Bruce Johnson (Cornell) has recently begun to lead efforts to integrate existing technology on text-based collaboration environments and multimedia tools such as CU-SeeMe. Jeff Winkler (SDSC) and Scott Brim (Cornell) expect to investigate integrating an active MOO environment with CU-SeeMe, and explore the differences between such an approach and other existing architectures (e.g., Juptier at Xerox Parc).

Network routing (leads: mahdavi@psc.edu and jjamison@mci.net)

As detailed in MCI's quarterly report, dynamic routing occurs between all vBNS sites, using OSPF among vBNS routers. The four NSF-supported NAPs are connected to the vBNS via DS-3 circuits, and MCI will utilize BGP4 peers with other Internet Service Providers as soon as non-SCC users for the vBNS are identified and approved by the NSF. MCI has established BGP4 peering between the vBNS and MCI's commercial IP network at three of the NAPs to test the NAP connections and import routes needed for network management access to the vBNS equipment.

Routing at the ATM level is static, facilitating reliability and predictability of routing. ATM switch configuration includes an over-full mesh of permanent virtual circuits among vBNS routers; each router knows about virtual circuits used for optimal `one-hop' paths to peer vBNS routers. In addition, where appropriate, they are configured to use other `secondary' virtual circuits to maintain connectivity in the event of circuit or switch failures in the ATM backbone. In some cases fall-back routes use an alternative VC through the backbone and maintain single-hop connectivity; in other cases the secondary route uses an additional router hop.

For IP level routing configuration, vBNS users can send route specifications directly to MCI engineering or may use the automated MCI vBNS route registry.

Network policy routing (lead: kc@nlanr.net)

Mathis (PSC) has provided several initial drafts of views for routing plans for the vBNS architecture that attempt to satisfy the constraints of a high performance, research, non-production network with the needs of the end-sites and the continual change that occurs in a developing network.

In order to deal with vBNS routing requirements of host routes and variable length subnets, Paul Zawada of NCSA configured OSPF in the NCSA LAN, and also configured a route server (Sun sparc) that utilizes gated and supplies routes to the internal NSC routers running OSPF.

Experimental router development (lead: csong@mci.net)

Researchers will also likely want to experiment with new IP functionality on the vBNS. MCI has discussed feasibility of doing this with NetStar, one of the current vBNS router vendors. NetStar has indicated its desire to support router experimentation on the vBNS and MCI drafted a proposed a software specification to allow experimental software to be written by MCI and incorporated into the vBNS NetStar GigaRouters. The proposed interface design supports IP Integrated Services, a set of IP functions currently under discussion in the IETF to provide multiple application-specific service types in an Internet environment. The major components, which could have alternative implementations tested on the vBNS, are a resource reservation protocol, a flow discriminator, and a service scheduler. We can evaluate the performance of these components in the vBNS in the face of both time-critical and non-time-critical applications, as well as test out integrated services mechanisms being developed within the IETF, e.g., link sharing, by Jacobson and Floyd at Lawrence Berkeley Laboratories (LBL). In conjunction with class-based queuing (an implementation of which UCL has developed), link sharing is a model for allowing policy control over bandwidth utilization by a given traffic (set of) source(s). NetStar has agreed in principle to the interface specification, and the vBNS user community now has the opportunity to review it as well.

This router development represents a major effort for MCI and the vBNS user community and is dependent on feedback from the research community and cooperation and resource committed by NetStar. MCI is pursuing this project due to the level of interest from the Internet community in IP Integrated Services, reflected at the NSF Workshop for vBNS Researchers held in June 1995.

Priority queueing

At Supercomputing'95 MCI used the NetStar routers in the vBNS to demonstrate priority queueing effects on the performance of multimedia applications. A video-capable workstation in Reston on the vBNS test network captured video scenes using nv/vat software and sent two traffic streams to a workstation on the SC'95 showfloor. The two traffic streams generated from the Reston workstation traversed the testnet where they were subjected to traffic congestion. Queues with different priority on the GigaRouter's ATM card supported the two traffic streams. Under congestion, the stream in the lower priority rate queue experienced longer queuing delay and eventually higher loss rates than the stream in the high priority rate queue. Two windows on the display workstation showed the same video scenes, one with good performance due to higher priority queueing of its traffic, and the other with poor performance due to the lesser service priority.

Advanced switches

MCI has begun to test OC-12-capable ATM switches, targeting switch throughput capacity, response to congestion, SVC support, Virtual Path switching, and manageability as major goals. MCI and NLANR participants will collaborate in testing frame-oriented drop strategies such as Early Packet Discard and Partial Packet Discard and alternative queue disciplines for their utility in carrying IP traffic. We will use simulation to extrapolate on laboratory and test network measurements to show the improvement that can be obtained with specific product modifications.

Network security (lead: ddrew@mci.net)

Ken Rowe at NCSA and his group are developing some central audit software for the I-WAY. Ken is also the coordinator for I-WAY Security Incident Response team. NLANR hopes to leverage the I-POP/I-WAY experience in scheduling vBNS resources in a secure manner. NLANR participants have also written a Metacenter Security Add-on proposal, which may include vBNS activity.

MCI is supporting ssh software on some vBNS systems, and porting it to some vBNS architectures. Dale Drew (MCI) is the lead on this effort.

Middleware: caching (leads: kc@nlanr.net and wessels@nlanr.net )

Claffy and Braun (SDSC) have proposed to develop and deploy a scalable caching and replication system across strategic network locations within the United States. Digital Equipment Corp. had committed to significant cost sharing and collaboration, as an integral part of this awarded proposal. The goal is to facilitate the evolution of U.S. and international information provisioning with an efficient national architecture for handling highly popular information.

Each NLANR site is participating with a caching server at their site as part of this NLANR caching project. NASA/Ames is also collaborating, supporting a cache at FIX-West.

TCP analysis (leads: mathis@psc.edu and mahdavi@psc.edu )

Mathis and Mahdavi are investigating analysis and simulation of current and proposed modifications to aspects of algorithms in the Reno and Vegas TCP versions for fast retransmit effects, RED/EPD, graceful interaction with suboptimally compatible higher (http) and low (ATM) layer protocols, and other issues under discussion on end2end-interest mailing list. In addition the use of tools specifically for scientific computing such as PVM/MPI, adds another layer of interaction to traffic workloads that requires investigation of the dynamics and effect on performance.

Network monitoring and statistics (lead: kc@nlanr.net)

NLANR interest in statistics collection holds a variety of motivations:

The community has still not converged on a standard set of measurements that would allow comparisons; we are organizing an NSF workshop on network analysis apects to be held in February at SDSC for vendors, researchers, and ISPs to gather and share objectives and constraints.

ATM OC-3 monitoring development

As part of the vBNS project, MCI is planning the development of a cell-level OC-3 monitoring device to trace the capture of IP headers from ATM cells in order to study IP-over-ATM traffic dynamics with microsecond granularity. MCI is developing the device on an inexpensive standard computing platform for wide affordability and programmability to support new data collection and analysis.

We hope to integrate valued functionality of the current NNStat and flow-based statistics into this monitoring device, as well as use captured traffic traces or theoretically developed traffic models to generate traffic for simulations and testing. The monitoring device has an important role for the vBNS of extending the NNStat-style monitoring to include the HIPPI traffic at each SCC (the latter currently unavailable since NNStat relies on the use of promiscuous mode by the Digital Alpha workstation on the vBNS FDDI).

Treno

Mathis and Mahdavi (PSC) have developed a tool to test the performance of network infrastructure under TCP-like loads. Because it is based on TCP windowing algorithms, it has low impact on the network, yet is able to provide accurate benchmarks. Mathis and Mahdavi hope to add an ATM interface to the Treno server and attach it to the ATM infrastructure of the vBNS, allowing the server to appear at the four NAPs via the vBNS infrastructure. The server will be of use to the Internet community as well as researchers of TCP performance.

External demonstrations

Mahdavi (PSC) assisted MCI in demonstrating the vBNS technology at the AFCEA by working with the visualization staff at PSC to provide a live, distributed demonstration of the use of supercomputers in volume rendering applications.

NCAR and NCSA worked with MCI to support vBNS-based demonstrations for Comdex in Atlanta in April 1995 and Telecomm'95 in Geneva in August 1995.

NLANR sites have also been involved in supporting the I-WAY project, which yielded a significant base of test applications to help define how to best use the vBNS once the `friendly access period' ends. Support has included: deployment of an IPOP platform; establishing additional routes; and working with the various applications users who were developing and testing distributed applications at the Supercomputing'95 show.


NLANR research applications

Many of the applications on the vBNS thus far were intended for demonstration at Supercomputing'95. We classify them by scientific category.

Physics

visualizing Einstein's gravitational waves

Ed Seidel (NCSA) in collaboration with researchers at UIUC, is using the vBNS to demonstrate a distributed, heterogeneous, scientific solution of the full 3-dimensional Einstein equations for gravitational fields, while simultaneously exploiting the resources at various vBNS sites and displaying the results as they are computed in the CAVE.

This application continues work demonstrated at both Supercomputing '93 and SIGgraph '94, showing both the usefulness of the CAVE for scientific visualization and the feasibility of using high-speed networks to distribute a complex scientific application across local HIPPI or FDDI networks. For Supercomputing'95 they used the I-WAY ATM network to distribute the calculation across the largest machines available across the vBNS, yielding the largest simulation of Einstein's equations ever attempted, and also to refine the scientific visualization capabilities in the CAVE. The experiment provides an important testbed for development of other large-scale distributed scientific applications, as our numerical algorithms are common to many scientific and engineering disciplines, therefore having direct application to other fields of computational science.

Earth system science

NCAR submitted four friendly user vBNS applications that are operational: and an additional application that is pending approval of attachment of a non-vBNS site (the University of Colorado) to NCAR: In addition, several atmospheric applications were demonstrated at SC'95. Louis Rossi (Northwestern) showed an interactive simulation of contaminant evolution through porous media, which computes the concentrations and pressure fields of the contaminants remotely on at least one massively parallel supercomputer, and displays the results in the CAVE as a user interactively adjusts environmental parameters.

Bob Wilhelmson (NCSA) and Crystal Shaw illustrated the visualization of simulated tornados associated with supercell and non-supercell convection.

Charles Goodrich (U. Maryland) demonstrated real-time and near-real-time space weather forecasting at SC'95 through coupled computer simulations using the combined resources of PSC, CTC, and Goddard Space Flight Center (GSFC). A real-time global MHD simulation of the magnetosphere on the T3D at PSC provided results to continually update an ion kinetic simulation running in tandem on the SP-2 at CTC.

Glen Wheless demonstrated the Chesapeake Bay Virtual Ecosystem (CBVE) to incorporate linked submodels of larval behavior, primary and secondary production, benthic processes, dissolved organic carbon, dissolved inorganic nitrogen and dissolved oxygen, all of which will be spatially and temporally linked via their existing 3D Bay circulation model.

Astronomy

In their demonstration at Supercomputing'95, in connection with the Grand Challenge Cosmology Consortium's project, Michael Norman (NCSA) , Joel Welling (PSC) (NCSA, UIUC, and the Beckman Institute) and colleagues simulated the merger of the Milky Way and Andromeda galaxies using multiple massively parallel supercomputers located at all four NSF centers. The project illustrated the concept of collaborative distributed computing, wherein two simulation codes of different type are coupled into a single multiphysics model. From their conference report:
A simulation of the collision of two galaxies was run on the PSC's Cray T3D, and coordinates for star clusters were transmitted in real time across the vBNS for display. For SC'95 these coordinates were displayed on a Silicon Graphics Onyx at the conference in San Diego, in a virtual reality environment, as part of the GII Testbed project. Technical problems prevented successfully displaying the results on the NII Wall at the show, but we were able to run the program to an Immersadesk in the MetaCenter booth on the exhibition floor. The GC3 group won an High Performance Computing Challenge award for the software system of which this demo was a part.

This application demands high bandwidth. 12 bytes or roughly 100 bits must be transmitted per point; a T3D processor can update roughly 1000 points per second using the algorithms involved. Thus running on 128 processors would have required a bandwidth of 10 Mbits/sec. This was the configuration in which we had planned to transmit to SC'95, but network difficulties at the show floor forced us to use drastically reduced bandwidth of roughly 3 Mbits/sec.

This particular application will continue to be used in the normal course of GC3 operations, to transmit data from the PSC T3D to a display at NCSA. The software we developed for the demo are quite general, and will allow us to transmit a variety of scientific visualization geometries. With an appropriate machine on the receiving end, these transmissions are very likely to be bandwidth-limited. vBNS was used for one such transmission at SC'95, when geometries from a molecular dynamics simulation were sent from the PSC's C90 to the MetaCenter booth. I'm afraid I can't offer bandwidth figures for this demo.

Additionally, Doug Roberts and Dick Crutcher (NCSA, UIUC) demonstrated a three-part radio astronomy synthesis array observing the center of our Galaxy. This application combines large-scale parallel computation, data mining and transfer, and advanced 3-D visualization comparing the computed and mined data.

Biology

Supercomputing'95 also provided Wolfgang F. Kraske an opportunity to demonstrate a distributed biomedical visualization system using the 140 Gflops processing throughput from a HIPPI connected Cray T3D, Intel Paragon and Intel Delta linked via a 622 Mbps asynchronous transfer mode (ATM) OC12 line to a remote (12 miles away) ATM switched workstation cluster. They connect to the IBM SP2 at Cornell and one at Argonne National Laboratory via OC3 linkage, and connect to the Maui IBM SP2 via the ACTS satellite. At the workstation cluster, the OC12 line is ATM multicast switched into a two level tree configuration servicing 19 workstations linked by OC3 lines. The terminals of the workstation cluster are located at 7 hospitals across the Los Angeles metropolitan area. Residual bandwidths on the workstation cluster are used to support video tele-conferencing and distributed patient data queries between the 7 facilities with distributed local computation providing image processing and display. The 90 GFLOP load on the supercomputers satisfies the requirements of asynchronous tasks such as virtual reality, interactive volumetric visualization and NGC tissue classification algorithm processing. The data sets comprise two GigaByte of Biomedical scan data acquired with x-ray computed tomography (CT), magnetic resonance imaging (MRI) and positron emission tomography (PET) scanners located at the 7 USC hospitals.

Other demonstrations in computational biology include

(Several applications were intended to run at SC'95 but could not do so due to I-WAY operational difficulties. However these users did have the opportunity to use their applications on the vBNS following the conference.)

Virtual reality (CAVE)

Argonne and NCSA are investigating performance models of interactive, immersive visualization for scientific applications, looking at methods of removing the barrier of distance while sharing a virtual experience. Argonne and NCSA have proposed to NSF that the vBNS be extended via the Ameritech NAP to Argonne National Laboratories, where connecting to the CAVE there would enhance the progress of this research effort.

At Supercomputing'95, experiments in VR included those of:

Visualization

Several visualization projects are using vBNS resources, some of which were demonstrated at SC'95.

In pursuit of increased effectiveness of wide area networks for real scientific computing applications, Adam Beguelin (PSC, CMU) is using communications libraries that adapt to underlying network topology, and heterogeneous workstation clustering, The latter application requires selective routing of traffic because needed workstation clusters are separated from the DMZ at some sites by one or more routers.

Joel Welling (PSC) has explored the distribution of an algorithm for object-oriented parallel volume rendering (VFleet) over geographically distributed supercomputing resources. They have found thus far that VFleet is less efficient in this mode than when run on a single machine (e.g., PSC's T3D), so the results have not been of great practical use yet.

Two visualization demonstrations at Supercomputing'95 were:

Performance measurement

Application level usage statistics

Von Welch (NCSA) provided the vBNS with a library for logging application level network usage information.

SNMP visualization

Daniel A. Reed (UIUC) and colleagues are demonstrating applications for visualization of network performance and usage statistics.

ATM traffic measurement

David Tipper and Sujata Banerjee are testing an ATM broadband traffic generator/measurement device donated by Hewlett-Packard, to characterize and experiment with generation of ATM traffic workloads.

Miscellaneous

New personnel

NCAR hired Chris Fair to perform the functions of the NLANR engineer funded by the NLANR, effective 1 November 1995.

NCSA hired Mike Haberman for NLANR activities, effective September 1995.

Feedback from the user community

outreach

Charlie Catlett has been serving as Technical Advisor to NSF (Steve Goldstein, NCRI), representing the U.S. in G7 high speed global network testbed planning, one of eleven projects outlined at the G7 Summit in Brussels in early 1995. He participated in two meetings this quarter with G7 representatives and telecommunications carriers from each country.

Claffy began liaison efforts with the research community, including at University of Colorado, LBL, USC/ISI, Digital, Xerox Parc, Bellcore, Cisco, and Stanford University.

Acceptable use policy

Several potential and active vBNS users have expressed concern regarding the nature of the vBNS as an experimental testbed. A realistic wide-area test of many research applications, such as those for real-time data distribution systems, require considerable investment by participating sites to get connected, an unlikely prospect if the vBNS were not available on a routine basis. One researcher noted:
How many of the really useful experiments would have been done on the original NSFnet if the ground rules had been similar to those of the vBNS. Quite likely that the NSFnet would not have progressed very far beyond high speed connections among supercomputer centers.
The vBNS is chartered for research in a variety of networking and application areas, much of which is most suited to infrastructure that is well coupled to the commercial production Internet. This close coupling, however, presents potential policy problems with use of the vBNS, described further in the next section.

NSF sponsored vBNS workshop

NSF hosted a workshop entitled, vBNS and Networking and Application Researchers in Washington, DC on 22-23 June 1995. The purpose of the meeting was to begin a dialogue among networking researchers, applications researchers, NSF staff and MCI staff concerning emerging opportunities associated with the vBNS. The workshop was chaired by Dr. Jonathan Turner (Washington University, St. Louis), and Dr. Paul Messina (Caltech). NLANR representatives (Randal Butler (NCSA), K. Claffy and Hans-Werner Braun (SDSC), Charlie Catlett (NCSA), Richard Loft (NCAR), and Jamshid Mahdavi (PSC)) represented their sites at the workshop. The sites were expecting an NSF report of the workshop shortly after, in order to respond to feedback from the research community articulated therein, but this report has not yet finalized. However, several concerns clearly emerged during the workshop itself, including the need for
  1. extending vBNS connectivity to a larger number of other network testbeds, research sites, and users, in order to build a more potent community of network researchers
  2. researchers to have more open access to network internal configuration parameters, APIs, and software
  3. permissiveness with respect to the Acceptable Use Policy (AUP) to encourage innovative developments, focusing less on high bandwidth as a criterion
  4. support for a large bandwidth range of researcher needs, specifically a valuation policy that recognizes that meritorious traffic is not necessarily high bandwidth
  5. strong push within R&D community to deploy as soon as possible technology that would facilitate more graceful connection between policy-restricted networks and the commodity Internet infrastructure (e.g., link sharing, routing based on more than destination address, possibly source address, ports, type-of-service fields)

From what we can see, the NSF is definitely listening to and preparing appropriate policy responses to this and all valuable feedback from the research community, and is determined to address concerns that could prevent the maximum return on the vBNS as an investment in the research community. Furthermore, MCI is also undertaking collaborative efforts with its equipment vendors to maximize the attractiveness (i.e., programmability, open architecture) of vBNS equipment to the network research community. MCI has also consistently been very collaborative with the NLANR group, so that vBNS and the NLANR becoming quite integral to each other. We are quite optimistic about the likelihood of both reasonable application of the AUP, and rapid expansion of the vBNS to other experimental network testbeds, sites, and offsite users. We look forward to a cooperative and productive (but busy) coming year.


contributors to this report included: Randy Butler, Charlie Catlett, Bilal Chinoy, Bruce B. Johnson, Jamshid Mahdavi, Matthew Mathis, and Marla Meehl

[NLANR home page] info@nlanr.net 27 dec 1995