RFC1224
Network Working Group L. Steinberg Request for Comments: 1224 IBM Corporation
May 1991
Techniques for Managing Asynchronously Generated Alerts
Status of this Memo
This memo defines common mechanisms for managing asynchronously produced alerts in a manner consistent with current network management protocols.
This memo specifies an Experimental Protocol for the Internet community. Discussion and suggestions for improvement are requested. Please refer to the current edition of the "IAB Official Protocol Standards" for the standardization state and status of this protocol. Distribution of this memo is unlimited.
Abstract
This RFC explores mechanisms to prevent a remotely managed entity from burdening a manager or network with an unexpected amount of network management information, and to ensure delivery of "important" information. The focus is on controlling the flow of asynchronously generated information, and not how the information is generated.
(d) Reduced configuration requirements to manage remote entity... 4
5. Closed Loop "Feedback" Alert Reporting with a "Pin" Sliding
Contents
Introduction
This memo defines mechanisms to prevent a remotely managed entity from burdening a manager or network with an unexpected amount of network management information, and to ensure delivery of "important" information. The focus is on controlling the flow of asynchronously generated information, and not how the information is generated. Mechanisms for generating and controlling the generation of asynchronous information may involve protocol specific issues.
There are two understood mechanisms for transferring network management information from a managed entity to a manager: request- response driven polling, and the unsolicited sending of "alerts". Alerts are defined as any management information delivered to a manager that is not the result of a specific query. Advantages and disadvantages exist within each method. They are detailed in section 2 below.
Alerts in a failing system can be generated so rapidly that they adversely impact functioning resources. They may also fail to be delivered, and critical information maybe lost. Methods are needed both to limit the volume of alert transmission and to assist in delivering a minimum amount of information to a manager.
It is our belief that managed agents capable of asynchronously generating alerts should attempt to adopt mechanisms that fill both of these needs. For reasons shown in section 2.4, it is necessary to fulfill both alert-management requirements. A complete alert-driven system must ensure that alerts are delivered or their loss detected with a means to recreate the lost information, AND it must not allow itself to overburden its manager with an unreasonable amount of information.
Problem Definition
The following discusses the relative advantages and disadvantages of polled vs. alert driven management.
Polling Advantages
(a) Reliable detection of failures.
A manager that polls for all of its information can
more readily determine machine and network failures;
a lack of a response to a query indicates problems
with the machine or network. A manager relying on
notification of problems might assume that a faulty
system is good, should the alert be unable to reach
its destination, or the managed system be unable to
correctly generate the alert. Examples of this
include network failures (in which an isolated network
cannot deliver the alert), and power failures (in which
a failing machine cannot generate an alert). More
subtle forms of failure in the managed entity might
produce an incorrectly generated alert, or no alert at
all.
(b) Reduced protocol complexity on managed entity
The use of a request-response based system is based on
conservative assumptions about the underlying transport
protocol. Timeouts and retransmits (re-requests) can
be built into the manager. In addition, this allows
the manager to affect the amount of network management
information flowing across the network directly.
(c) Reduced performance impact on managed entity
In a purely polled system, there is no danger of having
to often test for an alert condition. This testing
takes CPU cycles away from the real mission of the
managed entity. Clearly, testing a threshold on each
packet received could have unwanted performance effects
on machines such as gateways. Those who wish to use
thresholds and alerts must choose the parameters to be
tested with great care, and should be strongly
discouraged from updating statistics and checking values
frequently.
(d) Reduced Configuration Requirements to manage remote
entity
Remote, managed entities need not be configured
with one or more destinations for reporting information.
Instead, the entity merely responds to whomever
makes a specific request. When changing the network
configuration, there is never a need to reconfigure
all remote manageable systems. In addition, any number
of "authorized" managers (i.e., those passing any
authentication tests imposed by the network management
protocol) may obtain information from any managed entity.
This occurs without reconfiguring the entity and
without reaching an entity-imposed limit on the maximum
number of potential managers.
Polling Disadvantages
(a) Response time for problem detection
Having to poll many MIB [2] variables per machine on
a large number of machines is itself a real
problem. The ability of a manager to monitor
such a system is limited; should a system fail
shortly after being polled there may be a significant
delay before it is polled again. During this time,
the manager must assume that a failing system is
acceptable. See Appendix A for a hypothetical
example of such a system.
It is worthwhile to note that while improving the mean
time to detect failures might not greatly improve the
time to correct the failure, the problem will generally
not be repaired until it is detected. In addition,
most network managers would prefer to at least detect
faults before network users start phoning in.
(b) Volume of network management traffic
Polling many objects (MIB variables) on many machines
greatly increases the amount of network management
traffic flowing across the network (see Appendix A).
While it is possible to minimize this through the use
of hierarchies (polling a machine for a general status
of all the machines it polls), this aggravates the
response time problem previously discussed.
Alert Advantages
(a) Real-time Knowledge of Problems
Allowing the manager to be notified of problems
eliminates the delay imposed by polling many objects/
systems in a loop.
(b) Minimal amount of Network Management Traffic
Alerts are transmitted only due to detected errors.
By removing the need to transfer large amounts of status
information that merely demonstrate a healthy system,
network and system (machine processor) resources may be
freed to accomplish their primary mission.
Alert Disadvantages
(a) Potential Loss of Critical Information
Alerts are most likely not to be delivered when the
managed entity fails (power supply fails) or the
network experiences problems (saturated or isolated).
It is important to remember that failing machines and
networks cannot be trusted to inform a manager that
they are failing.
(b) Potential to Over-inform the Manager
An "open loop" system in which the flow of alerts to
a manager is fully asynchronous can result in an excess
of alerts being delivered (e.g., link up/down messages
when lines vacillate). This information places an extra
burden on a strained network, and could prevent the
manager from disabling the mechanism generating the
alerts; all available network bandwidth into the manager
could be saturated with incoming alerts.
Most major network management systems strive to use an optimal combination of alerts and polling. Doing so preserves the advantages of each while eliminating the disadvantages of pure polling.
Specific Goals of this Memo
This memo suggests mechanisms to minimize the disadvantages of alert usage. An optimal system recognizes the potential problems associated with sending too many alerts in which a manager becomes ineffective at managing, and not adequately using alerts (especially given the volumes of data that must be actively monitored with poor scaling). It is the author's belief that this is best done by allowing alert mechanisms that "close down" automatically when over- delivering asynchronous (unexpected) alerts, and that also allow a flow of synchronous alert information through a polled log. The use of "feedback" (with a sliding window "pin") discussed in section 5 addresses the former need, while the discussion in section 6 on "polled, logged alerts" does the latter.
This memo does not attempt to define mechanisms for controlling the asynchronous generation of alerts, as such matters deal with specifics of the management protocol. In addition, no attempt is made to define what the content of an alert should be. The feedback mechanism does require the addition of a single alert type, but this is not meant to impact or influence the techniques for generating any other alert (and can itself be generated from a MIB object or the management protocol). To make any effective use of the alert mechanisms described in this memo, implementation of several MIB objects is required in the relevant managed systems. The location of these objects in the MIB is under an experimental subtree delegated to the Alert-Man working group of the Internet Engineering Task Force (IETF) and published in the "Assigned Numbers" RFC [5]. Currently, this subtree is defined as
alertMan ::= { experimental 24 }.
Compatibility With Existing Network Management Protocols
It is the intent of this document to suggest mechanisms that violate neither the letter nor the spirit of the protocols expressed in CMOT [3] and SNMP [4]. To achieve this goal, each mechanism described will give an example of its conformant use with both SNMP and CMOT.
Closed Loop "Feedback" Alert Reporting with a "Pin" Sliding
Window Limit
One technique for preventing an excess of alerts from being delivered involves required feedback to the managed agent. The name "feedback" describes a required positive response from a potentially "over- reported" manager, before a remote agent may continue transmitting alerts at a high rate. A sliding window "pin" threshold (so named for the metal on the end of a meter) is established as a part of a
user-defined SNMP trap, or as a managed CMOT event. This threshold defines the maximum allowable number of alerts ("maxAlertsPerTime") that may be transmitted by the agent, and the "windowTime" in seconds that alerts are tested against. Note that "maxAlertsPerTime" represents the sum total of all alerts generated by the agent, and is not duplicated for each type of alert that an agent might generate. Both "maxAlertsPerTime" and "windowTime" are required MIB objects of SMI [1] type INTEGER, must be readable, and may be writable should the implementation permit it.
Two other items are required for the feedback technique. The first is a Boolean MIB object (SMI type is INTEGER, but it is treated as a Boolean whose only value is zero, i.e., "FALSE") named "alertsEnabled", which must have read and write access. The second is a user defined alert named "alertsDisabled". Please see Appendix B for their complete definitions.
Use of Feedback
When an excess of alerts is being generated, as determined by the total number of alerts exceeding "maxAlertsPerTime" within "windowTime" seconds, the agent sets the Boolean value of "alertsEnabled" to "FALSE" and sends a single alert of type "alertsDisabled".
Again, the pin mechanism operates on the sum total of all alerts generated by the remote system. Feedback is implemented once per agent and not separately for each type of alert in each agent. While it is also possible to implement the Feedback/Pin technique on a per alert-type basis, such a discussion belongs in a document dealing with controlling the generation of individual alerts.
The typical use of feedback is detailed in the following steps:
(a) Upon initialization of the agent, the value of
"alertsEnabled" is set to "TRUE".
(b) Each time an alert is generated, the value of
"alertsEnabled" is tested. Should the value be "FALSE",
no alert is sent. If the value is "TRUE", the alert is
sent and the current time is stored locally.
(c) If at least "maxAlertsPerTime" have been generated, the
agent calculates the difference of time stored for the
new alert from the time associated with alert generated
"maxAlertsPerTime" previously. Should this amount be
less than "windowTime", a single alert of the type
"alertsDisabled" is sent to the manager and the value of
"alertsEnabled" is then set to "FALSE".
(d) When a manager receives an alert of the type "Alerts-
Disabled", it is expected to set "alertsEnabled" back
to "TRUE" to continue to receive alert reports.
Example
In a sample system, the maximum number of alerts any single managed entity may send the manager is 10 in any 3 second interval. A circular buffer with a maximum depth of 10 time of day elements is defined to accommodate statistics keeping.
After the first 10 alerts have been sent, the managed entity tests the time difference between its oldest and newest alerts. By testing the time for a fixed number of alerts, the system will never disable itself merely because a few alerts were transmitted back to back.
The mechanism will disable reporting only after at least 10 alerts have been sent, and the only if the last 10 all occurred within a 3 second interval. As alerts are sent over time, the list maintains data on the last 10 alerts only.
Notes on Feedback/Pin Usage
A manager may periodically poll "alertsEnabled" in case an "alertsDisabled" alert is not delivered by the network. Some implementers may also choose to add COUNTER MIB objects to show the total number of alerts transmitted and dropped by "alertsEnabled" being FALSE. While these may yield some indication of the number of lost alerts, the use of "Polled, Logged Alerts" offers a superset of this function.
Testing the alert frequency need not begin until a minimum number of alerts have been sent (the circular buffer is full). Even then, the actual test is the elapsed time to get a fixed number of alerts and not the number of alerts in a given time period. This eliminates the need for complex averaging schemes (keeping current alerts per second as a frequency and redetermining the current value based on the previous value and the time of a new alert). Also eliminated is the problem of two back to back alerts; they may indeed appear to be a large number of alerts per second, but the fact remains that there are only two alerts. This situation is unlikely to cause a problem for any manager, and should not trigger the mechanism.
Since alerts are supposed to be generated infrequently, maintaining the pin and testing the threshold should not impact normal performance of the agent (managed entity). While repeated testing
may affect performance when an excess of alerts are being transmitted, this effect would be minor compared to the cost of generating and sending so many alerts. Long before the cost of testing (in CPU cycles) becomes relatively high, the feedback mechanism should disable alert sending and affect savings both in alert sending and its own testing (note that the list maintenance and testing mechanisms disable themselves when they disable alert reporting). In addition, testing the value of "alertsEnabled" can limit the CPU burden of building alerts that do not need to be sent.
It is advised that the implementer consider allowing write access to both the window size and the number of alerts allowed in a window's time. In doing so, a management station has the option of varying these parameters remotely before setting "alertsEnabled" to "TRUE". Should either of these objects be set to 0, a conformant system will disable the pin and feedback mechanisms and allow the agent to send all of the alerts it generates.
While the feedback mechanism is not high in CPU utilization costs, those implementing alerts of any kind are again cautioned to exercise care that the alerts tested do not occur so frequently as to impact the performance of the agent's primary function.
The user may prefer to send alerts via TCP to help ensure delivery of the "alerts disabled" message, if available.
The feedback technique is effective for preventing the over-reporting of alerts to a manager. It does not assist with the problem of "under-reporting" (see "polled, logged alerts" for this).
It is possible to lose alerts while "alertsEnabled" is "FALSE". Ideally, the threshold of "maxAlertsPerTime" should be set sufficiently high that "alertsEnabled" is only set to "FALSE" during "over-reporting" situations. To help prevent alerts from possibly being lost when the threshold is exceeded, this method can be combined with "polled, logged alerts" (see below).
Polled, Logged Alerts
A simple system that combines the request-response advantages of polling while minimizing the disadvantages is "Polled, Logged Alerts". Through the addition of several MIB objects, one gains a system that minimizes network management traffic, lends itself to scaling, eliminates the reliance on delivery, and imposes no potential over-reporting problems inherent in pure alert driven architectures. Minimizing network management traffic is affected by reducing multiple requests to a single request. This technique does not eliminate the need for polling, but reduces the amount of data
transferred and ensures the manager either alert delivery or notification of an unreachable node. Note again, the goal is to address the needs of information (alert) flow and not to control the local generation of alerts.
Use of Polled, Logged Alerts
As alerts are generated by a remote managed entity, they are logged locally in a table. The manager may then poll a single MIB object to determine if any number of alerts have been generated. Each poll request returns a copy of an "unacknowledged" alert from the alert log, or an indication that the table is empty. Upon receipt, the manager might "acknowledge" any alert to remove it from the log. Entries in the table must be readable, and can optionally allow the user to remove them by writing to or deleting them.
This technique requires several additional MIB objects. The alert_log is a SEQUENCE OF logTable entries that must be readable, and can optionally have a mechanism to remove entries (e.g., SNMP set or CMOT delete). An optional read-only MIB object of type INTEGER, "maxLogTableEntries" gives the maximum number of log entries the system will support. Please see Appendix B for their complete definitions.
The typical use of Polled, Logged Alerts is detailed below.
(a) Upon initialization, the agent builds a pointer to a log
table. The table is empty (a sequence of zero entries).
(b) Each time a local alert is generated, a logTable entry
is built with the following information:
SEQUENCE {
alertId INTEGER,
alertData OPAQUE
}
(1) alertId number of type INTEGER, set to 1 greater
than the previously generated alertId. If this is
the first alert generated, the value is initialized
to 1. This value should wrap (reset) to 1 when it
reaches 2**32. Note that the maximum log depth
cannot exceed (2**32)-1 entries.
(2) a copy of the alert encapsulated in an OPAQUE.
(c) The new log element is added to the table. Should
addition of the element exceed the defined maximum log
table size, the oldest element in the table (having the
lowest alertId) is replaced by the new element.
(d) A manager may poll the managed agent for either the next
alert in the alert_table, or for a copy of the alert
associated with a specific alertId. A poll request must
indicate a specific alertId. The mechanism for obtaining
this information from a table is protocol specific, and
might use an SNMP GET or GET NEXT (with GET NEXT
following an instance of zero returning the first table
entry's alert) or CMOT's GET with scoping and filtering
to get alertData entries associated with alertId's
greater or less than a given instance.
(e) An alertData GET request from a manager must always be
responded to with a reply of the entire OPAQUE alert
(SNMP TRAP, CMOT EVENT, etc.) or a protocol specific
reply indicating that the get request failed.
Note that the actual contents of the alert string, and
the format of those contents, are protocol specific.
(f) Once an alert is logged in the local log, it is up to
the individual architecture and implementation whether
or not to also send a copy asynchronously to the
manager. Doing so could be used to redirect the focus
of the polling (rather than waiting an average of 1/2
the poll cycle to learn of a problem), but does not
result in significant problems should the alert fail to
be delivered.
(g) Should a manager request an alert with alertId of 0,
the reply shall be the appropriate protocol specific
error response.
(h) If a manager requests the alert immediately following
the alert with alertId equal to 0, the reply will be the
first alert (or alerts, depending on the protocol used)
in the alert log.
(i) A manager may remove a specific alert from the alert log
by naming the alertId of that alert and issuing a
protocol specific command (SET or DELETE). If no such
alert exists, the operation is said to have failed and
such failure is reported to the manager in a protocol
specific manner.
Example
In a sample system (based on the example in Appendix A), a manager must monitor 40 remote agents, each having between 2 and 15 parameters which indicate the relative health of the agent and the network. During normal monitoring, the manager is concerned only with fault detection. With an average poll request-response time of 5 seconds, the manager polls one MIB variable on each node. This involves one request and one reply packet of the format specified in the XYZ network management protocol. Each packet requires 120 bytes "on the wire" (requesting a single object, ASN.1 encoded, IP and UDP enveloped, and placed in an ethernet packet). This results in a serial poll cycle time of 3.3 minutes (40 nodes at 5 seconds each is 200 seconds), and a mean time to detect alert of slightly over 1.5 minutes. The total amount of data transferred during a 3.3 minute poll cycle is 9600 bytes (120 requests and 120 replies for each of 40 nodes). With such a small amount of network management traffic per minute, the poll rate might reasonably be doubled (assuming the network performance permits it). The result is 19200 bytes transferred per cycle, and a mean time to detect failure of under 1 minute. Parallel polling obviously yields similar improvements.
Should an alert be returned by a remote agent's log, the manager notifies the operator and removes the element from the alert log by setting it with SNMP or deleting it with CMOT. Normal alert detection procedures are then followed. Those SNMP implementers who prefer to not use SNMP SET for table entry deletes may always define their log as "read only". The fact that the manager made a single query (to the log) and was able to determine which, if any, objects merited special attention essentially means that the status of all alert capable objects was monitored with a single request.
Continuing the above example, should a remote entity fail to respond to two successive poll attempts, the operator is notified that the agent is not reachable. The operator may then choose (if so equipped) to contact the agent through an alternate path (such as serial line IP over a dial up modem). Upon establishing such a connection, the manager may then retrieve the contents of the alert log for a chronological map of the failure's alerts. Alerts undelivered because of conditions that may no longer be present are still available for analysis.
Notes on Polled, Logged Alerts
Polled, logged alert techniques allow the tracking of many alerts while actually monitoring only a single MIB object. This dramatically decreases the amount of network management data that must flow across the network to determine the status. By reducing
the number of requests needed to track multiple objects (to one), the poll cycle time is greatly improved. This allows a faster poll cycle (mean time to detect alert) with less overhead than would be caused by pure polling.
In addition, this technique scales well to large networks, as the concept of polling a single object to learn the status of many lends itself well to hierarchies. A proxy manager may be polled to learn if he has found any alerts in the logs of the agents he polls. Of course, this scaling does not save on the mean time to learn of an alert (the cycle times of the manager and the proxy manager must be considered), but the amount of network management polling traffic is concentrated at lower levels. Only a small amount of such traffic need be passed over the network's "backbone"; that is the traffic generated by the request-response from the manager to the proxy managers.
Note that it is best to return the oldest logged alert as the first table entry. This is the object most likely to be overwritten, and every attempt should be made ensure that the manager has seen it. In a system where log entries may be removed by the manager, the manager will probably wish to attempt to keep all remote alert logs empty to reduce the number of alerts dropped or overwritten. In any case, the order in which table entries are returned is a function of the table mechanism, and is implementation and/or protocol specific.
"Polled, logged alerts" offers all of the advantages inherent in polling (reliable detection of failures, reduced agent complexity with UDP, etc.), while minimizing the typical polling problems (potentially shorter poll cycle time and reduced network management traffic).
Finally, alerts are not lost when an agent is isolated from its manager. When a connection is reestablished, a history of conditions that may no longer be in effect is available to the manager. While not a part of this document, it is worthwhile to note that this same log architecture can be employed to archive alert and other information on remote hosts. However, such non-local storage is not sufficient to meet the reliability requirements of "polled, logged alerts".
Compatibility with SNMP [4] and CMOT [3]
Closed Loop (Feedback) Alert Reporting
Use of Feedback with SNMP
At configuration time, an SNMP agent supporting Feedback/Pin is loaded with default values of "windowTime" and "maxAlerts-PerTime", and "alertsEnabled" is set to TRUE. The manager issues an SNMP GET to determine "maxAlertsPerTime" and "windowTime", and to verify the state of "alertsEnabled". Should the agent support setting Pin objects, the manager may choose to alter these values (via an SNMP SET). The new values are calculated based upon known network resource limitations (e.g., the amount of packets the manager's gateway can support) and the number of agents potentially reporting to this manager.
Upon receipt of an "alertsDisabled" trap, a manager whose state and network are not overutilized immediately issues an SNMP SET to make "alertsEnabled" TRUE. Should an excessive number of "alertsDisabled" traps regularly occur, the manager might revisit the values chosen for implementing the Pin mechanism. Note that an overutilized system expects its manager to delay the resetting of "alertsEnabled".
As a part of each regular polling cycle, the manager includes a GET REQUEST for the value of "alertsEnabled". If this value is FALSE, it is SET to TRUE, and the potential loss of traps (while it was FALSE) is noted.
Use of Feedback with CMOT
The use of CMOT in implementing Feedback/Pin is essentially identical to the use of SNMP. CMOT GET, SET, and EVENT replace their SNMP counterparts.
Polled, Logged Alerts
Use of Polled, Logged alerts with SNMP
As a part of regular polling, an SNMP manager using Polled, logged alerts may issue a GET_NEXT Request naming { alertLog logTableEntry(1) alertId(1) 0 }. Returned is either the alertId of the first table entry or, if the table is empty, an SNMP reply whose object is the "lexicographical successor" to the alert log.
Should an "alertId" be returned, the manager issues an SNMP GET naming { alertLog logTableEntry(1) alertData(2) value } where "value"
is the alertId integer obtained from the previously described GET NEXT. This returns the SNMP TRAP encapsulated within an OPAQUE.
If the agent supports the deletion of table entries through SNMP SETS, the manager may then issue a SET of { alertLog logTableEntry(1) alertId(1) value } to remove the entry from the log. Otherwise, the next GET NEXT poll of this agent should request the first "alertId" following the instance of "value" rather than an instance of "0".
Use of Polled, Logged Alerts with CMOT
Using polled, logged alerts with CMOT is similar to using them with SNMP. In order to test for table entries, one uses a CMOT GET and specifies scoping to the alertLog. The request is for all table entries that have an alertId value greater than the last known alertId, or greater than zero if the table is normally kept empty by the manager. Should the agent support it, entries are removed with a CMOT DELETE, an object of alertLog.entry, and a distinguishing attribute of the alertId to remove.
Multiple Manager Environments
The conflicts between multiple managers with overlapping administrative domains (generally found in larger networks) tend to be resolved in protocol specific manners. This document has not addressed them. However, real world demands require alert management techniques to function in such environments.
Complex agents can clearly respond to different managers (or managers in different "communities") with different reply values. This allows feedback and polled, logged alerts to appear completely independent to differing autonomous regions (each region sees its own value). Differing feedback thresholds might exist, and feedback can be actively blocking alerts to one manager even after another manager has reenabled its own alert reporting. All of this is transparent to an SNMP user if based on communities, or each manager can work with a different copy of the relevant MIB objects. Those implementing CMOT might view these as multiple instances of the same feedback objects (and allow one manager to query the state of another's feedback mechanism).
The same holds true for polled, logged alerts. One manager (or manager in a single community/region) can delete an alert from its view without affecting the view of another region's managers.
Those preferring less complex agents will recognize the opportunity to instrument proxy management. Alerts might be distributed from a manager based alert exploder which effectively implements feedback
and polled, logged alerts for its subscribers. Feedback parameters are set on each agent to the highest rate of any subscriber, and limited by the distributor. Logged alerts are deleted from the view at the proxy manager, and truly deleted at the agent only when all subscribers have so requested, or immediately deleted at the agent with the first proxy request, and maintained as virtual entries by the proxy manager for the benefit of other subscribers.
Summary
While "polled, logged alerts" may be useful, they still have a limitation: the mean time to detect failures and alerts increases linearly as networks grow in size (hierarchies offer shorten individual poll cycle times, but the mean detection time is the sum of 1/2 of each cycle time). For this reason, it may be necessary to supplement asynchronous generation of alerts (and "polled, logged alerts") with unrequested transmission of the alerts on very large networks.
Whenever systems generate and asynchronously transmit alerts, the potential to overburden (over-inform) a management station exists. Mechanisms to protect a manager, such as the "Feedback/Pin" technique, risk losing potentially important information. Failure to implement asynchronous alerts increases the time for the manager to detect and react to a problem. Over-reporting may appear less critical (and likely) a problem than under-informing, but the potential for harm exists with unbounded alert generation.
An ideal management system will generate alerts to notify its management station (or stations) of error conditions. However, these alerts must be self limiting with required positive feedback. In addition, the manager should periodically poll to ensure connectivity to remote stations, and to retrieve copies of any alerts that were not delivered by the network.
10. References
[1] Rose, M., and K. McCloghrie, "Structure and Identification of
Management Information for TCP/IP-based Internets", RFC 1155, Performance Systems International and Hughes LAN Systems, May 1990.
[2] McCloghrie, K., and M. Rose, "Management Information Base for
Network Management of TCP/IP-based internets", RFC 1213, Hughes LAN Systems, Inc., Performance Systems International, March 1991.
[3] Warrier, U., Besaw, L., LaBarre, L., and B. Handspicker, "Common
Management Information Services and Protocols for the Internet
(CMOT) and (CMIP)", RFC 1189, Netlabs, Hewlett-Packard, The Mitre Corporation, Digital Equipment Corporation, October 1990.
[4] Case, J., Fedor, M., Schoffstall, M., and C. Davin, "Simple
Network Management Protocol" RFC 1157, SNMP Research, Performance Systems International, Performance Systems International, MIT Laboratory for Computer Science, May 1990.
[5] Reynolds, J., and J. Postel, "Assigned Numbers", RFC 1060,
USC/Information Sciences Institute, March 1990.
11. Acknowledgements
This memo is the product of work by the members of the IETF Alert-Man Working Group and other interested parties, whose efforts are gratefully acknowledged here:
Amatzia Ben-Artzi Synoptics Communications Neal Bierbaum Vitalink Corp. Jeff Case University of Tennessee at Knoxville John Cook Chipcom Corp. James Davin MIT Mark Fedor Performance Systems International, Inc. Steven Hunter Lawrence Livermore National Labs Frank Kastenholz Clearpoint Research Lee LaBarre Mitre Corp. Bruce Laird BBN, Inc Gary Malkin FTP Software, Inc. Keith McCloghrie Hughes Lan Systems David Niemi Contel Federal Systems Lee Oattes University of Toronto Joel Replogle NCSA Jim Sheridan IBM Corp. Steve Waldbusser Carnegie-Mellon University Dan Wintringham Ohio Supercomputer Center Rich Woundy IBM Corp.
Appendix A
Example of polling costs
The following example is completely hypothetical, and arbitrary. It assumes that a network manager has made decisions as to which systems, and which objects on each system, must be continuously monitored to determine the operational state of a network. It does not attempt to discuss how such decisions are made, and assumes that they were arrived at with the full understanding that the costs of polling many objects must be weighed against the
level of information required.
Consider a manager that must monitor 40 gateways and hosts on a single network. Further assume that the average managed entity has 10 MIB objects that must be watched to determine the device's and network's overall "health". Under the XYZ network management protocol, the manager may get the values of up to 4 MIB objects with a single request (so that 3 requests must be made to determine the status of a single entity). An average response time of 5 seconds is assumed, and a lack of response within 30 seconds is considered no reply. Two such "no replies" are needed to declare the managed entity "unreachable", as a single packet may occasionally be dropped in a UDP system (those preferring to use TCP for automated retransmits should assume a longer timeout value before declaring the entity "unreachable" which we will define as 60 seconds).
We begin with the case of "sequential polling". This is defined as awaiting a response to an outstanding request before issuing any further requests. In this example, the average XYZ network management protocol packet size is 300 bytes "on the wire" (requesting multiple objects, ASN.1 encoded, IP and UDP enveloped, and placed in an ethernet packet). 120 request packets are sent each cycle (3 for each of 40 nodes), and 120 response packets are expected. 72000 bytes (240 packets at 300 bytes each) must be transferred during each poll cycle, merely to determine that the network is fine.
At five seconds per transaction, it could take up to 10 minutes to determine the state of a failing machine (40 systems x 3 requests each x 5 seconds per request). The mean time to detect a system with errors is 1/2 of the poll cycle time, or 5 minutes. In a failing network, dropped packets (that must be timed out and resent) greatly increase the mean and worst case times to detect problems.
Note that the traffic costs could be substantially reduced by combining each set of three request/response packets in a single request/response transaction (see section 6.1.1 "Example").
While the bandwidth use is spread over 10 minutes (giving a usage of 120 bytes/second), this rapidly deteriorates should the manager decrease his poll cycle time to accommodate more machines or improve his mean time to fault detection. Conversely, increasing his delay between polls reduces traffic flow, but does so at the expense of time to detect problems.
Many network managers allow multiple poll requests to be "pending"
at any given time. It is assumed that such managers would not normally poll every machine without any delays. Allowing "parallel polling" and initiating a new request immediately following any response would tend to generate larger amounts of traffic; "parallel polling" here produces 40 times the amount of network traffic generated in the simplistic case of "sequential polling" (40 packets are sent and 40 replies received every 5 seconds, giving 80 packets x 300 bytes each per 5 seconds, or 4800 bytes/second). Mean time to detect errors drops, but at the cost of increased bandwidth. This does not improve the timeout value of over 2 minutes to detect that a node is not responding.
Even with parallel polling, increasing the device count (systems to manage) not only results in more traffic, but can degrade performance. On large networks the manager becomes bounded by the number of queries that can be built, tracked, responses parsed, and reacted to per second. The continuous volume requires the timeout value to be increased to accommodate responses that are still in transit or have been received and are queued awaiting processing. The only alternative is to reduce the poll cycle. Either of these actions increase both mean time to detect failure and worst case time to detect problems.
If alerts are sent in place of polling, mean time to fault detection drops from over a minute to as little as 2.5 seconds (1/2 the time for a single request-response transaction). This time may be increased slightly, depending on the nature of the problem. Typical network utilization is zero (assuming a "typical" case of a non-failing system).
Appendix B
All defined MIB objects used in this document reside
under the mib subtree:
alertMan ::= { iso(1) org(3) dod(6) internet(1)
experimental(3) alertMan(24) ver1(1) }
as defined in the Internet SMI [1] and the latest "Assigned
Numbers" RFC [5]. Objects under this branch are assigned
as follows:
RFC 1224-MIB DEFINITIONS ::= BEGIN
alertMan OBJECT IDENTIFIER ::= { experimental 24 }
ver1 OBJECT IDENTIFIER ::= { alertMan 1 }
feedback OBJECT IDENTIFIER ::= { ver1 1 }
polledLogged OBJECT IDENTIFIER ::= { ver1 2 }
END
1) Feedback Objects
OBJECT:
------
maxAlertsPerTime { feedback 1 }
Syntax:
Integer
Access:
read-write
Status:
mandatory
OBJECT:
------
windowTime { feedback 2 }
Syntax:
Integer
Access:
read-write
Status:
mandatory
OBJECT:
------
alertsEnabled { feedback 3 }
Syntax:
Integer
Access:
read-write
Status:
mandatory
2) Polled, Logged Objects
OBJECT:
------
alertLog { polledLogged 1 }
Syntax:
SEQUENCE OF logTableEntry
Access:
read-write
Status:
mandatory
OBJECT:
------
logTableEntry { alertLog 1 }
Syntax:
logTableEntry ::= SEQUENCE {
alertId
INTEGER,
alertData
OPAQUE
}
Access:
read-write
Status:
mandatory
OBJECT:
------
alertId { logTableEntry 1 }
Syntax:
Integer
Access:
read-write
Status:
mandatory
OBJECT:
------
alertData { logTableEntry 2 }
Syntax:
Opaque
Access:
read-only
Status:
mandatory
OBJECT:
------
maxLogTableEntries { polledLogged 2 }
Syntax:
Integer
Access:
read-only
Status:
optional
Security Considerations
Security issues are not discussed in this memo.
Author's Address
Lou Steinberg IBM NSFNET Software Development 472 Wheelers Farms Rd, m/s 91 Milford, Ct. 06460
Phone: 203-783-7175 EMail: [email protected]
