Joel Apisdorf <firstname.lastname@example.org>
MCI Telecommunications Corporation
k claffy <email@example.com>
National Laboratory for Applied Network Research
Kevin Thompson <firstname.lastname@example.org>
Rick Wilder <email@example.com>
MCI Telecommunications Corporation
The Internet is rapidly growing in number of users, traffic volume, and topological complexity. At the same time it is increasingly driven by economic competition. These developments render it more difficult, and yet more critical, to characterize network usage and workload trends, and point to the need for a high-performance monitoring system that can provide workload data to Internet users and administrators. To ensure the practicality of using the monitor at variety of locations, implementation of low-cost commodity hardware is a necessity.
In its role as the network service provider for NSF's vBNS (very high speed Backbone Network Service) project, MCI has undertaken the development of an OC3-based monitor to meet these needs. We will describe and demonstrate our current prototype. The goal of the project is to specifically accommodate three incompatible trends:
Specific design goals that led to the current prototype are
The vBNS has deployed the monitor at all vBNS sites as of January 1997. The software is freely available to others for use elsewhere, and both the flow analysis code and monitor architecture are public domain.
We attach the OC3MON ATM NICs to an OC3 fiber pair carrying IP traffic, connecting the receive port of each ATM card to the monitor port of an optical splitter, which carries 5 percent or 10 percent (depending on the type of splitter used) of the light from each fiber to the receive port of one NIC (as of January 1997, the dual splitter cost is about $800; the NICs run about $1,200). Attached to an OC3 trunk terminated on a switching device (e.g., ATM switch or router), one of the OC3MON NICs sees all traffic received by the switching device and the other NIC sees all traffic transmitted by the switching device. In the vBNS, we will attach an OC3MON to each connection from the wide area ATM backbone to the primary nodes at the supercomputer centers. (See figure 1.)
Figure 1: National Science Foundation very high speed Backbone Network Service (vBNS) topology. The backbone connects the five supercomputer centers at Cornell, Pittsburgh, Urbana-Champaign, Boulder, and San Diego at OC3, with T3 connectivity to the four original NAPs in Washington, DC, Pennsauken, Chicago, and San Francisco. MCI is currently upgrading the backbone to OC12 in 1997.
The DOS-based software running on the host PC consists of device drivers and a TCP/IP stack combined into a single executable; higher level software performs the real-time flow analysis. Several design constraints motivated our decision to use DOS-based functionality rather than a Unix kernel. First, the Texas Instruments cards in the original OC3MON design required polling the interface card for cells at 1/128 the cell rate in order to obtain accurate timestamp granularity at full OC3 rate, since the card itself did not timestamp the cells. Monitoring a full duplex link requires two cards in the machine, which meant that we had to reprogram the timer interrupt to occur every 128/(2*353207.5) = 1/5518 second. Because Unix has a higher interrupt latency than DOS, we were better off with DOS at that point.
Our latest design uses Fore cards that can attach timestamps to the cells on their own; the host no longer needs to poll the card at all. We need to interrupt only at most every 1/40 second (i.e., if both links received 40 byte packets simultaneously), so low latency is no longer a constraint. However, we would not have gotten a prototype working without the control that DOS provided.
Second, we needed the ability to monopolize the entire machine, which is easier with DOS than Unix. OC3MON needs to provide the hardware with large blocks of contiguous physical memory, so we did not want the operating system to have to maintain knowledge about the memory and possibly fragment, resulting in lower efficiency on card-to-host buffer transfers. We did not want the kernel to suddenly decide it needed to capture the PCI bus to swap a page to disk, nor did we want the analysis software to fall behind because the kernel scheduled another process. We also wanted more control than Unix provides over when TCP/IP could have a time slice. These concerns will become even more important for OC12MON, which faces four times the data rate, or even if we dedicate a separate machine to each direction of traffic faces twice the data rate.
The disadvantage of DOS is its blocking I/O routines, whereas Unix would provide nonblocking I/O. Ports of OC3MON to Linux and NetBSD are now under way by vBNS researchers at NCSA and SDSC, respectively.
The host software directs each ATM NIC to perform AAL5 (ATM Adaptation Layer) reassembly on a specified range of virtual circuit and virtual path identifiers (VCI/VPI) (default is 7 bits of VPI and 13 bits of VCI, zeroing unused bits of both). Note that Cisco routers and Fore switches also support AAL3/4, but MCI does not use it on either the vBNS or its commodity infrastructure because it consumes an additional 4 bytes from each cell (above the 5 already used for the ATM header) to support submultiplexed channels within a given VP/VC. Since the LLC/SNAP 8-byte-per-frame header, which RFC1577 says a router should insert by default in front of all packets, already includes a 2-byte ethertype field that allows, if needed, multiplexing of different protocols (IP, IBM SNA, Novell IPX, etc.) on the same VC, including AAL3/4 support in the design would not have been beneficial. Note that the addition of any bytes to a simple (no data attached) TCP ACK causes the 8-byte ATM AAL5 trailer to require another cell, doubling the bandwidth used by such packets. Also note that the AAL5 trailer already has room for a user-defined 16-bit field, and the router does not care about protocol information until after receiving a complete packet, so supporting protocol multiplexing did not require the per-packet, variable-length, upfront overhead of an LLC/SNAP header.
AAL5 makes use of a user-defined single-bit field in the ATM header to indicate whether a cell is the last in a frame. AAL5 also assumes that cells for a given frame will not be interspersed with cells for another frame within the same VP/VC pair--no multiplexing occurs inside a VC. Combined with a single bit of state per VP/VC pair maintained by the receiver, which indicates that the cell is in the middle of a frame for that VP/VC pair, there is enough information to reassemble the frame.
The receiving card normally also needs a pointer to the location in host memory (or card memory if the card were to buffer received frames before DMAing them to the host, which ours does not) where it has put previous cell payloads for incomplete frames, so that it can store future cells contiguously, or at least maintain a linked list. Once a SAR (segmentation and reassembly) engine design involves this leap from one bit to the size of a pointer, most go even further and use several more words for management purposes. VC table entries on the order of 16 to 32 bytes are not uncommon. Thus most ATM NICs are limited to on the order of 1024 VC/VP combinations active at a time.
Since OC3MON has no need for data beyond the first cell, and since it already maintains per-flow state on the host, we chose to limit the per-VC state on the card to the bare minimum: 1 bit (2 bits when we implement up-to-3-cell capture for OC12MON). Although the Fore cards have 256KB of memory, some of it is used for the i960 code (about 32K), the OS, reassembly engine data structures, and the stack. Since the VP/VC lookup needs an exact power of two, the largest we could get was 128KB. (Single-bit state for 2^20 VP/VC combos = 2^17 bytes = 131072 bytes = 128KB.)
The cards for OC12MON will have 4MB of memory, all of which will be available for state information, allowing us to keep single-bit state for all 2^24 = 16777215 possible ATM UNI (user to network interface) VPI/VCI combinations. Using 2 bits of state per VPI/VCI combination (for example, if we wanted 3 cells per packet sent to the host) leaves room for only half as many VPI/VCI combinations. Note that when we copy multiple cells of the same packet to the host, the card will not place them near each other, so the host must do further reassembly using the ATM headers.
Examining 20 bits of VCI/VPI information allows OC3MON to monitor over one million VCs simultaneously. The host controls exactly how many bits of the VCI this 20-bit index will include; the rest derive from the VPI. The host also specifies at startup what to expect for the remaining bits of the VPI/VCI, i.e., those not used for indexing into the card's state table. The card can then complain about, or at least drop, nonconforming cells.
Many SAR engines choose to completely ignore the VPI and any bits of the VCI not used for indexing. When presented with the arbitrary VPI/VCI combinations we expect to see on a general-purpose monitor, inevitable aliasing will cause collisions in reassembly state among VPI/VCI pairs. OC3MON avoids this situation by (1) using a large number of VPI and VCI bits for its table lookup, leading to more successful reassemblies in the presence of arbitrary channel usage; and (2) comparing the bits it does not use for indexing with the expected values as described above, which keeps unsuccessful reassemblies from corrupting successful ones.
Since we want OC3MON to be able to see traffic on (almost) any VPI/VCI without prior knowledge of which circuits are active, and because the fast SRAM (static random access memory) used on such ATM cards for state tables is expensive and not amenable to modification by the consumer, this design turned out to be extremely advantageous.
The AAL reassembly logic is customized to capture and make available to the host only the first cell of each frame. The 48 bytes of payload from these cells typically contain the LLC/SNAP header (8 bytes), IP and TCP header (typically 20 bytes each). Copying the 5-byte ATM header also allows us the flexibility of doing ATM-based analysis in the future. The SAR engine discards the rest of each AAL5 protocol data unit (PDU, equivalent to a frame or IP packet), limiting the amount of data transferred from the NICs over the PCI bus to the host. Although as yet unimplemented, one could increase the amount collected to accommodate IP options or larger packet headers as specified for IP version 6. Currently, however, the cards only pass the first cell of each packet, so when IP layer options push part of the TCP header into the second cell, these latter portions will not be seen by the host. Although this is suboptimal, we decided the savings in PCI (peripheral component interconnect) bus cycles, host memory, and CPU usage justified this decision.
Each NIC (network interface card) has two 1MB buffers in host memory to hold IP header data. These cards are bus masters, able to DMA (direct memory access) header data from each AAL5 PDU into the host memory buffers with their own PCI bus transfer. This capability eliminates the need for host CPU intervention except when a buffer fills, at which point the NIC generates an interrupt to the host, signaling it to process that buffer up to memory while the NIC fills the other buffer with more header data. This design allows the host to have a long interrupt latency without risking loss of monitored data. The NICs add timestamps to the header data as they prepare to transfer it to host memory. Clock granularity is 40 nanoseconds, about 1/70 of the OC3 cell transmission time.
The resulting trace is conducive to various kinds of analysis. One could just collect a raw timestamped packet-level trace in host memory, and then dump the trace to disk. This technique is useful for capturing a detailed view of traffic over a relatively brief interval for extensive future study. However, because we currently use the DOS-supplied disk I/O routines, which are blocking, we cannot write to disk simultaneously with performing flow analysis. Therefore one can only collect a trace as big as the size of host memory, which in our case would be 128 million bytes, and then must stop OC3MON header collection to let OC3MON transfer the memory buffer to disk. In the future we hope to develop separate I/O routines that directly use the hardware, allowing us to keep up with OC3 line rate, assuming some minimum packet size.
Because the amount of data captured in a packet-level trace and the time needed for our disk I/O inhibit continuous operational header capture, the default mode of OC3MON operation is to maintain IP flow statistics that do not require the storage of each header. In this mode of operation, concurrently with the interrupt driven header capture, software runs on the host CPU to convert the packet headers to flows, which are analyzed and stored at regular intervals for remote querying via a Web interface. The query engine we use is similar to that found in the NLANR FIX West workload query interface (http://www.nlanr.net/NA/). We will describe the methodology for deriving flow information in the second half of the paper, followed by some example snapshot statistics taken with OC3MON.
We tested OC3MON on a concatenated OC3, or OC3c, link fully occupied with single cell packets (as would occur in the admittedly unlikely event of continuous TCP ACKs with no data and LLC/SNAP disabled on the routers), which yields 353207.5 packets per second (or in the single-cell packet case, the same number of cells) across each half-duplex link. Each header, including timestamp, ATM, LLC/SNAP, IP, and TCP headers, consumes 60 bytes, so the internal bus bandwidth required would be 353207.5 * 2 * 60 * 8 = 339 Mbits. The 32-bit, 33MHz bus in the PC is slated at 1.056 gigabits, so we do not expect bus bandwidth to be the bottleneck until we need to support OC12. There are already extensions to the PCI standard to double the bus width and speed, so when we need to support the worst OC12 workload (i.e., single-celled traffic), the bus technology will likely be available. (Digital has already demonstrated the 64-bit part.)
We do not currently support sampling in the capture or flow analysis software. Although sampling is one option to avoid losing gaps of data during traffic burst, it changes the statistics of a timeout-based flow analysis in a most unclear way. For simply collecting packet headers, or to venture into the murky statistics zone that flows of sampled packet streams would involve, we could modify the card to support sampling in the future. Testing OC3MON with single-cell packets on OC3 indicated that we do not lose packets, so supporting sampling is not a high priority for us at this time.
Secure access is a problem for any machine, especially monitoring-capable ones. On both the vBNS and the commodity MCI Internet backbone, the monitoring machines live in locked machine rooms or secure terminal facilities. One can also require the monitor to accept packets only from certain IP addresses, or configure nearby routers to block packets from unknown addresses from reaching it.
We obtain the flow data summary from OC3MON via a password-protected remote query to a port on the machine. This level of security is equivalent to that provided by most SNMP implementations. The query process triggers OC3MON to clear the current flow summary, but OC3MON retains the active flows in memory.
An ANNEX terminal server supports console one-time password access to OC3MON.
In deriving flow profile information from packets, we need to establish a definition of what constitutes a flow. Since the appropriate constraints to put on what one labels a flow depend on the analysis objective, our methodology specifies a set of parameters that are configurable based on the analysis requirements.
We specifically do not restrict ourselves to the TCP connection definition, i.e., SYN/FIN-based, of a flow. Instead, we define a flow based on traffic satisfying specified temporal and spatial locality conditions, as observed at an internal point of the network (e.g., where OC3MON sits). That is, a flow represents actual traffic activity from one or both of its transmission endpoints as perceived at a given network measurement point. A flow is active as long as observed packets that meet the flow specification arrive separated in time by less than a specified timeout value, as figure 2 illustrates. The lower half of the figure depicts multiple independent flows, of which hundreds of thousands may be active simultaneously at WAN transit points.
Figure 2: Defining a flow based on timeout during idle periods
This approach to flow characterization allows one to assess statistics relevant to issues such as route caching, resource reservation at multiple service levels, usage-based accounting, and the integration of IP traffic over an ATM fabric. Our definition of the timeout is similar to that used in other studies of timeout-based traffic behavior [4, 6, 8]. Jain and Routhier originally selected for their investigation of local network traffic a timeout of 500 milliseconds. Wide area traffic studies of the transport layer have typically used longer timeouts, between 15 and 20 minutes [8, 10]. Caceres et al. used a 20-minute timeout, motivated by the ftp idle timeout value of 15 minutes, and after comparison to a 5-minute timeout yielded minimal differences. Estrin and Mitzel  also compared timeouts of 5 and 15 minutes and found little difference in conversation duration at the two values, but chose to use a timeout of 5 minutes. Acharya and Bhalla  used a 15-minute timeout.
We explored a range of timeouts in Claffy et al. , and found that 64 seconds was a reasonable compromise between the size of the flow table and the amount of work setting up and tearing down flows between the same points. The timeout parameter is configurable in OC3MON; we have used the default of 64 seconds for the measurements in this paper. Initial tests with timeouts as large as 10 minutes did not significantly increase the number of flows, but we have not yet tested it under heavier data streams.
This timeout-based flow definition allows flexibility in how one further specifies a flow. There are other aspects that structure a flow specification: directionality, one sided vs. two sided, endpoint granularity, and functional layer.
Network administrators may want to define flows at a coarser granularity, such as aggregating network number pairs for which they create virtual circuits across their transit network. For example, an ATM cloud may bundle many finer grained IP flows within each ATM circuit. Conversely, a finer granularity would be necessary for providing special service to a single instance of an application (e.g., a videoconference).
These examples illustrate the importance of flexibility in the parameterization of a flow model and the need to ground a flow specification in the requirements of the network, and even allow at any point in the network for multiple simultaneous flow specifications. One may want to assume flows by destination network address for routing; by process pair for accounting; by source address for accounting and policy routing; by destination address or host or network address pair for bundling flows across ATM virtual circuits; or by address plus precedence information for flows at multiple priority levels.
Several factors motivate our decision to restrict ourselves to an observed state model, all reflective of one circumstance: The Internet is inherently a connectionless datagram environment, and thus connection-oriented information cannot always be assumed available. We provide further details in an earlier study .
These four aspects--directionality, one-sided vs. two-sided aggregation, endpoint granularity, and functional layer--provide a framework for specifying a flow profile structure. We designed the flows analysis software in OC3MON as flexibly as possible. One can specify a specific flow timeout (in seconds), an endpoint granularity (network, host, or host/port), and one-sided or two-sided flows (source, destination, or pair). One can also restrict OC3MON to analyzing flows for a specific transport protocol, port number, or host address.
For our flow profiling we use host pair plus source and destination application identifier (i.e., UDP/TCP port number), if they exist. That is, for the measurements in this paper, OC3MON considers a flow unique based on its protocol, source IP address, destination IP address, source port and destination port, and a 64-second timeout. A packet is considered to belong to the same flow if no more than 64 seconds have passed since the last packet with the same flow attributes. When flows time out, they are passed up to the statistics routines that update accumulators for remote querying via the Ethernet interface at regular intervals. The results of these queries, still in raw flow format, are then stored on a Web server that supports a menu-driven interface. The menus, illustrated in figures 4, 5, and 6, allow users to customize graphs of the data according to their interest.
The current release of OC3MON supports the granularity of classless network using a CIDR-aware IP-to-AS-path mapping derived from a periodically updated core Internet backbone routing table. This includes flow conversion to the AS granularity, which allows for assessment of traffic flow at a more convenient macroscopic level. Table 1 shows a sample AS report. Note that the actual data shown represents one hour of vBNS traffic collection, and is not necessarily representative of long-term traffic patterns on the vBNS.
Table 1: AS Report from OC3MON
Flow Report Top AS Pairs sorted by bytes
Site Location: NCAR
Date: 01/30/97, Time: 16:05:00
|#||AS src name||AS src number||AS dest name||AS dest number||#Flows||#Packets||#Bytes||Cum Percentage of Total Trunk Traffic|
|6||XX XXXXXXXX XXXXXX||xxx||XXXXXXXXX||xxxxx||115||95079||4.15e+07||67.25|
On the microscopic level, it is useful to identify the heavy traffic sources and destinations. OC3MON reports on "heavy hitters": IP addresses that are most frequently seen according to flows, packets, and bytes.
Table 2: Heavy-Hitter Report from OC3MON
Top Heavy Hitters sorted by bytes
Site Location: NCAR
Date: 01/30/97, Time: 14:35:01 a 5-minute sample
|#||Source IP address||Hostname||proto||port||#Flows||#Packets||#Bytes||Cum Percentage of Total Trunk Traffic|
OC3MON also reports active flows in conjunction with a remote real-time graphic display. In addition to reporting based on expired flows, OC3MON can periodically multicast summary information, including source/destination IP addresses, of active flows. These multicast updates allow for monitoring the data at multiple locations. The vBNS has this feed activated on its five OC3MONs throughout the network. Each monitor is given a multicast address on the command line on which to transmit, and a tunnel endpoint used as the destination for encapsulating the IP multicast. The vBNS Cisco router at each node acts as the tunnel endpoint, forwarding multicast data to vBNS-connected destinations. The receiving process is a Web-based Java applet that displays overall average and instantaneous bit rates of IP address pairs. Figure 3 shows a screen capture of the program displaying the active feed from the vBNS OC3MON at PSC.
Figure 3: OC3MON active flow display
To illustrate the kind of graphs and tables one can retrieve, we provide sample graphs of OC3MON measurements on an OC3 trunk of MCI's IP backbone and MCI's vBNS during the period between 25 January and 30 January 1997. Space restrictions prevent us from showing every graph type in this paper; we provide only a small set of possible plots to illustrate the utility of the tool.
Figures 4, 5, and 6 show the main menu and two submenu form interfaces for the query engine, respectively.
Figure 4: The main menu lets you choose either a snapshot of a specific monitoring interval (configured for 5 minutes here), or one of the submenus.
Figure 5: Time-series menu for OC3MON query engine (invoked upon clicking option 1, Time-Series, in main menu)
Figure 6: 2-D profiles menu for OC3MON query engine (invoked upon clicking suboption C, Graphic profile, option in main menu)
The main menu lets you choose either a snapshot of a specific monitoring interval (configured for 5 minutes here) or one of the submenus. The snapshot returns a page similar to that shown in figures 7 and 8.
Figure 7: Flow assessment snapshot of traffic during single interval of OC3MON collection--page 1 (invoked upon clicking suboption A, Printed summary, in the main menu)
Figure 8: Flow assessment snapshot of traffic during single interval of OC3MON collection--page 2 (invoked upon clicking suboption A, Printed summary, in the main menu)
Figures 9, 10, 11, and 12 show packets, bytes, flows, and average per-second packet size for the six-day period. Rates for packets and bytes are measured for each direction on the OC3 trunk. During this measurement interval, the average number of packets per second cycled between 10 and 45 thousand, depending on direction. (MCI has also installed parallel OC3s where traffic demands require them, in addition to OC12s, into its IP infrastructure.)
The average number of flows per second goes from around 150,000 at night to over 450,000 in the case of Tuesday, 28 January. Note that the average packet size goes in the opposite direction--the per-second average packet size gets larger at night, presumably due to less interactive traffic and likely occurrence of automatic backups.
Figure 9: Average packets per second over 5-minute intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
Figure 10: Average bits per second over 5-minute intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
Figure 11: Average and maximum flows per second over 5-minute intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
Figure 12: Average and maximum packet size over 5-minute intervals on MCI IP OC3 backbone trunk, Tue 28 Jan
OC3MON also supports analysis by TCP/UDP application type. Figure 13 illustrates the proportion of traffic from Web servers using the well-known http port 80 (Web servers can also use other ports, whose traffic will not be reflected in the graph) measured in packets, bytes, and flows. Note that Web traffic consumes approximately the same proportion of flows as it does packets, but a somewhat larger proportion of bytes, indicating the use of larger packet sizes relative to other Internet traffic.
Figure 13: Proportion of Web server-to-client traffic, i.e., from port 80 to any port, measured in packets, bytes, and flows over 5-minute sample intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
Figure 14 plots flows in the opposite direction, from clients to Web servers; these flows have much lower byte proportions, being mostly query and acknowledgement traffic, slightly lower packet traffic, but similar flow counts.
Figure 14: Proportion of Web client-to-server traffic, i.e., to port 80 from any port, measured in packets, bytes, and flows over 5-minute sample intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
Domain name system (dns) traffic is also characterized by short query/response packets and thus, as shown in figure 15, comprises a huge proportion of (single packet, 40-80 byte) flows, but less than 3 percent of the byte traffic.
Figure 15: Proportion of dns traffic measured in packets, bytes, and flows over 5-minute sample intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
We can also look at the traffic by transport layer protocol; figure 16 is the proportion of UDP packets, bytes, and flows (which includes all of the dns traffic plotted in figure 12).
Figure 16: Proportion of udp traffic measured in packets, bytes, and flows over 5-minute sample intervals on MCI IP OC3 backbone trunk, Sat 25 Jan - Thur 30 Jan 97
Figure 17 shows the absolute counts of IPIP traffic, again measured in packets, bytes, and flows. In this case a 24-hour period is shown. IPIP (IP protocol 4) traffic includes Mbone tunnel traffic, where very few flows each typically consume a substantial proportion of packets and bytes.
Figure 17: Counts of IPIP (IP protocol 4) traffic measured in packets, bytes, and flows on MCI IP OC3 backbone trunk, Thur 30 Jan 97
Although each Mbone flow seems to consume an inordinate amount of resources, note that in the expected case, the Mbone flows represent tunneled multicast traffic, and thus potentially serve a larger number of customers than just the single flow depicts. In contrast, the cuseeme audio/video teleconferencing application, plotted in figure 18 with a profile similar to the Mbone flow profile, is not multicast, and so poses a definite threat to Internet service providers trying to grow, or even maintain, a (largely flat-priced) customer base.
Figure 18: Counts of cuseeme traffic measured in packets, bytes, and flows on MCI IP OC3 backbone trunk, Thur 30 Jan 97
We might also want to know an average of how many packets and bytes are in a flow of a given type; figure 19 shows this metric for cuseeme traffic.
Figure 19: Average packet size of cuseeme traffic on MCI IP OC3 backbone trunk, Thur 30 Jan 97
The use of ports as an application classifier limits us to applications that use a single port. Realaudio is an emerging application that uses more than one port: TCP port 7070 and UDP ports 6970 through 7170. Because we were particularly interested in the growth of this application, we modified the post-processing analysis script to support a query for this set of ports. (Note that this will be an upper bound, since other applications may also use these ports. In particular, afs uses port 7000.)
Figure 20: Counts of realaudio traffic (which uses a set of ports: TCP port 7070 and UDP ports 6970 through 7170) measured in packets, bytes, and flows on MCI IP OC3 backbone trunk, Thur 30 Jan 97
OC3MON's design is conducive to several extensions. Enhancements to the analysis of packet trace and flows data and the Web interface to the data are limited only by the imaginations of software developers. Our most immediate concerns are procuring OC12c interface cards for OC12MON, and as soon as possible being able to process both IP/ATM/SONET and IP/PPP/SONET encapsulations at OC12 rates with the same reasonably priced hardware.
We are also investigating moving more of the functionality of OCnMON, such as flows extraction from the packet header trace onto the interface card, in order to offload the host processor. This optimization would be increasingly useful at OC12c and OC48c speeds, where buses and host CPUs run out of steam. We believe that Field Programmable Gate Arrays (FPGAs) can provide this migration with a high degree of parallelism, without sacrificing the iterative design process and flexibility of software. We are also considering writing routines to access enhanced integrated drive electronics (EIDE) controller and the DMA engine on the Intel PCI ISA accelerator (PIIX3) directly to obtain much better asynchronous disk I/O.
Other extensions we examine with interest include:
We have described the design, implementation, and use of a high-performance yet affordable Internet monitoring tool. We have also described and shown examples from the Web-based interface to the associated library of post-processing analysis utilities for characterizing network usage and workload trends. By using low-cost, commodity hardware, we have ensured the practicality of using the monitor at a wide range of locations. Our network flow analysis tools have proven useful to us in understanding, verifying, debugging, or spotting traffic behavior anomalies in the locations where we have deployed it.
The Web-based query engine for vBNS OC3MON data is at http://www.vbns.net/stats/flows/html/level0/. An electronic HTML version of this paper and pointers to the OC3MON software are at http://www.nlanr.net/NA/Oc3mon/. The software itself is available via FTP from ftp://nlanr.net/Software/Oc3mon. The original prototype for the Web query engine, written by Hans-Werner Braun for the FIX-West FDDI, is still housed at http://www.nlanr.net/NA/.
It is being tested with:
Optical Splitters (available from ADC Telecommunications or AMP):
ADC also makes racks to mount 8 to 12 splitter modules (if you tap multiple links at same site)
We are grateful to Hans-Werner Braun for prototyping the original flow statistics software and making it freely available.
This material is based on work sponsored by the National Science Foundation, grants NCR-9415666 and NCR-9321047. The very high speed Backbone Network Service (vBNS) project is managed and coordinated by MCI Communications Corporation under sponsorship of the National Science Foundation. The government has certain rights to this material. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
 K. Claffy and H. W. Braun. "Post-NSFNET statistics collection." Proceedings INET'95.
 K. Claffy and H. W. Braun. Web traffic characterization: An assessment of the impact of caching documents from the NCSA's Web server. World Wide Web conference, 1994, Chicago, Illinois.
 K. Claffy, Hans-Werner Braun, and George C. Polyzos. A parameterizable methodology for Internet traffic flow profiling. IEEE JSAC.
 R. Jain and S. A. Routhier. "Packet trains: Measurement and a new model for computer network traffic." IEEE Journal on Selected Areas in Communications, vol. 4, no. 6, pp. 986-995, September 1986.
 J. Mogul. "Observing TCP dynamics in real networks." In Proceedings of ACM SIGCOMM '92, August 1992, pp. 305-317.
 M. Acharya, R. Newman-Wolfe, H. Latchman, R. Chow, and B. Bhalla. "Real-time hierarchical traffic characterization of a campus area network." In Proceedings of the Sixth International Conference on Modeling Techniques and Tools for Computer Performance Evaluation, 1992, University of Florida.
 M. Acharya and B. Bhalla. "A flow model for computer network traffic using real-time measurements." In Second International Conference on Telecommunications Systems, Modeling and Analysis, March 24-27, 1994.
 R. Caceres, P. Danzig, S. Jamin, and D. Mitzel. "Characteristics of wide-area TCP/IP conversations." In Proceedings of ACM SIGCOMM '91, September 1991, pp. 101-112.
 C. Partridge. A proposed flow specification, Internet Request for Comments Series RFC 1363, September 1992.
 D. Estrin and D. Mitzel. "An assessment of state and lookup overhead in routers." Proc. Infocom '92.