Postfix Bottleneck Analysis


Purpose of this document

This document is an introduction to Postfix queue congestion analysis. It explains how the qshape(1) program can help to track down the reason for queue congestion. qshape(1) is bundled with Postfix 2.1 and later source code, under the "auxiliary" directory. This document describes qshape(1) as bundled with Postfix 2.4.

This document covers the following topics:

Introducing the qshape tool

When mail is draining slowly or the queue is unexpectedly large, run qshape(1) as the super-user (root) to help zero in on the problem. The qshape(1) program displays a tabular view of the Postfix queue contents.

For example, in the output below we see the top 10 lines of the (mostly forged) sender domain distribution for captured spam in the "hold" queue:

$ qshape -s hold | head
                         T  5 10 20 40 80 160 320 640 1280 1280+
                 TOTAL 486  0  0  1  0  0   2   4  20   40   419
             yahoo.com  14  0  0  1  0  0   0   0   1    0    12
  extremepricecuts.net  13  0  0  0  0  0   0   0   2    0    11
        ms35.hinet.net  12  0  0  0  0  0   0   0   0    1    11
      winnersdaily.net  12  0  0  0  0  0   0   0   2    0    10
           hotmail.com  11  0  0  0  0  0   0   0   0    1    10
           worldnet.fr   6  0  0  0  0  0   0   0   0    0     6
        ms41.hinet.net   6  0  0  0  0  0   0   0   0    0     6
                osn.de   5  0  0  0  0  0   1   0   0    0     4

When the output is a terminal intermediate results showing the top 20 domains (-n option) are displayed after every 1000 messages (-N option) and the final output also shows only the top 20 domains. This makes qshape useful even when the deferred queue is very large and it may otherwise take prohibitively long to read the entire deferred queue.

By default, qshape shows statistics for the union of both the incoming and active queues which are the most relevant queues to look at when analyzing performance.

One can request an alternate list of queues:

$ qshape deferred
$ qshape incoming active deferred

this will show the age distribution of the deferred queue or the union of the incoming active and deferred queues.

Command line options control the number of display "buckets", the age limit for the smallest bucket, display of parent domain counts and so on. The "-h" option outputs a summary of the available switches.

Trouble shooting with qshape

Large numbers in the qshape output represent a large number of messages that are destined to (or alleged to come from) a particular domain. It should be possible to tell at a glance which domains dominate the queue sender or recipient counts, approximately when a burst of mail started, and when it stopped.

The problem destinations or sender domains appear near the top left corner of the output table. Remember that the active queue can accommodate up to 20000 ($qmgr_message_active_limit) messages. To check whether this limit has been reached, use:

$ qshape -s active       (show sender statistics)

If the total sender count is below 20000 the active queue is not yet saturated, any high volume sender domains show near the top of the output.

With oqmgr(8) the active queue is also limited to at most 20000 recipient addresses ($qmgr_message_recipient_limit). To check for exhaustion of this limit use:

$ qshape active          (show recipient statistics)

Having found the high volume domains, it is often useful to search the logs for recent messages pertaining to the domains in question.

# Find deliveries to example.com
#
$ tail -10000 /var/log/maillog |
        egrep -i ': to=<.*@example\.com>,' |
        less

# Find messages from example.com
#
$ tail -10000 /var/log/maillog |
        egrep -i ': from=<.*@example\.com>,' |
        less

You may want to drill in on some specific queue ids:

# Find all messages for a specific queue id.
#
$ tail -10000 /var/log/maillog | egrep ': 2B2173FF68: '

Also look for queue manager warning messages in the log. These warnings can suggest strategies to reduce congestion.

$ egrep 'qmgr.*(panic|fatal|error|warning):' /var/log/maillog

When all else fails try the Postfix mailing list for help, but please don't forget to include the top 10 or 20 lines of qshape(1) output.

Example 1: Healthy queue

When looking at just the incoming and active queues, under normal conditions (no congestion) the incoming and active queues are nearly empty. Mail leaves the system almost as quickly as it comes in or is deferred without congestion in the active queue.

$ qshape        (show incoming and active queue status)

                 T  5 10 20 40 80 160 320 640 1280 1280+
          TOTAL  5  0  0  0  1  0   0   0   1    1     2
  meri.uwasa.fi  5  0  0  0  1  0   0   0   1    1     2

If one looks at the two queues separately, the incoming queue is empty or perhaps briefly has one or two messages, while the active queue holds more messages and for a somewhat longer time:

$ qshape incoming

                 T  5 10 20 40 80 160 320 640 1280 1280+
          TOTAL  0  0  0  0  0  0   0   0   0    0     0

$ qshape active

                 T  5 10 20 40 80 160 320 640 1280 1280+
          TOTAL  5  0  0  0  1  0   0   0   1    1     2
  meri.uwasa.fi  5  0  0  0  1  0   0   0   1    1     2

Example 2: Deferred queue full of dictionary attack bounces

This is from a server where recipient validation is not yet available for some of the hosted domains. Dictionary attacks on the unvalidated domains result in bounce backscatter. The bounces dominate the queue, but with proper tuning they do not saturate the incoming or active queues. The high volume of deferred mail is not a direct cause for alarm.

$ qshape deferred | head

                         T  5 10 20 40 80 160 320 640 1280 1280+
                TOTAL 2234  4  2  5  9 31  57 108 201  464  1353
  heyhihellothere.com  207  0  0  1  1  6   6   8  25   68    92
  pleazerzoneprod.com  105  0  0  0  0  0   0   0   5   44    56
       groups.msn.com   63  2  1  2  4  4  14  14  14    8     0
    orion.toppoint.de   49  0  0  0  1  0   2   4   3   16    23
          kali.com.cn   46  0  0  0  0  1   0   2   6   12    25
        meri.uwasa.fi   44  0  0  0  0  1   0   2   8   11    22
    gjr.paknet.com.pk   43  1  0  0  1  1   3   3   6   12    16
 aristotle.algonet.se   41  0  0  0  0  0   1   2  11   12    15

The domains shown are mostly bulk-mailers and all the volume is the tail end of the time distribution, showing that short term arrival rates are moderate. Larger numbers and lower message ages are more indicative of current trouble. Old mail still going nowhere is largely harmless so long as the active and incoming queues are short. We can also see that the groups.msn.com undeliverables are low rate steady stream rather than a concentrated dictionary attack that is now over.

$ qshape -s deferred | head

                     T  5 10 20 40 80 160 320 640 1280 1280+
            TOTAL 2193  4  4  5  8 33  56 104 205  465  1309
    MAILER-DAEMON 1709  4  4  5  8 33  55 101 198  452   849
      example.com  263  0  0  0  0  0   0   0   0    2   261
      example.org  209  0  0  0  0  0   1   3   6   11   188
      example.net    6  0  0  0  0  0   0   0   0    0     6
      example.edu    3  0  0  0  0  0   0   0   0    0     3
      example.gov    2  0  0  0  0  0   0   0   1    0     1
      example.mil    1  0  0  0  0  0   0   0   0    0     1

Looking at the sender distribution, we see that as expected most of the messages are bounces.

Example 3: Congestion in the active queue

This example is taken from a Feb 2004 discussion on the Postfix Users list. Congestion was reported with the active and incoming queues large and not shrinking despite very large delivery agent process limits. The thread is archived at: http://groups.google.com/groups?th=636626c645f5bbde

Using an older version of qshape(1) it was quickly determined that all the messages were for just a few destinations:

$ qshape        (show incoming and active queue status)

                           T   A   5  10  20  40  80 160 320 320+
                 TOTAL 11775 9996  0   0   1   1  42  94 221 1420
  user.sourceforge.net  7678 7678  0   0   0   0   0   0   0    0
 lists.sourceforge.net  2313 2313  0   0   0   0   0   0   0    0
        gzd.gotdns.com   102    0  0   0   0   0   0   0   2  100

The "A" column showed the count of messages in the active queue, and the numbered columns showed totals for the deferred queue. At 10000 messages (Postfix 1.x active queue size limit) the active queue is full. The incoming was growing rapidly.

With the trouble destinations clearly identified, the administrator quickly found and fixed the problem. It is substantially harder to glean the same information from the logs. While a careful reading of mailq(1) output should yield similar results, it is much harder to gauge the magnitude of the problem by looking at the queue one message at a time.

Example 4: High volume destination backlog

When a site you send a lot of email to is down or slow, mail messages will rapidly build up in the deferred queue, or worse, in the active queue. The qshape output will show large numbers for the destination domain in all age buckets that overlap the starting time of the problem:

$ qshape deferred | head

                    T   5  10  20  40   80  160 320 640 1280 1280+
           TOTAL 5000 200 200 400 800 1600 1000 200 200  200   200
  highvolume.com 4000 160 160 320 640 1280 1440   0   0    0     0
             ...

Here the "highvolume.com" destination is continuing to accumulate deferred mail. The incoming and active queues are fine, but the deferred queue started growing some time between 1 and 2 hours ago and continues to grow.

If the high volume destination is not down, but is instead slow, one might see similar congestion in the active queue. Active queue congestion is a greater cause for alarm; one might need to take measures to ensure that the mail is deferred instead or even add an access(5) rule asking the sender to try again later.

If a high volume destination exhibits frequent bursts of consecutive connections refused by all MX hosts or "421 Server busy errors", it is possible for the queue manager to mark the destination as "dead" despite the transient nature of the errors. The destination will be retried again after the expiration of a $minimal_backoff_time timer. If the error bursts are frequent enough it may be that only a small quantity of email is delivered before the destination is again marked "dead". In some cases enabling static (not on demand) connection caching by listing the appropriate nexthop domain in a table included in "smtp_connection_cache_destinations" may help to reduce the error rate, because most messages will re-use existing connections.

The MTA that has been observed most frequently to exhibit such bursts of errors is Microsoft Exchange, which refuses connections under load. Some proxy virus scanners in front of the Exchange server propagate the refused connection to the client as a "421" error.

Note that it is now possible to configure Postfix to exhibit similarly erratic behavior by misconfiguring the anvil(8) service. Do not use anvil(8) for steady-state rate limiting, its purpose is (unintentional) DoS prevention and the rate limits set should be very generous!

If one finds oneself needing to deliver a high volume of mail to a destination that exhibits frequent brief bursts of errors and connection caching does not solve the problem, there is a subtle workaround.

The effect of this surprising configuration is that up to 2000 consecutive errors are tolerated without marking the destination dead, while the total concurrency remains reasonable (10-20 processes). This trick is only for a very specialized situation: high volume delivery into a channel with multi-error bursts that is capable of high throughput, but is repeatedly throttled by the bursts of errors.

When a destination is unable to handle the load even after the Postfix process limit is reduced to 1, a desperate measure is to insert brief delays between delivery attempts.

This solution forces the Postfix smtp(8) client to wait for $smtp_connect_timeout seconds between deliveries. The solution depends on Postfix connection management details, and needs to be updated when SMTP connection caching is introduced.

Hopefully a more elegant solution to these problems will be found in the future.

Postfix queue directories

The following sections describe Postfix queues: their purpose, what normal behavior looks like, and how to diagnose abnormal behavior.

The "maildrop" queue

Messages that have been submitted via the Postfix sendmail(1) command, but not yet brought into the main Postfix queue by the pickup(8) service, await processing in the "maildrop" queue. Messages can be added to the "maildrop" queue even when the Postfix system is not running. They will begin to be processed once Postfix is started.

The "maildrop" queue is drained by the single threaded pickup(8) service scanning the queue directory periodically or when notified of new message arrival by the postdrop(1) program. The postdrop(1) program is a setgid helper that allows the unprivileged Postfix sendmail(1) program to inject mail into the "maildrop" queue and to notify the pickup(8) service of its arrival.

All mail that enters the main Postfix queue does so via the cleanup(8) service. The cleanup service is responsible for envelope and header rewriting, header and body regular expression checks, automatic bcc recipient processing, milter content processing, and reliable insertion of the message into the Postfix "incoming" queue.

In the absence of excessive CPU consumption in cleanup(8) header or body regular expression checks or other software consuming all available CPU resources, Postfix performance is disk I/O bound. The rate at which the pickup(8) service can inject messages into the queue is largely determined by disk access times, since the cleanup(8) service must commit the message to stable storage before returning success. The same is true of the postdrop(1) program writing the message to the "maildrop" directory.

As the pickup service is single threaded, it can only deliver one message at a time at a rate that does not exceed the reciprocal disk I/O latency (+ CPU if not negligible) of the cleanup service.

Congestion in this queue is indicative of an excessive local message submission rate or perhaps excessive CPU consumption in the cleanup(8) service due to excessive body_checks, or (Postfix ≥ 2.3) high latency milters.

Note, that once the active queue is full, the cleanup service will attempt to slow down message injection by pausing $in_flow_delay for each message. In this case "maildrop" queue congestion may be a consequence of congestion downstream, rather than a problem in its own right.

Note, you should not attempt to deliver large volumes of mail via the pickup(8) service. High volume sites should avoid using "simple" content filters that re-inject scanned mail via Postfix sendmail(1) and postdrop(1).

A high arrival rate of locally submitted mail may be an indication of an uncaught forwarding loop, or a run-away notification program. Try to keep the volume of local mail injection to a moderate level.

The "postsuper -r" command can place selected messages into the "maildrop" queue for reprocessing. This is most useful for resetting any stale content_filter settings. Requeuing a large number of messages using "postsuper -r" can clearly cause a spike in the size of the "maildrop" queue.

The "hold" queue

The administrator can define "smtpd" access(5) policies, or cleanup(8) header/body checks that cause messages to be automatically diverted from normal processing and placed indefinitely in the "hold" queue. Messages placed in the "hold" queue stay there until the administrator intervenes. No periodic delivery attempts are made for messages in the "hold" queue. The postsuper(1) command can be used to manually release messages into the "deferred" queue.

Messages can potentially stay in the "hold" queue longer than $maximal_queue_lifetime. If such "old" messages need to be released from the "hold" queue, they should typically be moved into the "maildrop" queue using "postsuper -r", so that the message gets a new timestamp and is given more than one opportunity to be delivered. Messages that are "young" can be moved directly into the "deferred" queue using "postsuper -H".

The "hold" queue plays little role in Postfix performance, and monitoring of the "hold" queue is typically more closely motivated by tracking spam and malware, than by performance issues.

The "incoming" queue

All new mail entering the Postfix queue is written by the cleanup(8) service into the "incoming" queue. New queue files are created owned by the "postfix" user with an access bitmask (or mode) of 0600. Once a queue file is ready for further processing the cleanup(8) service changes the queue file mode to 0700 and notifies the queue manager of new mail arrival. The queue manager ignores incomplete queue files whose mode is 0600, as these are still being written by cleanup.

The queue manager scans the incoming queue bringing any new mail into the "active" queue if the active queue resource limits have not been exceeded. By default, the active queue accommodates at most 20000 messages. Once the active queue message limit is reached, the queue manager stops scanning the incoming (and deferred, see below) queue.

Under normal conditions the incoming queue is nearly empty (has only mode 0600 files), with the queue manager able to import new messages into the active queue as soon as they become available.

The incoming queue grows when the message input rate spikes above the rate at which the queue manager can import messages into the active queue. The main factors slowing down the queue manager are disk I/O and lookup queries to the trivial-rewrite service. If the queue manager is routinely not keeping up, consider not using "slow" lookup services (MySQL, LDAP, ...) for transport lookups or speeding up the hosts that provide the lookup service. If the problem is I/O starvation, consider striping the queue over more disks, faster controllers with a battery write cache, or other hardware improvements. At the very least, make sure that the queue directory is mounted with the "noatime" option if applicable to the underlying filesystem.

The in_flow_delay parameter is used to clamp the input rate when the queue manager starts to fall behind. The cleanup(8) service will pause for $in_flow_delay seconds before creating a new queue file if it cannot obtain a "token" from the queue manager.

Since the number of cleanup(8) processes is limited in most cases by the SMTP server concurrency, the input rate can exceed the output rate by at most "SMTP connection count" / $in_flow_delay messages per second.

With a default process limit of 100, and an in_flow_delay of 1s, the coupling is strong enough to limit a single run-away injector to 1 message per second, but is not strong enough to deflect an excessive input rate from many sources at the same time.

If a server is being hammered from multiple directions, consider raising the in_flow_delay to 10 seconds, but only if the incoming queue is growing even while the active queue is not full and the trivial-rewrite service is using a fast transport lookup mechanism.

The "active" queue

The queue manager is a delivery agent scheduler; it works to ensure fast and fair delivery of mail to all destinations within designated resource limits.

The active queue is somewhat analogous to an operating system's process run queue. Messages in the active queue are ready to be sent (runnable), but are not necessarily in the process of being sent (running).

While most Postfix administrators think of the "active" queue as a directory on disk, the real "active" queue is a set of data structures in the memory of the queue manager process.

Messages in the "maildrop", "hold", "incoming" and "deferred" queues (see below) do not occupy memory; they are safely stored on disk waiting for their turn to be processed. The envelope information for messages in the "active" queue is managed in memory, allowing the queue manager to do global scheduling, allocating available delivery agent processes to an appropriate message in the active queue.

Within the active queue, (multi-recipient) messages are broken up into groups of recipients that share the same transport/nexthop combination; the group size is capped by the transport's recipient concurrency limit.

Multiple recipient groups (from one or more messages) are queued for delivery grouped by transport/nexthop combination. The destination concurrency limit for the transports caps the number of simultaneous delivery attempts for each nexthop. Transports with a recipient concurrency limit of 1 are special: these are grouped by the actual recipient address rather than the nexthop, yielding per-recipient concurrency limits rather than per-domain concurrency limits. Per-recipient limits are appropriate when performing final delivery to mailboxes rather than when relaying to a remote server.

Congestion occurs in the active queue when one or more destinations drain slower than the corresponding message input rate.

Input into the active queue comes both from new mail in the "incoming" queue, and retries of mail in the "deferred" queue. Should the "deferred" queue get really large, retries of old mail can dominate the arrival rate of new mail. Systems with more CPU, faster disks and more network bandwidth can deal with larger deferred queues, but as a rule of thumb the deferred queue scales to somewhere between 100,000 and 1,000,000 messages with good performance unlikely above that "limit". Systems with queues this large should typically stop accepting new mail, or put the backlog "on hold" until the underlying issue is fixed (provided that there is enough capacity to handle just the new mail).

When a destination is down for some time, the queue manager will mark it dead, and immediately defer all mail for the destination without trying to assign it to a delivery agent. In this case the messages will quickly leave the active queue and end up in the deferred queue (with Postfix < 2.4, this is done directly by the queue manager, with Postfix ≥ 2.4 this is done via the "retry" delivery agent).

When the destination is instead simply slow, or there is a problem causing an excessive arrival rate the active queue will grow and will become dominated by mail to the congested destination.

The only way to reduce congestion is to either reduce the input rate or increase the throughput. Increasing the throughput requires either increasing the concurrency or reducing the latency of deliveries.

For high volume sites a key tuning parameter is the number of "smtp" delivery agents allocated to the "smtp" and "relay" transports. High volume sites tend to send to many different destinations, many of which may be down or slow, so a good fraction of the available delivery agents will be blocked waiting for slow sites. Also mail destined across the globe will incur large SMTP command-response latencies, so high message throughput can only be achieved with more concurrent delivery agents.

The default "smtp" process limit of 100 is good enough for most sites, and may even need to be lowered for sites with low bandwidth connections (no use increasing concurrency once the network pipe is full). When one finds that the queue is growing on an "idle" system (CPU, disk I/O and network not exhausted) the remaining reason for congestion is insufficient concurrency in the face of a high average latency. If the number of outbound SMTP connections (either ESTABLISHED or SYN_SENT) reaches the process limit, mail is draining slowly and the system and network are not loaded, raise the "smtp" and/or "relay" process limits!

When a high volume destination is served by multiple MX hosts with typically low delivery latency, performance can suffer dramatically when one of the MX hosts is unresponsive and SMTP connections to that host timeout. For example, if there are 2 equal weight MX hosts, the SMTP connection timeout is 30 seconds and one of the MX hosts is down, the average SMTP connection will take approximately 15 seconds to complete. With a default per-destination concurrency limit of 20 connections, throughput falls to just over 1 message per second.

The best way to avoid bottlenecks when one or more MX hosts is non-responsive is to use connection caching. Connection caching was introduced with Postfix 2.2 and is by default enabled on demand for destinations with a backlog of mail in the active queue. When connection caching is in effect for a particular destination, established connections are re-used to send additional messages, this reduces the number of connections made per message delivery and maintains good throughput even in the face of partial unavailability of the destination's MX hosts.

If connection caching is not available (Postfix < 2.2) or does not provide a sufficient latency reduction, especially for the "relay" transport used to forward mail to "your own" domains, consider setting lower than default SMTP connection timeouts (1-5 seconds) and higher than default destination concurrency limits. This will further reduce latency and provide more concurrency to maintain throughput should latency rise.

Setting high concurrency limits to domains that are not your own may be viewed as hostile by the receiving system, and steps may be taken to prevent you from monopolizing the destination system's resources. The defensive measures may substantially reduce your throughput or block access entirely. Do not set aggressive concurrency limits to remote domains without coordinating with the administrators of the target domain.

If necessary, dedicate and tune custom transports for selected high volume destinations. The "relay" transport is provided for forwarding mail to domains for which your server is a primary or backup MX host. These can make up a substantial fraction of your email traffic. Use the "relay" and not the "smtp" transport to send email to these domains. Using the "relay" transport allocates a separate delivery agent pool to these destinations and allows separate tuning of timeouts and concurrency limits.

Another common cause of congestion is unwarranted flushing of the entire deferred queue. The deferred queue holds messages that are likely to fail to be delivered and are also likely to be slow to fail delivery (time out). As a result the most common reaction to a large deferred queue (flush it!) is more than likely counter-productive, and typically makes the congestion worse. Do not flush the deferred queue unless you expect that most of its content has recently become deliverable (e.g. relayhost back up after an outage)!

Note that whenever the queue manager is restarted, there may already be messages in the active queue directory, but the "real" active queue in memory is empty. In order to recover the in-memory state, the queue manager moves all the active queue messages back into the incoming queue, and then uses its normal incoming queue scan to refill the active queue. The process of moving all the messages back and forth, redoing transport table (trivial-rewrite(8) resolve service) lookups, and re-importing the messages back into memory is expensive. At all costs, avoid frequent restarts of the queue manager (e.g. via frequent execution of "postfix reload").

The "deferred" queue

When all the deliverable recipients for a message are delivered, and for some recipients delivery failed for a transient reason (it might succeed later), the message is placed in the deferred queue.

The queue manager scans the deferred queue periodically. The scan interval is controlled by the queue_run_delay parameter. While a deferred queue scan is in progress, if an incoming queue scan is also in progress (ideally these are brief since the incoming queue should be short), the queue manager alternates between looking for messages in the "incoming" queue and in the "deferred" queue. This "round-robin" strategy prevents starvation of either the incoming or the deferred queues.

Each deferred queue scan only brings a fraction of the deferred queue back into the active queue for a retry. This is because each message in the deferred queue is assigned a "cool-off" time when it is deferred. This is done by time-warping the modification time of the queue file into the future. The queue file is not eligible for a retry if its modification time is not yet reached.

The "cool-off" time is at least $minimal_backoff_time and at most $maximal_backoff_time. The next retry time is set by doubling the message's age in the queue, and adjusting up or down to lie within the limits. This means that young messages are initially retried more often than old messages.

If a high volume site routinely has large deferred queues, it may be useful to adjust the queue_run_delay, minimal_backoff_time and maximal_backoff_time to provide short enough delays on first failure (Postfix ≥ 2.4 has a sensibly low minimal backoff time by default), with perhaps longer delays after multiple failures, to reduce the retransmission rate of old messages and thereby reduce the quantity of previously deferred mail in the active queue. If you want a really low minimal_backoff_time, you may also want to lower queue_run_delay, but understand that more frequent scans will increase the demand for disk I/O.

One common cause of large deferred queues is failure to validate recipients at the SMTP input stage. Since spammers routinely launch dictionary attacks from unrepliable sender addresses, the bounces for invalid recipient addresses clog the deferred queue (and at high volumes proportionally clog the active queue). Recipient validation is strongly recommended through use of the local_recipient_maps and relay_recipient_maps parameters. Even when bounces drain quickly they inundate innocent victims of forgery with unwanted email. To avoid this, do not accept mail for invalid recipients.

When a host with lots of deferred mail is down for some time, it is possible for the entire deferred queue to reach its retry time simultaneously. This can lead to a very full active queue once the host comes back up. The phenomenon can repeat approximately every maximal_backoff_time seconds if the messages are again deferred after a brief burst of congestion. Perhaps, a future Postfix release will add a random offset to the retry time (or use a combination of strategies) to reduce the odds of repeated complete deferred queue flushes.

Credits

The qshape(1) program was developed by Victor Duchovni of Morgan Stanley, who also wrote the initial version of this document.