Network Working Group N. Brownlee Request for Comments: 2123 The University of Auckland Category: Informational March 1997
Traffic Flow Measurement: Experiences with NeTraMet
Status of this Memo
This memo provides information for the Internet community. This memo does not specify an Internet standard of any kind. Distribution of this memo is unlimited.
Abstract
This memo records experiences in implementing and using the Traffic Flow Measurement Architecture and Meter MIB. It discusses the implementation of NeTraMet (a traffic meter) and NeMaC (a combined manager and meter reader), considers the writing of meter rule sets and gives some guidance on setting up a traffic flow measurement system using NeTraMet.
Early in 1992 my University needed to develop a system for recovering the costs of its Internet traffic. In March of that year I attended the Internet Accounting Working Group's session at the San Diego IETF, where I was delighted to find that the Group had produced a detailed architecture for measuring network traffic and were waiting for someone to try implementing it.
During 1992 I produced a prototype measurement system, using balanced binary trees to store information about traffic flows. This work was reported at the Washington IETF in November 1992. The prototype performed well, but it made no attempt to recover memory from old flows, and the overheads in managing the balanced trees proved to be unacceptably high. I moved on to develop a production-quality system, this time using hash tables to index the flow information.
This version was called NeTraMet (the Network Traffic Meter), and was released as free software in October 1993. Since then I have continued working on NeTraMet, producing new releases two or three times each year. NeTraMet is now in production use at many sites around the world. It is difficult to estimate the number of sites, but there is an active NeTraMet mailing list, which had about 130 subscribers in March 1996.
Early in 1996 the Realtime Traffic Flow Measurement Working Group (RTFM) was chartered to move the Traffic Flow Measurement Architecture on to the IETF standards track. This document records traffic flow measurement experience gained through three years experience with NeTraMet.
The Traffic Flow Architecture document [1] describes four components:
- METERS, which are attached to the network at the points where it is desired to measure the traffic,
- METER READERS, which read data from meters and store it for later use,
- MANAGERS, which configure meters and control meter readers, and
- ANALYSIS APPLICATIONS, which process the data from meter readers so as to produce whatever reports are required.
NeTraMet is a computer program which implements the Traffic Meter, stores the measured flow data in memory, and provides an SNMP agent so as to make it available to Meter Readers. The NeTraMet distribution files include NeMaC, which is a combined Manager and Meter Reader capable of managing an arbitrary number of meters, each of which may be using its own rule set, and having its flow data collected at its own specified intervals. The NeTraMet distribution also includes several rudimentary Analysis Applications, allowing users to produce simple plots from NeMaC's flow data files (fd_filter and fd_extract) and to monitor - in real time - the flows at a remote meter (nm_rc and nifty).
Since the first release the Traffic Meter MIB [2] has been both improved and simplified. Significant changes have included better ways to specify traffic flows (i.e. more actions and better control structures for the Packet Matching Engine), and computed attributes (class and kind). These changes have been prompted by operational requirements at sites using NeTraMet, and have been tested extensively in successive versions of NeTraMet.
NeTraMet is widely used to collect usage data for Internet Service Providers. This is especially so in Australia and New Zealand, but there are also active users at sites around the world, for example in Canada, France, Germany and Poland.
NeTraMet is very useful as a tool for understanding exactly where traffic is flowing in large networks. Since the Traffic Meters perform considerable data reduction (as specified by their rule sets) they significantly reduce the volume of data to be read by Meter Readers. This characteristic makes NeTraMet particularly effective for networks with many remote sites. An example of this (the Kawaihiko network) is briefly described below.
Brownlee Informational [Page 3]
RFC 2123 Traffic Flow Measurement March 1997
As well as providing data for post-observation analysis, NeTraMet can be used for real-time network monitoring and trouble-shooting. The NeTraMet distribution includes 'nifty,' an X/Motif application which monitors traffic flows and attempts to highlight those which are 'interesting.'
As pointed out in the Architecture document [1], the goal of the Realtime Traffic Flow Measurement Working Group is to develop a standard for the Traffic Meter, with the goal of seeing it implemented in network devices such as hubs, switches and routers. Until the Architecture is well enough developed to allow this, it has sufficed to implement the meter as a program running on a general- purpose computer system.
The choice of computer system for NeTraMet was driven by the need to choose one which would be widely available within the Internet community. One strong possibility was a Unix system, since these are commonly used for a variety of network support and management tasks. For the initial implementation, however, Unix would have had some disadvantages:
- The wide variety of different Unix systems can increase the difficulties of software support.
- The cost of a Unix system as a meter is too high to allow users to run meters simultaneously at many points within their networks.
Brownlee Informational [Page 4]
RFC 2123 Traffic Flow Measurement March 1997
Another factor in choosing the platform was system performance. When I first started implementing NeTraMet it was impossible to predict how much processing workload was needed for a viable meter. Similarly, I had no idea how much memory would be required for code or data. I therefore chose to implement NeTraMet on a DOS PC. This was because:
- It is a minimum system in all respects. If the meter works well on such a system, it can be implemented on almost any hardware (including routers, switches, etc.)
- It is an inexpensive system. Sites can easily afford to have many meters around their networks.
- It is a simple system, and one which I had complete control over. This allowed me to implement effective instrumentation to monitor the meter's performance, and to include a wide variety of performance optimisations in the code.
Once the meter was running I needed a manager to download rule files to it. Since a single manager and meter reader can effectively support a large number of meters, a Unix environment for NeMaC was a natural choice. There are fewer software support problems for NeMaC than for NeTraMet since NeMaC has minimal support needs - it only needs to open a UDP socket to the SNMP port on each controlled meter.
Early NeTraMet distributions contained only the PC meter and Unix manager. In later releases I ported NeTraMet (the meter) to Unix, and extended the control features of NeMaC (the combined manager and meter reader). I have also experimented with porting NeMaC to the DOS system. This is not difficult, but doesn't seem to be worth pursuing.
The current version of NeTraMet is a production-quality traffic measurement system which has been in continuous use at the University of Auckland for nearly two years.
For the PC I chose to use Ethernet as the physical network medium. This is simply an initial choice, being the medium used within the University of Auckland's data network. Interfaces for other media could easily be added as they are needed.
In the PC environment a variety of 'generalised' network interfaces are available. I considered those available from companies such as Novell, DEC and Microsoft and decided against them, partly because they are proprietary, and partly because they did not appear to be particularly easy to use. Instead I chose the CRYNWR Packet Drivers [3] . These are available for a wide variety of interface cards and are simple and clearly documented. They support Ethernet's promiscuous mode, allowing one to observe headers for every passing packet in a straightforward manner. One disadvantage of the Packet Drivers is that it is harder to use them with newer user shells (such as Microsoft Windows), but this was irrelevant since I intended to run the meter as the only program on a dedicated machine.
Timing on the PC presented a challenge since the BIOS timer routines only provide a clock tick about 18 times each second, which limits the available time resolution. Initially I made do with a timing resolution of one second for packets, since I believed that most flows existed for many seconds. In recent years it has become apparent that many flows have lifetimes well under a second. To measure them properly with the Traffic Flow Meter one needs times resolved to 10 millisecond intervals, this being the size of TimeTicks, the most common time unit within SNMP [4]. Since all the details of the original PC are readily available [5], it was not difficult to understand the underlying hardware. I have written PC timer routines for NeTraMet which read the hardware timer with 256 times the resolution of the DOS clock ticks, i.e. about 5 ticks per millisecond.
There are many TCP/IP implementations available for DOS, but most of them are commercial software. Instead I chose Waterloo TCP [6], since this was available (including full source code) as public domain software. This was necessary since I needed to modify it to allow me to save incoming packet headers at the same time as forwarding packets destined for the meter to the IP handler routines. For SNMP I chose CMU SNMP [7], since again this was available (with full source code) as public domain software. This made it fairly simple to port it from Unix to the PC.
Brownlee Informational [Page 6]
RFC 2123 Traffic Flow Measurement March 1997
Finally, for the NeTraMet development I used Borland's Turbo C and Turbo Assembler. Although many newer C programming environments are now available, I have been perfectly happy with Turbo C version 2 for the NeTraMet project!
In implementing the Unix meter, the one obvious problem was 'how do I get access to packet headers?' Early versions of the Unix meter were implemented using various system-specific interfaces on a SunOS 4.2 system. Later versions use libpcap [8], which provides a portable method of obtaining access to packet headers on a wide range of Unix systems. I have verified that this works very well for ethernet interfaces on Solaris, SunOS, Irix, DEC Unix and Linux, and for FDDI interfaces on Solaris. libpcap provides timestamps for each packet header with resolution determined by the system clock, which is certainly better than 10 ms!
All Unix systems provide TCP/IP capabilities, so that was not an issue. For SNMP I used CMU SNMP, exactly as on the PC.
This section briefly discusses the data structures used by the meter, and the packet matching process. One very strong concern during the evolution of NeTraMet has been the need for the highest possible level of meter performance. A variety of interesting optimisations have been developed to achieve this; as discussed below. Another particular concern was the need for efficient and effective memory managent; this is discussed in detail below.
All the programs in NeTraMet, NeMaC and their supporting utility programs are written in C, partly because C and its run-time libraries provides good access to the underlying hardware, and partly because I have found it to be a highly portable language.
Brownlee Informational [Page 7]
RFC 2123 Traffic Flow Measurement March 1997
The data for each flow is stored in a C structure. The structure includes all the flow's attribute values (including packet and byte counts), together with a link field which can be used to link flows into lists. NeTraMet assumes that Adjacent addresses are 802 MAC Addresses, which are all six bytes long. Similarly, Transport addresses are assumes to be two bytes long, which is the case for port numbers in IP. Peer addresses are normally four bytes or less in length. They may, however, be as long as 20 bytes (for CLNS). I have chosen to use a fixed Peer address size, defined at compile time, so as to avoid the complexity of having variable-sized flow structures.
The flow table itself is an array of pointers to flow data structures, which allows indexed access to flows via their flow numbers. There is also a single large hash table, referred to in the Architecture document [1] as the flow table's 'search index'. Each hash value in the table points to a circular chain of flows. To find a flow one computes its hash value then searches that value's flow chain.
The meter stores each rule in a C structure. All the rule components have fixed sizes, but address fields must be wide enough to hold any type of address - Adjacent, Peer or Transport. The rule address width is defined at compile time, in the same way as flow Peer addresses. Each rule set is implemented as an array of pointers to rule data structures, and the rule table is an array of pointers to the rule sets. The size of each rule set is specified by NeMaC (before it begins downloading the rule set), but the maximum number of rule sets is defined at compile time.
Packet matching is carried out in NeTraMet exactly as described in the Architecture document [1]. Each incoming packet header is analysed so as to determine its attribute values. These values are stored in a structure which is passed to the Packet Matching Engine. To facilitate matching with source and destination reversed this structure contains two substructures, one containing the source Adjacent, Peer and Transport address values, the other containing the destination address values.
As described in the Architecture [1] each rule's address will usually be tested, i.e. ANDed with the rule's mask and compared with the rule's value. If the comparison fails, the next rule in sequence is executed. This allows one to write rule sets which use a group of rules to test an incoming packet to see whether one of its addresses
Brownlee Informational [Page 8]
RFC 2123 Traffic Flow Measurement March 1997
- e.g. its SourcePeerAddress - is one of a set of specified IP addresses. Such groups of related rules can grow quite large, containing hundreds of rules. It was clear that sequential execution of such groups of rules would be slow, and that something better was essential.
The optimisation implemented in NeTraMet is to find groups of rules which test the same attribute with the same mask, and convert them into a single hashed search of their values. The overhead of setting up hash tables (one for each group) is incurred once, just before the meter starts running a new rule set. When a 'group' test is to be performed, the meter ANDs the incoming attribute value, computes a hash value from it, and uses this to search the group's hash table. Early tests showed that the rule hash chains were usually very short, usually having only one or two members. The effect is to reduce large sequences of tests to a hash computation and lookup, with a very small number of compares; in short this is an essential optimisation for any traffic meter!
There is, of course, overhead associated with performing the hashed compare. NeTraMet handles this by having a minimum group size defined at compile time. If the group is too small it is not combined into a hashed group.
In early versions of NeTraMet I did not allow Gotos into a hashed group of rules, which proved to be an unnecessarily conservative position. NeTraMet stores each group's hash table in a separate memory area, and keeps a pointer to the hash table in the first rule of the group. (The rules data structure has an extra component to hold this hash table pointer). Rules within the group don't have hash table pointers; when they are executed as the target of a Goto rule they behave as ordinary rules, i.e. their tests are performed normally.
When the Packet Matching Engine has decided that an incoming packet belongs to a flow which is to be measured, it searches the flow table to determine whether or not the flow is already present. It does this by computing a hash from the packet and using it for access to the flow table's search index.
When designing a hash table, one normally assumes that the objects in the table have a constant size. For NeTraMet's flow table this would mean that each flow would contain a value for every attribute. This, however, is not the case, since only those attribute values 'pushed' by rules during packet matching are stored for a flow.
Brownlee Informational [Page 9]
RFC 2123 Traffic Flow Measurement March 1997
To demonstrate this problem , let us assume that every flow in the table contains a value for only one attribute, SourcePeerAddress, and that the rule set decides whether flows belong to a specified list of IP networks, in which case only their network numbers are pushed. The rules perform this test using a variety of masks, since the network number allocations range from 16 to 24 bits in width. In searching the flow table, the meter must distinguish between zeroes in the address and 'don't care' bits which had been ANDed out. To achieve this it must store SourcePeerMask values in the flow table as well as the ANDed SourcePeerAddress values.
In early versions of NeTraMet this problem was side-stepped by using multiple hash tables and relying on the user to write rules which used the same set of attributes and masks for all the flows in each table. This was effective, but clumsy and difficult to explain. Later versions changed to using a single hash table, and storing the mask values for all the address attributes in each flow.
The current version of the meter stores the address masks in compressed form. After examining a large number of rule sets I realised that although a rule set may have many rules, it usually has a very small number of address masks. It is a simple matter to build a table of address masks, and store an index to this 'mask table' instead of a complete mask. NeTraMet's maximum number of masks is defined at compile time, up to a maximum of 256. This allows me to use a single byte for each mask in the flow data structure, significantly reducing the structure's size. As well as this size reduction, two masks can be compared by comparing their indices in the mask table, i.e. it reduces to a single-byte comparison. Overall, using a mask table seems to provide useful improvements in storage efficiency and execution speed.
As described in the Architecture document [1], every incoming packet is tested against the current rule set by the Packet Matching Engine. This section explains my efforts to improve NeTraMet performance on the PC by reducing the amount of processing required by each incoming packet.
On the PC each incoming packet causes an interrupt, which NeTraMet must process so as to collect information about the packet. In early versions I used a ring buffer with 512 slots for packet headers, and simply copied each packet's first 64 bytes into the next free slot. The packet headers were later taken from the buffer, attribute values were extracted from them, and the resulting 'incoming attribute values' records were passed to the Packet Matching Engine.
Brownlee Informational [Page 10]
RFC 2123 Traffic Flow Measurement March 1997
I modified the interrupt handling code to extract the attribute values and store them in a 'buffer slot.' This reduced the amount of storage required in each slot, allowing more space for storing flows. It did increase slightly the amount of processing done for each packet interrupt, but this has not caused any problems.
In later versions I realised that if one is only interested in measuring IP packets, there is no point in storing (and later processing) Novell or EtherTalk packets! It is a simple matter for the meter to inspect a rule set and determine which Peer types are of interest. If there are PushRule rules which test SourcePeerType (or DestPeerType), they specify which types are of interest. If there are no such rules, every packet type is of interest. The PC NeTraMet has a set of Boolean variables, one for each protocol it can handle. The values of these 'protocol' variables are determined when the meter begins running a new rule set. For each incoming packet, the interrupt handler determines the Peer type. If the protocol is not of interest, no further processing is done - the packet is simply ignored. In a similar manner, if Adjacent addresses are never tested there is no point in copying them into the packet buffer slot.
The overall effect of these optimisations is most noticeable for rule files which measure IP flows on a network segment which also carries a lot of traffic for other network protocols; this situation is common on multiprotocol Local Area networks. On the Unix version of NeTraMet the Operating System does all the packet interrupt processing, and libpcap [8] delivers packet headers directly to NeTraMet. The 'protocol' and 'adjacent address' optimisations are still performed, at the point when NeTraMet receives the packet headers from libpcap.
The Architecture document [1] explains that a flow data record must be held in the meter until its data has been read by a meter reader. A meter must therefore have a reliable way of deciding when flow data has been read. The problem is complicated by the fact that there may be more than one meter reader, and that meter readers collect their data asynchronously.
Brownlee Informational [Page 11]
RFC 2123 Traffic Flow Measurement March 1997
Early versions of NeTraMet solved this problem by having a single MIB variable which a meter reader could set to indicate that it was beginning a data collection. In response to such an SNMP SET request, NeTraMet would update its 'collectors' table. This had an entry for each meter reader, and variables recording the start time for the last two collections. The most recent collection might still be in progress, but its start time provides a safe estimate of the time when the one before it actually finished. Space used for flows which have been idle since the penultimate collection started can be recovered by the meter's garbage collector, as described below.
The Meter MIB [2] specifies a more general table of meter reader information. A meter reader wishing to collect data from a meter must inform the meter of its intention by creating a row in the table, then setting a LastTime variable in that row to indicate the start of a collection. The meter handles such a SET request exactly as described above. If there are multiple meter readers the meter can easily find the earliest time any of them started its penultimate collection, and may recover flows idle since then. Should a meter reader fail, NeTraMet will eventually time out its entry in the meter reader info table, and delete it. This avoids a situation where the meter can't recover flows until they have been collected by several meter readers, one of which has failed.
In principle, the size of the flow table (i.e. the maximum number of flows) could be changed dynamically. This would involve allocating space for the flow table's new pointer array and copying the old pointers into it. NeTraMet does not implement this. Instead the maximum number of flows is set from the command line when it starts execution. If no maximum is specified, a compile-time default number is used.
Memory for flow data structures (i.e. 'flows') is allocated dynamically. NeTraMet requests the C run-time system for blocks of several hundred flows, and links them into a free list. When a new flow is needed NeTraMet gets memory space from the free list, then searches the flow table's pointer array for an unused flow pointer. In practice a 'last-allocated' index is used to point to the flow table, so a simple linear search suffices. The flow index is saved in the flow's data record, and its other attribute values are set to zero.
Brownlee Informational [Page 12]
RFC 2123 Traffic Flow Measurement March 1997
To release a flow data record it must first be removed from any hash list it is part of - this is straightforward since those lists are circular. The flow's entry in the flow table pointer array is then set to zero (NULL pointer), and its space is returned to the free list.
Once a flow data record is created it could continue to exist indefinitely. In time, however, the meter would run out of space. To deal with this problem NeTraMet uses an incremental garbage collector to reclaim memory.
At regular intervals specified by a 'GarbageCollectInterval' variable the garbage collector procedure is invoked. This searches through the flow table looking for flows which might be recovered. To control the resources consumed by garbage collection there are limits on the number of in-use and idle flows which the garbage collector may inspect these are set either when NeTraMet is started (as options on the command line) or dynamically by NeMaC (using variables in an Enterprise MIB for NeTraMet)
To decide whether a flow can be recovered, the garbage collector considers how long it has been idle (no packets in either direction), and when its data was last collected. If it has been collected by all known meter readers since its LastTime, its memory may be recovered. This alogrithm is implemented using a variable called 'GarbageCollectTime,' which normally contains the meter's UpTime when the penultimate collection (i.e. the one before last) was started. See the section on observing meter reader activity (above) for more details.
Should flows not be collected often enough the meter could run out of space. NeTraMet attempts to prevent this by having a low-priority background process check the percentage of flows active and compare it with the HighWaterMark MIB variable. If the percentage of active flows is greater than the high-water mark, 'GarbageCollectTime' is incremented by the current value of the InactivityTimeout MIB variable.
The Meter MIB [2] specifies that a meter should switch to using a 'standby' rule set if the percentage of active flows rises above HighWaterMark. In using NeTraMet to measure traffic flows to and from the University of Auckland it has not been difficult to create standby rules which are very similar to the 'production' rule file, differing only in that they push much less information about flows. This has, on several occasions, allowed the meter to continue running for one or two days after the meter reader failed. When the meter reader restarted, it was able to collect all the accumulated flow data!
Brownlee Informational [Page 13]
RFC 2123 Traffic Flow Measurement March 1997
The MIB also specifies that the meter should take some action when the active flow percentage rises above its FloodMark value. If this were not done, the meter could spend a rapidly increasing proportion of its time garbage collecting, to the point where its ability to respond to requests from its manager would be compromised. NeTraMet switches to the default rule set when its FloodMark is reached.
A potentially large number of new flows may be created when the meter switches to a standby rule set. It is important to set a HighWaterMark so as to allow enough flow table space for this. In practice, a HighWaterMark of 65% and a FloodMark of 95% seem to work well.
As explained above, a meter reader wishing to collect flows begins each collection by setting the LastTime variable in its ReaderInfoTable row, then works its way through the flow table collecting data. A number of algorithms can be used to examine the flow table; these are presented below.
The simplest approach is a linear scan of the table, reading the LastTime variable for each row. If the read fails the row is inactive. If it succeeds, it is of interest if its LastTime value is greater than the time of the last collection. Although this method is simple it is also rather slow, requiring an SNMP GET request for every possible flow; this renders it impractical.
Early versions of NeTraMet used two 'windows' into the flow table to find flows which were of interest. Both windows were SNMP tables, indexed by a variable which specified a time. A succession of GETNEXT requests on one of these windows allowed NeMaC (the meter reader) to find the flow indices for all flows which had been active since the specified time. The two windows were the ActivityTime window (which located active flows), and the CreateTime window (which located new flows). Knowing the index of an active flow, the meter reader can GET the values for all the attributes of interest. NeMaC allows the user to specify which these are, rather than simply read all the attributes.
Having the two windows allowed NeMaC to read attributes which remain constant - such as the flow's address attributes - when the flow is created, but to only read attributes which change with time - such as its packet and byte counts - during later collections. Experience has shown, however, that many flows have rather short lifetimes; one effect of this is that the improved efficiency of using two windows does not result in any worthwhile improvement in collection performance.
Brownlee Informational [Page 14]
RFC 2123 Traffic Flow Measurement March 1997
The current version of the Meter MIB [2] uses a TimeFilter variable in the flow table entries. This can be used with GETNEXT requests to find all flows which have been active since a specified time directly, without requiring the extra 'window' SNMP variables. It can be combined with SNMPv2's GETBULK request to further reduce the number of SNMP packets needed for each collection; I have yet to implement this in NeTraMet.
A disadvantage of using SNMP to collect data from the meter is that SNMP packets impose a high overhead. For example, if we wish to read an Integer32 variable (four bytes of data), it will be returned with its object identifier, type and length, i.e. at least ten bytes of superfluous data. One way to reduce this overhead is to use an Opaque object to return a collection of data. NeTraMet uses this approach to retrieve 'column activity data' from the meter, as follows.
Each packet of column activity data contains data values for a specified attribute, and each value is preceded by its flow number. The flow table can be regarded as a two-dimensional array, with a column for each flow attribute. Column activity data objects allow the meter reader to read columns of the flow table, so as to collect only those attributes specified by the user. The actual implementation is complicated by the fact that since the flow table is read column by column, rows can become active after the first column has been read. NeMaC reads the widest columns (those with greatest size in bytes, e.g. PeerAddress) first, and ignores any rows which appear in later columns. Newly active rows will, of course, be read in the next collection.
Using Opaque objects in this way dramatically reduces the number of SNMP packets required to read a meter. This has proved worthwhile in situations where the number of flows is large (for example on busy LANs), and where the meter(s) are physically dispersed over slow WAN links. It has the disadvantage that general-purpose MIB browsers cannot understand the column activity variables, but this seems a small price to pay for the improved data collection performance.
If a meter fails, for example because of a power failure, it will restart and begin running rule set 1, the default rule set which is built into the meter. Its manager must recognise that this has happened, and respond with some suitable action.
NeMaC allows the user to specify a 'keepalive' interval. After every such interval NeMaC reads the meter's sysUptime and compares it with the last sysUptime. If the new sysUptime is less than the last one,
Brownlee Informational [Page 15]
RFC 2123 Traffic Flow Measurement March 1997
NeMaC decides that the meter has restarted. It downloads the meter's backup rule set and production rule set, then requests the meter to start running the production rule set. In normal use we use a keepalive interval of five minutes and a collection interval of 15 minutes. If a meter restarts, we lose up to five minutes data before the rules sets are downloaded.
Having the meter run the default rule set on startup is part of the Traffic Flow Measurement Architecture [1], in keeping with the notion that meters are very simple devices which do not have disk storage. Since disks are now very cheap, it may be worth considering whether the architecture should allow a meter to save its configuration (including rule sets) on disk.
The PC version of the meter, NeTraMet, continually measures how much processor time is being used. Whenever there is no incoming packet data to process, 'dummy' packets are generated and placed in the input buffer. These packets are processed normally by the Packet Matching Engine; they have a PeerType of 'dummy.' The numbers of dummy and normal packets are counted by the meter; their ratio is used as an estimate of the processor time which is 'idle,' i.e. not being used to process incoming packets. The Unix version is intended to run as a process in a multiprocessing system, so it cannot busy- wait in this way.
The meter also collects several other performance measures; these can be displayed on the meter console in response to keyboard requests.
The PC meter can be used with a 10 MHz 286 machine, on which it can handle a steady load of about 750 packets per second. On a 25 MHz 386SX it will handle about 1250 packets per second. Users have reported that a 40 MHz 486 can handle peaks of about 3,000 packets per second without packet loss. The Unix meter has been tested metering traffic on a (lightly loaded) FDDI interface; it uses about one percent of the processor time on a SPARC 10 system running Solaris.
The Traffic Meter provides a versatile device for measuring a user- specified set of traffic flows, and performing useful data reduction on them. This data reduction capability not only minimises the volume of data to be collected by meter readers, but also simplifies the later processing of traffic flow data.
Brownlee Informational [Page 16]
RFC 2123 Traffic Flow Measurement March 1997
The flows of interest, and the processing to be performed, are specified in a 'rule set' which is downloaded to the meter (NeTraMet) by the manager (NeMaC). This section explains what is involved in writing rule sets.
NeTraMet is limited to metering packets observed on a network segment. This means that for all the observed flows, Source and Dest Type attributes (e.g. SourcePeerType and DestPeerType) have the same value.
The NeTraMet implementation uses single variables in its flow data structure for AdjacentType, SourceType and TransType. Nonetheless, the rule sets discussed below push values for both Source and Dest Type attributes; this make sure that packet matching works properly with the directions reversed, even for a meter which allows Source and Dest Type values to be different.
NeMaC reads rule sets from text files which contain the rules, the set number which the meter (and meter reader) will identify them by, and a 'format,' i.e. a list specifying which attributes the meter reader should collect and write to the flow data file. The # character indicates the start of a comment; NeMaC ignores the rest of the line.
The first rule tests the incoming packet's SourcePeerType to see whether it is 'dummy.' If it is, the packet is ignored, otherwise the next rule is executed.
The second rule tests the Null attribute. Such a test always succeeds, so the rule simply jumps to the action of the next rule. (The keyword 'next' is converted by NeMaC into the number of the following rule.)
The third rule pushes the packet's SourcePeerType value, then jumps to the action of the next rule. The user does not know in advance what the value of PushPkt rules will be, which is why the value appearing in them is always zero. The user must take care not to write rule sets which try to perform the test in a PushPkt rule. This is a very common error in a rule set, so NeMaC tests for it and displays an error message.
The following rules push a series of attribute values from the packet, and the last rule also Counts the packet, i.e. it tells the Packet Matching Engine (PME) that the packet has been successfully matched. The PME responds by searching the flow table to see whether the flow is already current (i.e. in the table), creating a new flow data record for it should this be necessary, and incrementing its packet and byte counters.
Overall this rule set simply classifies the packet (i.e. decides whether or not it is to be counted), then pushes all the Peer and Transport attribute values for it. It makes no attempt to specify a direction for the flow - this is left to the PME, as described in [1]. The resulting flow data file will have each flow's source and destination addresses in the order of the first packet the meter observed for the flow.
3.2 Specifying flow direction, using computed attributes
As indicated above, the Packet Matching Engine will reliably determine the flow, and the direction within that flow, for every packet seen by a meter. If the rule set does not specify a direction for the flow, the PME simply assumes that the first packet observed for a flow is travelling forward, i.e. from source to destination. In later analysis of the flow data, however, one is usually interested in traffic to or from a particular source.
One can achieve this in a simple manner by writing a rule set to specify the source for flows. All that is required is to have rules which succeed if the packet is travelling in the required direction, and which execute a 'Fail' action otherwise. This is demonstrated in the following two examples.
Brownlee Informational [Page 18]
RFC 2123 Traffic Flow Measurement March 1997
(Note that early versions of NeMaC allowed 'Retry' as a synonym for 'Fail.' The current version also allows 'NoMatch,' which seems a better way to imply "fail, allowing PME to try a second match with directions reversed.")
The rule labelled ip_pkt tests whether the packet came from network 130.216. If it did not, the test fails and the following rule executes a NoMatch action, causing the PME to retry the match with the directions reversed. If the second match fails the packet did not have 130.216 as an end-point, and is ignored.
The next rule set meters IP traffic on a network segment which connects two routers, g1 and g2. It classifies flows into three groups - those travelling from g1 to g2, those whose source is g1 and those whose source is g2.
The first two rules ignore non-IP packets. The next two rules Goto s1 if the packet's source was g1, or to s2 otherwise. The rule labelled s2 tests whether the packet's source was g2; if not a NoMatch action is executed, allowing the PME to try the match with the packet's direction reversed. If the match fails on the second try the packet didn't come from (or go to) g1 or g2, and is ignored.
Brownlee Informational [Page 20]
RFC 2123 Traffic Flow Measurement March 1997
Packets which come from g1 are tested by the rule labelled s1, and the PME will Goto either g3 or g1.
Packets which came from g2 are tested by the rule labelled s3. If they are not going to g1 the PME will Goto g2. If they are going to g1 a NoMatch action is executed - we want them counted as backward- travelling packets for the g1-g2 flow.
The rules at g1, g2 and g3 push the value 1, 2 or 3 from their rule into the flow's FlowClass attribute. This value can be used by an Analysis Application to separate the flows into the three groups of interest. FlowClass is an example of a 'computed' attribute, i.e. one whose value is Pushed by the PME during rule matching.
The remaining rules Push the values of other attributes required for later analysis, then Count the flow.
Subroutines are implemented in the PME in much the same way as in BASIC. A subroutine body is just a sequence of statements, supported by the GoSub and Return actions. 'GoSub' saves the PME's running environment and jumps to the first rule of the subroutine body. Subroutine calls may be nested as required - NeTraMet defines the maximum nesting at compile time. 'Return n' restores the environment and jumps to the action part of the nth rule after the Gosub, where n is the index value from the Return rule.
The Return action provides a way of influencing the flow of control in a rule set, rather like a FORTRAN Computed Goto. This is one way in which a subroutine can return a result. The other way is by Pushing a value in either a computed attribute (as demonstrated in the preceding section), or in a flow attribute.
One common use for a subroutine is to test whether a packet attribute matches one of a set of values. Such a subroutine becomes much more useful if it can be used to test one of several attributes. The PME architecture provides for this by using 'meter variables' to hold the names of the attributes to be tested. The meter variables are called V1, V2, V3, V4 and V5, and the Assign action is provided to set their values. If, for example, we need a subroutine to test either SourcePeerAddress or DestPeerAddress, we write its rules to test V1 instead. Before calling the subroutine we Assign SourcePeerAddress to V1; later tests of V1 are converted by the PME into tests on SourcePeerAddress. Note that since meter variables may be reassigned in a subroutine, their values are part of the environment which must be saved by a Gosub action.
Brownlee Informational [Page 21]
RFC 2123 Traffic Flow Measurement March 1997
The following rule set demonstrates the use of a subroutine ..
The subroutine begins at the rule labelled classify (shown below). It returns to the first, second or third rule after the invoking Gosub rule, depending on whether the tested PeerAddress is in the UA, AIT, or 'other' group of networks. In the listing below only one network is tested in each of the groups - it is trivial to add more rules (one per network) into either of the first two groups. In this example the subroutine Pushes the network number from the packet into the tested attribute before returning.
Brownlee Informational [Page 22]
RFC 2123 Traffic Flow Measurement March 1997
The first invocation of classify (above) begins at the rule labelled ip_pkt. It Assigns SourcePeerAddress to V1 then executes a Gosub action. Classify returns to one of the three following rules. They will Goto from_ua or from_ait if the packet came from the UA or AIT groups, otherwise the PME will retry the match. This means that matched flows will have a UA or AIT network as their source, and flows between other networks will be ignored.
The next two invocations of 'classify' test the packet's DestPeerAddress. Packets from AIT to UA are Retried, forcing them to be counted as AU to AIT flows. Packets from UA to UA are ignored, as are packets from AIT to AIT.
The next example demonstrates a way of grouping IP flows together depending on their Transport Address, i.e. their IP port number. Simply Pushing every flow's SourceTransAddress and DestTransAddress would produce a large number of flows, most of which differ only in one of their transport addresses (the one which is not a well-known port).
Instead we Push the well-known port number into each flow's SourceTransAddress; its DestTransAddress will be zero by default.
The first few rules ignore dummy packets, select IP packets for further processing, and count packets for other protocols in a single flow for each PeerType. TCP and UDP packets cause the PME to Push their TransType and Goto tcp_udp. ICMP and OSPF packets are counted in flows which have only their TransType Pushed.
Brownlee Informational [Page 24]
RFC 2123 Traffic Flow Measurement March 1997
At tcp_udp the packets' SourceTransAddress is tested to see whether it is included in a set of 'interesting' port numbers. If it is, the port number is pushed from the rule into the SourceTransAddress attribute, and the packet is counted at c_well_known. (NeMaC accepts Pushto as a synonym for PushRuleto).
This testing is repeated for the packet's DestTransAddress; if one of these tests succeeds the PME Goes to the corresponding rule above and Pushes the port number into the flow's SourceTransAddress. If these tests fail the packet is counted at c_unknown, where all the flow's Trans attributes are pushed. For production use more well-known ports would need to be included in the tests above - c_unknown is intended only for little-used exception flows!
Note that these rules only Push a value into a flow's SourceTransAddress, and they don't contain any NoMatch actions. They therefore don't specify a packet's direction, and they could be used in other rule sets to group together flows for well-known ports.
The last example (below) meters flows from a remote router, and demonstrates another approach to grouping well-known ports.
The first group of rules test incoming packet's AdjacentAddresses to see whether they belong to a flow with an end point at the specified router. Any which don't are ignored. Non-IP packets are counted in flows which only have their PeerType Pushed; these will produce one flow for each non-IP protocol. IP packets with TransTypes other than UDP and TCP are counted at not_wkp, where all their address attributes are pushed.
The high-order six bits of SourceTransAddress for UDP and TCP packets are compared with zero. If this succeeds their source port number is less than 1024, so they are from a well-known port. The port number is pushed from the rule into the flow's SourceTransAddress attribute, and the packet is counted at well_known_port. If the test fails, it is repeated on the packet's DestTransAddress. If the destination is a well-known port the match is Retried, and will succeed with the well-known port as the flow's source.
If later analysis were to show that a high proportion of the observed flows were from non-well-known ports, further pairs of rules could be added to perform a test in each direction for other heavily-used ports.
Although the Architecture document [1] specifies - in great detail - how the Traffic Flow Meter works, and how a meter reader should collect flow data from a meter, it does not say anything about how the collected data should be stored. NeMaC uses a simple, self- documenting file format, which has proved to be very effective in use.
There are two kinds of records in a flow data file: flow records and information records. Each flow record is simply a sequence of attribute values with separators (these can be specified in a NeMaC rule file) or spaces between them, terminated by a newline.
Information records all start with a cross-hatch. The file's first record begins with ##, and identifies the file as being a file of data from NeTraMet. It records NeMaC's parameters and the time this collection was started. The file's second record begins with #Format: and is a copy of the Format statement used by NeMaC to collect the data.
Brownlee Informational [Page 26]
RFC 2123 Traffic Flow Measurement March 1997
The rest of the file is a sequence of collected data sets. Each of these starts with a #Time: record, giving the time-of-day the collection was started, the meter name, and the range of meter times this collection represents. These from and to times are meter UpTimes, i.e. they are times in hundredths of seconds since the meter commenced operation. Most analysis applications have simply used the collection start times (which are ASCII time-of-day values), but the from and to times could be used to convert Uptime values to time-of- day. The flow records which comprise a data set follow the #Time record.
Several features of NeMaC's flow data files (as indicated above) are worthy of note:
- Collection times overlap slightly between samples. This allows for flows which were created after the collection started, and makes sure that flows are not missed from a collection.
Brownlee Informational [Page 28]
RFC 2123 Traffic Flow Measurement March 1997
- The rule set may change during a run. The above shows flows from rule set 1 - the default set - in the first collection, followed by the first flows created by rule set 3 (which has just been downloaded by NeMaC).
- FlowIndexes may be reused by the meter once their flows have been recovered by the garbage collector. The combination of FlowRuleSet, FlowIndex and StartTime are needed to identify a flow uniquely.
- Packet and Byte counters are 32-bit unsigned integers, and are never reset by the meter. Computing the counts occurring within a collection interval requires taking the difference between the collected count and its value when the flow was last collected. Note that counter wrap-around can be allowed for by simply performing an unsigned subtraction and ignoring any carry.
- In the sample flow data file above I have used double spaces as separators between the flow identifiers, peer addresses, pdu counts and packet counts.
- The format of addresses in the flow data file depends on the type of address. NeMaC always displays Adjacent addresses as six hex bytes separated by hyphens, and Transport addresses as (16-bit) integers. The format of a Peer address depends on its PeerType, e.g. dotted decimal for IP. To facilitate this NeMaC needs to know the PeerType for each flow; the user must request NeMaC to collect it.
When NeMaC first starts collecting from a meter, it reads the flow data for all active flows. This provides a starting point for analysis applications to compute the counts between successive collections.
From time to time the user needs to terminate a flow data file and begin a new one. For example, a user might need to generate a separate file for each day of metering. NeMaC provides for this by closing the file after each collection, then opening it and appending the data from the next collection. To terminate a file the user simply renames it. The Unix system will effect the name change either immediately (if the file was closed) or as soon as the current collection is complete (and the file is closed).
When NeMaC begins its next collection it observes that the file has disappeared, so it creates a new one and writes the # header records before writing the collected data.
Brownlee Informational [Page 29]
RFC 2123 Traffic Flow Measurement March 1997
There is one aspect of the above which requires some care on the user's part. The last data set in a file is not duplicated as the first data set of the next file. In other words, analysis applications must either look ahead at the first data set of the next file, or begin by reading the last data set of the previous file. If they fail to do this they will loose one collection's worth of flow data at each change of file.
Most analysis applications will be unique, taking data produced by a locally-developed rule set and producing reports to satisfy specific local requirements. The NeTraMet distribution files include three applications which are of general use, as follows:
- fd_filter computes data rates, i.e. the differences between successive data sets in a flow data file. It also allows the user to assign a 'tag' number to each flow; these are 'computed' attributes similar to FlowClass and FlowKind - the only difference is that they are computed from the collected data sets.
- fd_extract takes 'tagged' files from fd_filter and produces simple 'column list' files for use by other programs. One common use for fd_extract is to produce time-series data files which can be plotted by utilities like GNUPlot.
- nm_rc is a 'remote console' for a NeTraMet meter. It is a slightly simplified version of NeMaC combined with fd_filter. It can be used to monitor any meter, and will display (as lines of text characters) information about the n busiest flows observed during each collection interval.
- nifty is a traffic flow analyser, which (like nm_rc) displays data from a NeTraMet meter. nifty is an X/Motif application, which produces displays like 'Packet rate (pps) vs Flow lifetime (minutes),' so as to highlight those flows which are 'interesting.'
These applications are useful in themselves, and they provide a good starting point for users who wish to write their own analysis applications.
This section gives a brief summary of the steps involved in setting up a traffic measurement system using NeTraMet. These are:
- Decide what is to be measured. One good way to approach this is to specify exactly which flows are to be measured, and what reports will be required. Specifying the flows should make it obvious where meters will have to be placed so that the flows can be observed, whether PCs will be adequate for the task, etc..
- Install meters. As well as actually placing the meter hosts this includes making sure that they are configured correctly, with appropriate IP addresses, SNMP community strings, etc.
- Develop the rule set (and a standby rule set). The degree of difficulty here depends on how much is known in advance about the traffic. One possible approach is to start with the meter default rule set and measure how much traffic there is for each PeerType. (This is a good way to verify that NeTraMet and NeMaC are working properly). You can now add rules so as to increase the granularity of the flows; this will of course increase the number of flows to be collected, and force the meter's garbage collector to work harder. Another approach is to try a rule set with very fine granularity (i.e. one which Pushes all the address attributes), then observing how many flows are collected every few minutes.
- Develop a strategy for controlling meter reader. This means setting the meter's maximum number of flows, the collection interval, how breaks between flow data files will be handled, how often NeMaC should check that the meter is running, etc.
- Develop application(s) to process the collected flow data and produce the required files and reports.
- Test run. Monitor the system, then refine the rule sets and meter reading strategy until the overall system performance is satisfactory.
This process can take quite a long time, but the overall result is well worth the effort.
At the University of Auckland we run two sets of meters. One of these measures the traffic entering and leaving our University network, and generates usage reports for all our Internet users. This has been in production since early 1994.
Brownlee Informational [Page 31]
RFC 2123 Traffic Flow Measurement March 1997
The other set consists of meters which are distributed at Universities throughout New Zealand. They provide continuous traffic flow measurements at five-minute intervals for all the links making up the Universities' network (Kawaihiko); this system has been in production since January 1996, and has already proved very useful in planning the network's development.
The Kawaihiko Network provides IP connectivity for the New Zealand Universities. They are linked via a Frame Relay cloud, using a partial mesh of permanent virtual circuits. There is a NeTraMet meter at each site, metering inward and outward traffic. All the meters are managed from Auckland, and they all run copies of the same rule set.
The rule set has about 650 rules, most of which are in a single subroutine which classifies PeerAddresses into three categories - 'Kawaihiko network,' 'other New Zealand network' and 'non-New Zealand network.' Inside New Zealand IP addresses lie within six CIDR blocks, and there are about four hundred older networks which have addresses outside those blocks. The rules are arranged in groups by subnet size, i.e. all the /24 networks are tested first, then the /23 networks, etc, finishing with the /16 networks. This means that although there are about 600 networks, any PeerAddress can be classified with only nine tests.
The Kawaihiko rule set classifies flows, using computed attributes to indicate the network 'kind' (Kawaihiko / New Zealand / international) for each flow's SourcePeerAddress and DestPeerAddress, and to indicate whether the flow is a 'network news' flow or not.
Flow data is collected from all of the meters every five minutes, and is used to produce weekly reports, as follows:
- Traffic Plots. Plots of the 5-minute traffic rates for each site, showing international traffic in and out, news traffic in and out, and total traffic in and out of the site.
- Traffic Matrices. Two of these are produced, one for news traffic, the other for total traffic. They show the traffic rates from every site (including 'other New Zealand' and 'international') to every other site. The mean, third quartile and maximum are printed for every cell in the matrices.
This memo documents the implementation work on traffic flow measurement here at the University of Auckland. Many of my University colleagues have contributed significantly to this work, especially Russell Fulton (who developed the rules sets, Perl scripts and Cron jobs which produce our traffic usage reports automatically week after week) and John White (for his patient help in documenting the project).
[1] Brownlee, N., Mills, C., and G. Ruth, "Traffic Flow Measurement: Architecture", RFC 2063, The University of Auckland, Bolt Beranek and Newman Inc., GTE Laboratories, Inc, January 1997.
[2] Brownlee, N., "Traffic Flow Measurement: Meter MIB", RFC 2064, The University of Auckland, January 1997.
[4] Case J., McCloghrie K., Rose M., and Waldbusser S., "Structure of Management Information for version 2 of the Simple Network Managemenet Protocol", RFC 1902, SNMP Research Inc., Hughes LAN Systems, Dover Beach Consulting, Carnegie Mellon University, April 1993.
[5] IBM Corporation, "IBM PC Technical Reference Manual," 1984.
Security issues are not discussed in detail in this document. The meter's management and collection protocols are responsible for providing sufficient data integrity and confidentiality.