Check out the new USENIX Web site.
SummariesUSENIX

 

First Conference on Network Administration

Santa Clara, California,
April 7-10, 1999

Keynote
Session: Monitoring and Video
Session: Configuration Management and Security
Invited Talks

KEYNOTE ADDRESS

Home, Road, and Work Have Merged Via the Internet

Norm Schryer, AT&T Research
Summary by David Parter

"If the code and the comments disagree, both are probably wrong." — Norm Schryer

Norm Schryer has an OC-3 to his house. And a cable modem. Why? He works for AT&T Research, where his group has been investigating broadband network services. In fact, they have spent most of the last five years on the infrastructure required to do research on broadband services.

Schryer discussed many aspects of widespread high-speed multilocation networking and their implications for network managers:

  • Services are widely distributed.
  • Distributed responsibility among PCs, laptops, palmtops, etc.
  • Central fragility: The center of the network is subject to failure.
  • Security issues are important and cause problems too.
  • "The bigger and better the network, the easier it is to destroy itself."

WAH/WOOH: Work at Home / Work out of Home

Most members of Schryer's group have cable modems at home. All of them work all the time (or close to it) "fixing it, making it, or doing something with it." They are "on net" all the time. This makes for more loyal employees, some of whom move in order to get broadband services. Those who live in areas where cable modems or xDSL are not available are "geographically challenged."

VPNs

For both the home user and the road warrior, Virtual Private Networks (VPNs) are required for security. The VPN is tunnelled over the broadband service (cable modem, xDSL, etc.) using IPSEC. According to Schryer, VPN software clients are very fragile, and VPN hardware is great but expensive. VPN software is like "rolling ball bearings down a razor blade—if any fall off, IPSEC breaks." Several vendors provide VPN software, but it is often usable only on specific versions of the operating systems and, especially with Windows, often breaks when other software or patches are installed.

Their solution is a dedicated computer running Linux, which provides the VPN and IPSEC software. All the other computers at home do not need to be reconfigured in order to use the VPN.

For performance reasons, Schryer strongly recommended against creating a VPN over the commodity Internet. Instead, you should contract with your broadband service provider for a high-speed dedicated link to your corporate network.

At this point, about half of AT&T Research is using its VPN system from home.

Road Warriors

When travelling, most users have a laptop running a Windows operating system and have to use a VPN software client (instead of the "moat," the Linux tunneling system). Their experience is that the Windows IP stack is very fragile, and IPSEC shuts down and locks out the user when it detects what it thinks is an attack. Frequent reinstalls of the laptops result.

Schryer also discussed other challenges for the network administrator when dealing with high-speed connections to the home and road: high-bandwidth applications such as music; universal messaging; PDAs; and wireless links and network telephony, among others.

In summary, he identified the following principles, with the note that if you are horrified at the scope of the problem, that is good.

  • Very strong vendor management is needed: insist on standards, interoperability, and reliability; SNMP management of IPSEC is a disaster; "pin your vendor to the wall on reliability"; if they lie, you pay for it.
  • 7x24 customer care—with a global network, customers are always on—7x24 coverage is "dirt cheap" for keeping customers.
  • "Do not step in the Microsoft": Microsoft network software is extremely fragile—can break at any installation of software, even nonnetwork software; Windows is a nonmanaged platform; segment what you are doing—keep critical stuff on a separate platform that is managed.

REFEREED PAPERS

Session: Monitoring and Video

Summary by Ryan Donnelly

Driving by the Rear-View Mirror: Managing a Network with Cricket
Jeff Allen, WebTV Networks, Inc.

"Cricket is not the same as MRTG" was the theme upon which Jeff Allen began his discussion of the new enterprise network-management tool. Cricket was born out of a need to forecast network growth and plan for expansion—not merely react to significant changes in usage patterns. In addition, Cricket, like its predecessor, provides for instantaneous "is this link up" management features. Also, like MRTG, Cricket is a cgi-based management tool, featuring graphical displays of traffic trends over configurable time periods with comparison to other dates/ times. As such, it provides information for long-term analyses of traffic trends on a specific link or group of links.

The system, while appearing MRTG-like, has evolved in numerous ways. An increase in configurability, via a hierarchical configuration tree, allows for data gathering from scripts, SNMP, Perl procedures, etc. The data is obtained by means of a collector, which runs out of cron and stores data in RRD files, later to be interpreted by a graphing program. The RRD/ grapher program implements "graphing on demand," which generates results similar to those displayed by MRTG.

In addition to graphing data, Cricket also gathers application and host-based statistics to monitor such events as PVC states, cable modem signal strength, and router CPU load.

Cricket has taken its place as the successor to MRTG. While it can display much of the same information, its hierarchical configuration tree and improved code allow it to perform many more tasks in a more efficient manner. The Cricket home page is <http://www.munitions.com/~jra/cricket>.

Don't Just Talk About the Weather—Manage It! A System for Measuring, Monitoring, and Managing Internet Performance and Connectivity
Cindy Bickerstaff, Ken True, Charles Smothers, Tod Oace, and Jeff Sedayao, Intel Corporation; Clinton Wong, @Home Networks

The Internet Measurement and Control System (IMCS) was developed to provide quantitative measures of Internet performance. IMCS provides such statistics as HTTP GET requests and ICMP echo times to flagship Internet sites. It then uses such statistics to delineate process limits for a given data set.

Such limits are obtained by two Perl procedures, TimeIt and Imeter. TimeIt measures the total time to look up a URL, connect, transfer data, and disconnect. Throughout the transfer period it also logs both transfer and error rates. Imeter completes the statistics-gathering engine by adding an ICMP echo component. Pings are issued to such strategic locales as root name servers and large Web sites.

The monitoring and alert component of IMCS is Web-driven. As soon as IMCS determines that the process limits have been violated for a specific measure, an "all-in-one" Web page can alert network operations center staff to potential problems on a link and thus allows for a degree of preemptive network management.

To judge by Intel's experience, IMCS can provide a network administrator with rich performance statistics about specific network traffic. But the system faces the problem of process limit setting, by virtue of the lack of a consistent statistical model.

In the future, the authors plan to integrate an expanded set of services to be monitored, such as SMTP and DNS, and also plan to integrate flow data.

Supporting H.323 Video and Voice in an Enterprise Network
Randal Abler and Gail Wells, Georgia Institute of Technology

Randal Abler and Gail Wells have been exploring the possibilities of implementing H.323, a voice- and video-over-IP transport. The H.323 standard allows for the proliferation of two-way videoconferencing over the Internet and other applications such as distance education.

Developed as a LAN protocol, H.323 is UDP-based. As such, extensive H.323 traffic has the capability of overwhelming TCP traffic on a congested link. In trying to check such traffic levels, most H.323 applications contain the ability to limit client bandwidth usage. In addition, a network-based bandwidth-limiting solution, such as MPLS, could be warranted. As evidenced by testing, bandwidth limitation is crucial not only for network-congestion reasons but also for client-performance reasons. Programs such as Microsoft NetMeeting experienced degraded performance when used on a link with a speed inconsistent with the program's bandwidth setting. Testing also showed that conventional modem links do not supply sufficient bandwidth for acceptable realtime video usage. However, with the advent of digital connectivity to the home, the future of H.323 appears bright.

Session: Configuration Management and Security

Summary by David Parter

Network Documentation: A Web-Based Relational Database Approach
Wade Warner and Rajshekhar Sunderraman, Georgia State University

Wade Warner described Georgia State University's use of a relational database to track and document configuration information for its network. An online Web-accessible database was preferable to paper records because the Web interface allows for access from all platforms, and because paper records don't scale and are hard to use.

The implementation that was described is specific to the GSU network, but it is easy to add other types of devices. It makes extensive use of Query by Example (QBE), so that changes in the device types do not require redesign or rewriting of the user interfaces. Currently they are using Oracle, but they have an MSQL implementation freely available.

Items included in the database are installed software, printers, and specific information about the computers (such as amount of memory, serial numbers, and MAC addresses). They track computers by hostname and IP address, which has caused some issues for multi-homed hosts.

Several commercial tools are available for similar tasks, including populating the initial database. All suffer from the same problem: keeping the database current. The GSU group populates the database by extracting information from a variety of existing sources (such as the DNS system), manually entering the data, and using a Web form.

There are three access levels for viewing the data: network admins have all access; the middle layer can query specific network devices and generate preformed reports; and the users can see the availability of certain resources such as printers and software. All access is controlled with a user ID and password.

There were some interesting questions and comments from the audience about using port-scanning and SNMP to gather some of the initial data. (Wade said that they are working on that.) Also, USC had (or has?) a similar system, requiring the MAC address be entered into the database before they will assign an IP address.

Just Type Make! Managing Internet Firewalls Using Make and Other Publicly Available Utilities
Sally Hambridge, Charles Smothers, Tod Oace, and Jeff Sedayao, Intel Corp.

Jeff Sedayao gave a very interesting presentation on the management of Intel's distributed firewall environment, which relies heavily on make and other common publicly available UNIX utilities.

The Intel firewall architecture has the following characteristics:

  • a typical screened subnet architecture.
  • inner/outer firewall routers with a "DMZ" containing bastion hosts.
  • ISP routers on a common network segment outside the "outer" router (not in the DMZ).
  • geographically dispersed routers and firewalls, with failover from one firewall complex (DMZ site) to another as needed. (Currently there are eight firewall complexes.)

The bastion hosts need consistent configurations from one firewall complex to another, and this system has to be managed by limited staff.

Their solution is to view all the firewalls as a single distributed system rather than a collection of independent firewalls. A single set of source information is used to construct customized configuration files for each firewall (using make). Each DMZ network segment is described by a series of serializable segments of firewall access lists that can be combined in any order to function correctly. Each set of rules is stored in a separate file. For a given firewall, the constructed ACL is ordered for the best performance (most common rules first) for that particular firewall complex, preserving the required order of individual rules to enforce the desired access policy.

The bastion hosts are handled in a similar manner. Each is an exact copy of its counterparts at other firewall complexes. There is only one configuration for a given function (DNS, Web proxy, etc.). make is used to manage the configurations, and rdist (with ssh) is used to synchronize the bastion hosts with the master.

Their experience with this system has been positive, with the following lessons learned:

  • The system is scalable and robust for both support staff and users.
  • There is a fundamental need for discipline.
  • It is easy to ensure that a correct configuration is installed at all the firewall complexes. (For example, they had the "Melissa virus" check pushed out before their manager asked about it.)
  • It is also easy to propagate an error to all sites.
  • Admins must be trained to change the master, not the deployed configuration.
  • Version control is necessary.
  • Using RCS-like tools, one can easily look at changes from the previous configuration to try to identify the cause of a problem.
  • rdist can be used to compare the master configuration tree and the installed bastion host.
  • The router configurations are reloaded daily to ensure that the most current version is in use.

Future work is planned on automating testing of configurations before pushing to the bastion hosts, and on automating the firewall ACL optimization.

Tricks You Can Do If Your Firewall Is a Bridge
Thomas A. Limoncelli, Lucent Technologies, Bell Labs

Typical firewalls act as routers, even if the routing function is not otherwise necessary. The network on one side of the firewall is one IP network, and the network on the other side is another network. To applications on either end, the firewall looks like a router. Tom Limoncelli described his experiences with a firewall that acts as a bridge (transparent to the applications) instead.

The major advantage of the bridge firewall is that it can be inserted into the network without requiring a configuration change in any other devices. This also reduces the need for extra transit networks between the traditional (nonfirewall) router and the firewall, and allows for quicker installation (without downtime). In addition, Limoncelli argued that the software is simplified by not having to implement routing functions. (Not everyone agreed with this claim.)

According to Limoncelli, his is not a new idea—it was proposed (and perhaps built?) a long time ago, and then the concept died as people focused on router-like firewalls. He did not mention who had done the initial work.

He worked through several scenarios, including replacing a router-like firewall with a bridge-like firewall without downtime:

Step 1: Program the new firewall with the same rules.

Step 2: Insert into the network (which should take a few seconds and one extra network cable) on the local-network side of the existing firewall.

Step 3: Run both firewalls in series. Check the log files on the inner firewall—it should not filter (remove) any traffic at all. (The outer firewall should have eliminated all filtered traffic.)

Step 4: Remove the router-like firewall (removing the transit network). This will require short downtime, as the external router will have to be reconfigured from the transit network to the local network.

Additional rule testing can be added by deploying the bridge firewall first on the transit-network side. The existing firewall should show no packets filtered, as the outer (bridge) firewall should have removed them.

In addition to replacing existing router-like firewalls (or augmenting existing firewalls to provide multiple layers of defense), the bridge-like firewall can be used to add firewalls to existing networks without reconfiguration. For example, it is easy to add a firewall to protect a manager's office systems without disrupting the rest of the network. Unfortunately, it is also easy for someone to add a bridge-like firewall without consulting the network staff, since the firewall is transparent to the IP layer.

Many of the questions focused on details of the Lucent product, although Tom's experience is with the research prototype.

INVITED TALKS

New Challenges and Dangers for the DNS
Jim Reid, Origin TIS-INS

Summary by Ryan Donnelly

According to Jim Reid, the Domain Name System is definitely at an evolutionary crossroads. The introductions of IPv6, dynamic DNS, secure DNS, and Windows 2000 have sent DNS administrators around the globe scrambling to design more effective methods for managing the DNS infrastructure. In doing so they have encountered innumerable issues associated with redesigning DNS.

One of the primary concerns with DNS evolution is its hasty marriage to WINS, the Microsoft DNS-like system for Windows machines. In trying to integrate the two systems, IETF engineers have had to add multiple new records and features to the DNS system in order to make the migration as painless as possible. Two of the primary facilitators of the change are dynamic DNS and secure DNS.

Dynamic DNS is a process by which hosts are authorized to modify a given DNS zone file. This is necessary to continue to support the "plug-and-play" philosophy expected by Windows/WINS users. In implementing this system, engineers have encountered the problems of excessive zone transfers that are due to constant zone updates, and the problem of host authorization. While the excessive zone update problem can be solved by implementing a periodic zone-transfer schedule, the issue of authorization requires a more complicated solution: secure DNS.

Secure DNS is a method by which keys are assigned to hosts that are authorized to modify DNS data. While partially solving the problem of host authentication, the overhead involved in key management and encryption/decryption is substantial. In addition, the storage of such 1,024-bit keys in zone files has the capability of increasing the current zone files by as much as a factor of 10.

Reid concluded his talk by emphasizing the fact that much development work still has to be done, and that nothing is cast in stone. The presentation is available at <http://www.usenix.org/publications/library/proceedings/neta99/invitedtalks.html>.

Problems with World-Wide Networking
Holly Brackett Pease, Digital Isle

Summary by David Parter

Holly Brackett Pease described many of the challenges—technical, political, and logistical—of deploying a worldwide private IP backbone.

Digital Isle provides a worldwide IP backbone for its customers, bypassing the more common but often congested network access points and Internet exchanges. Instead, they contract for peering with the appropriate Internet service providers in a particular country as needed by their customers.

Technical challenges include understanding BGP and the way it is configured by the various ISPs and backbone providers to assure that traffic does in fact flow over the correct carrier and that routes are accepted (as appropriate) by peers and their providers. Another technical challenge is the differing telecommunications standards and terminology: X.21 is not a movie rating (but you still don't want to use it—G.703 is your friend). Actually, X.21 is an International Consultative Committee for Telephone and Telegraph (CCITT) standard defining the interface between packet-type data terminal equipment and a digital-interface unit. G.703 is a physical/electrical interface specification, used for leased lines (E1s, similar to T1s but all digital) in certain European countries and Australia, and supported by router manufacturers such as Cisco.

Logistical challenges include getting the network equipment to the site with the correct interfaces, cables, and everything else that is needed. Once the equipment has been shipped, it is often necessary to pay a hefty value-added tax to get it past customs. Pease pointed out that Fry's does not exist in most parts of the world, and if you can find a supplier for that missing cable, they won't take your credit card—they want a purchase order. (Anixter is a good supplier — they do have worldwide coverage.)

Political issues include local laws covering IP telephony and content. In one case, they had to register their networks as being from that country in order to be allowed to peer with networks in that country — despite being a partner of the largest ISP.

There were several questions:

Q: Are the peering agreements with the major players in each country peering of two equals, or on a payment basis?

A: Pay all partners—this is a technical peering and interchange agreement, not one ISP trading access to another. Digital Isle is providing a premium service for their customers; there is a cost. They don't peer at the local exchanges at all.

Q: Language issues?

A: Most of the ISPs that they have had to deal with have at least one or two staff with excellent English. If not, they use the AT&T international translation service.

Q: Content filtering?

A: Mostly not a problem, since Digital Isle already has content rules, and their customers are Fortune 1000 companies, not ISPs, and have no downstream customers. IP telephony rules vary from country to country, and it is necessary to make sure that the network enforces those rules.

The Little NIC That Could
Christopher J. Wargaski, RMS Business Systems

Summary by Ryan Donnelly

In an era of high-growth enterprise networks, NICs (network information centers) are becoming more and more expensive to run. Chris Wargaski's presentation on a cost-effective NIC encompassed the idea of a "project staffing model." In this model, the NIC is the first point of contact for all IT staff. Owned solely by the networking department, overall responsibility for the NIC is held by a middle-level manager. Below such a manager are the on-call manager, NIC analyst, and on-call analyst.

It is the responsibility of the on-call manager to track problem occurrence and resolution. Additionally, it is the responsibility of the on-call manager to report relevant problem-resolution statistics to upper management.

The NIC analyst is the primary point of contact within the NIC itself. The analyst is primarily responsible for answering phones and issuing trouble tickets. The analyst is also responsible for monitoring the overall health of the network in order to detect network problems preemptively. When a problem is detected, it is the responsibility of the NIC analyst to assign the ticket to the on-call analyst.

The on-call analyst is the technologist responsible for fixing the problem and updating the NIC regarding its status.

In the project staffing model, both the NIC analyst and the on-call analyst are obtained by rotating engineering/technical staff through the NIC, as determined by the NIC manager. The manager must consider extensive input from staff on scheduling conflicts and hold staff meetings on a regular basis to ensure a constant information flow. It should be emphasized to staff that training, while important and highly encouraged, is optional.

The presentation ended with several comments on the topic of NICs as hindrances as opposed to facilitators.

Splitting IP Networks: The 1999 Update
Thomas A. Limoncelli, Lucent Technologies, Bell Labs

Summary by David Parter

After a rap intro, Tom Limoncelli introduced his talk by commenting that when he and his co-authors first wrote the Bell Labs network splitting paper (1997), they were unsure if network splitting was interesting to anyone and if it would ever be done again. He then asked if anyone from Microsoft or Hewlett-Packard was in the audience — since they will (or may be) facing the same issues soon. In fact, the sysadmin group at Bell Labs has since been called upon to renumber more networks, involving more total computers, than in the original work.

The original problem was splitting the Bell Labs research networks (on a computer-by-computer basis) between AT&T and Lucent/Bell Labs, because of the AT&T "tri-vestiture" (splitting into AT&T, Lucent, and NCR). In addition, they took the opportunity to upgrade the networks, reorganize the file servers, and bring the workstations into a consistent configuration.

The project was a major success, mainly as a result of the amount of advance planning and effort spent "getting it right" before widespread implementation.

Prior to actually renumbering the networks, it was necessary to identify which workstations and file servers were destined for which networks. In some cases, this was not known for several months, because of uncertainty about which members of the staff were going to which company. In addition, data stored on the file servers had to be sorted and moved to a file server destined for the correct network. An important tool in this phase was a Web page that allowed users to check on the status of the project and read information about specific resources. Additions and corrections were supplied by the users. The users had less concern about potential disruptions since they were involved in the planning process.

The information about which workstations and file servers were destined for which network was also used to populate new NIS netgroups named for the new networks. The new netgroups were used to control filesystem exports on the newly reconfigured file servers. By generating the netgroups from the master list of which workstations were destined for which network, they eliminated a potential inconsistency.

The key technology used in the actual renumbering project was multi-netting all the Ethernets, so that all the original IP networks and the new networks were on the same physical networks. This allowed them to renumber workstations one at a time without disrupting all other workstations on the network.

Repatching data jacks from their current network hub or switch to a network switch designated for the correct future network was done in parallel with the renumbering. As all the switches and hubs were connected together, this was transparent to the users.

When the workstation and fileserver renumbering was thought to be complete, the connections between the two new (physical) networks was reduced to a single bridge. Traffic was monitored at the bridge in order to detect workstations and file servers that ended up on the wrong physical network. After all cross-traffic (other than broadcast) was eliminated, the bridge was removed, the two networks were independent, and the project was completed.

Network Management on the Cheap
Rob Wargaski, RMS Business Solutions

Summary by Ryan Donnelly

Rob Wargaski's talk on cheap network management included a fairly extensive summary of UNIX network-management power tools and their most effective uses. He began his talk by discussing two different types of network management: tactical and strategic. Tactical network management is essentially day-to-day reactive monitoring and firefighting. Strategic monitoring is more focused on data analysis and trend forecasting.

Some of the tactical tools mentioned included ping, traceroute and arp, used to assess baseline network connectivity. An additional tool is K-Arp-Ski, a comprehensive X-based sniffer and network analyzer. It can monitor an IP/MAC address combination, multiprotocol traffic, and NIC vendor information. He also mentioned several SNMP-based tools, such as SNMPSniff and Sniffit, along with the standard UNIX tcpdump.

Discussing strategic network-management tools, Rob first brought up the topic of enterprise network-management packages such as HP OpenView. He also mentioned some CMU/UCD character-based SNMP tools with API interfaces. These included snmpget, snmpgetnext and snmpwalk. The combination of these three tools allows a network administrator to retrieve either a select amount of data from a network device or the entire tree of device information. Because such data is atomic in nature, strategic management requires postprocessing of the information in order to determine trends and typical network behaviors.

The Evolution of VLAN/ELAN Architecture at Vanderbilt University
John J. Brassil, Vanderbilt University

Summary by David Parter

John Brassil described the Vanderbilt University network and its evolution from a "flat" bridged network to one making use of many virtual LANs (VLANs), mostly grouped by common services or administrative structure. Vanderbilt's network includes almost 10,000 computers and will grow soon with the addition of student dorms.

The talk included a large amount of background information on VLANs, emulated LANs (ELANs), and ATM LAN Emulation (LANE).

The network is continuing to evolve:

  • While the IP network is currently not routed, they will switch to a routed IP network soon. This is difficult because in the past they have had a lack of planning in IP address allocation.
  • They are encouraging users to use local Internet service providers for off-campus connectivity using cable modems and xDSL, and are phasing out their campuswide modem pool.
  • They have recently added a firewall at their Internet connection. It is currently being used only for traffic logging, not as a firewall.

Interoperable Virtual Private Networks (VPNs), Directory Services, and Security
Eric Greenberg, Seine Dynamics

Summary by Ryan Donnelly

Eric Greenberg presented several different uses and implementations of Virtual Private Networks (VPNs). Some potential uses include remote dial-in applications, private networks, and business-to-business secure networks.

There are several methods for implementing such VPNs. The first is the PPTP protocol, which provides for multiprotocol tunneling over PPP. While beneficial for moving such protocols as Appletalk and IPX over an IP link, PPTP does not introduce any new security features. L2TP, however, does.

L2TP was described as the successor to PPTP. It is transport-independent, making it a much more viable option for networks that implement various transports, and does not necessarily require IP. It is also UDP-based and has the added feature of enabling digital certificates.

X.509 certificates are yet another method of implementing VPN security. These allow machines within a network to authenticate one another by means of a unique certificate that is stored on a central directory server. An offshoot of this method is public key cryptography, which can be implemented by sending a copy of a public key with the certificate itself. Such certificates can also be issued by numerous emerging certificate authorities, such as Verisign.

IPSEC is an additional method of designing secure virtual private networks. It provides for IP-only tunnelling and is integrated into IPv6 via the "next-header method."

Greenberg also discussed some possible security-policy and key management issues. He suggested that certificate and key management be performed on a central directory server that would hold the certificates and their associated security policies.

In a VPN, however, security is not the only concern. As traffic levels continue to rise, quality-of-service (QoS) issues also arise. Multiprotocol Label Switching (MPLS) allows VPNs to influence QoS issues by assigning a priority to a certain type of traffic.

Greenberg concluded his talk by analyzing the features of PPTP, L2TP and IPSEC, and the environments in which each is most useful.

Internet Measurements
k. claffy, Cooperative Association for Internet Data Analysis at the San Diego Supercomputer Center, UCSD

Summary by David Parter

k.c. claffy gave an entertaining and blunt talk about Internet measurements, tools, and interesting findings from CAIDA. She also provided a lot of answers to what she called "cocktail party" questions — interesting (but not really useful) facts about the Internet. According to k.c., the current state of measuring the Internet is "abysmal science." Because of the growth of the commercial Internet, researchers can't get access to the network infrastructure and operational data as well as they could when the primary network was NSFnet. They do have access to the vBNS, which helps. (vBNS is very-high-performance Backbone Network Service, a product of a five-year cooperative agreement between MCI and the National Science Foundation. See <www.vbns.net>.)

She pointed out that this doesn't stop researchers from building "junk" and doesn't stop users from doing "random junk." There is a lot of measurement activity now (at least 11 organized groups), but the groups don't talk to one another, and at this time there is no correlation of the data sets (e.g., workload versus traffic).

She identified and described four areas of measurement:

  • topology
  • workload characterization (passive measurements)
  • performance evaluation (active measurements)
  • routing (dynamics)

One of the more interesting aspects of her talk was the visualization tools that CAIDA is using to explore the data. Unfortunately, they are hard to describe. Readers are advised to visit <www.caida.org> and look at the samples. Slides from the talk are online at <http://www.caida.org/Presentations/>.

Some of the "cocktail party" facts:

  • Half the packets on the Internet are about 40 bytes long.
  • Half the packets on the Internet have a full 1500-byte payload.

 

?Need help? Use our Contacts page.
First posted: 5 Nov. 1999 jr
Last changed: 5 Nov. 1999 jr
Conference index
;login: issue index
Proceedings index
USENIX home