Postfix Queue Scheduler


Disclaimer

Many of the transport-specific configuration parameters discussed in this document will not show up in "postconf" command output before Postfix version 2.9. This limitation applies to many parameters whose name is a combination of a master.cf service name such as "relay" and a built-in suffix such as "_destination_concurrency_limit".

Overview

The queue manager is by far the most complex part of the Postfix mail system. It schedules delivery of new mail, retries failed deliveries at specific times, and removes mail from the queue after the last delivery attempt. There are two major classes of mechanisms that control the operation of the queue manager.

Topics covered by this document:

Concurrency scheduling

The following sections document the Postfix 2.5 concurrency scheduler, after a discussion of the limitations of the earlier concurrency scheduler. This is followed by results of medium-concurrency experiments, and a discussion of trade-offs between performance and robustness.

The material is organized as follows:

Drawbacks of the existing concurrency scheduler

From the start, Postfix has used a simple but robust algorithm where the per-destination delivery concurrency is decremented by 1 after delivery failed due to connection or handshake failure, and incremented by 1 otherwise. Of course the concurrency is never allowed to exceed the maximum per-destination concurrency limit. And when a destination's concurrency level drops to zero, the destination is declared "dead" and delivery is suspended.

Drawbacks of +/-1 concurrency feedback per delivery are:

(*) A pseudo-cohort is a number of delivery requests equal to a destination's delivery concurrency.

The revised concurrency scheduler has a highly modular structure. It uses separate mechanisms for per-destination concurrency control and for "dead destination" detection. The concurrency control in turn is built from two separate mechanisms: it supports less-than-1 feedback per delivery to allow for more gradual concurrency adjustments, and it uses feedback hysteresis to suppress concurrency oscillations. And instead of waiting for delivery concurrency to throttle down to zero, a destination is declared "dead" after a configurable number of pseudo-cohorts reports connection or handshake failure.

Summary of the Postfix 2.5 concurrency feedback algorithm

We want to increment a destination's delivery concurrency when some (not necessarily consecutive) number of deliveries complete without connection or handshake failure. This is implemented with positive feedback g(N) where N is the destination's delivery concurrency. With g(N)=1 feedback per delivery, concurrency increases by 1 after each positive feedback event; this gives us the old scheduler's exponential growth in time. With g(N)=1/N feedback per delivery, concurrency increases by 1 after an entire pseudo-cohort N of positive feedback reports; this gives us linear growth in time. Less-than-1 feedback per delivery and integer truncation naturally give us hysteresis, so that transitions to larger concurrency happen every 1/g(N) positive feedback events.

We want to decrement a destination's delivery concurrency when some (not necessarily consecutive) number of deliveries complete after connection or handshake failure. This is implemented with negative feedback f(N) where N is the destination's delivery concurrency. With f(N)=1 feedback per delivery, concurrency decreases by 1 after each negative feedback event; this gives us the old scheduler's behavior where concurrency is throttled down dramatically after a single pseudo-cohort failure. With f(N)=1/N feedback per delivery, concurrency backs off more gently. Again, less-than-1 feedback per delivery and integer truncation naturally give us hysteresis, so that transitions to lower concurrency happen every 1/f(N) negative feedback events.

However, with negative feedback we introduce a subtle twist. We "reverse" the negative hysteresis cycle so that the transition to lower concurrency happens at the beginning of a sequence of 1/f(N) negative feedback events. Otherwise, a correction for overload would be made too late. This makes the choice of f(N) relatively unimportant, as borne out by measurements later in this document.

In summary, the main ingredients for the Postfix 2.5 concurrency feedback algorithm are a) the option of less-than-1 positive feedback per delivery to avoid overwhelming servers, b) the option of less-than-1 negative feedback per delivery to avoid giving up too fast, c) feedback hysteresis to avoid rapid oscillation, and d) a "reverse" hysteresis cycle for negative feedback, so that it can correct for overload quickly.

Summary of the Postfix 2.5 "dead destination" detection algorithm

We want to suspend deliveries to a specific destination after some number of deliveries suffers connection or handshake failure. The old scheduler declares a destination "dead" when negative (-1) feedback throttles the delivery concurrency down to zero. With less-than-1 feedback per delivery, this throttling down would obviously take too long. We therefore have to separate "dead destination" detection from concurrency feedback. This is implemented by introducing the concept of pseudo-cohort failure. The Postfix 2.5 concurrency scheduler declares a destination "dead" after a configurable number of pseudo-cohorts suffers from connection or handshake failures. The old scheduler corresponds to the special case where the pseudo-cohort failure limit is equal to 1.

Pseudocode for the Postfix 2.5 concurrency scheduler

The pseudo code shows how the ideas behind new concurrency scheduler are implemented as of November 2007. The actual code can be found in the module qmgr/qmgr_queue.c.

Types:
        Each destination has one set of the following variables
        int concurrency
        double success
        double failure
        double fail_cohorts

Feedback functions:
        N is concurrency; x, y are arbitrary numbers in [0..1] inclusive
        positive feedback: g(N) = x/N | x/sqrt(N) | x
        negative feedback: f(N) = y/N | y/sqrt(N) | y

Initialization:
        concurrency = initial_concurrency
        success = 0
        failure = 0
        fail_cohorts = 0

After success:
        fail_cohorts = 0
        Be prepared for feedback > hysteresis, or rounding error
        success += g(concurrency)
        while (success >= 1)            Hysteresis 1
            concurrency += 1            Hysteresis 1
            failure = 0
            success -= 1                Hysteresis 1
        Be prepared for overshoot
        if (concurrency > concurrency limit)
            concurrency = concurrency limit

Safety:
        Don't apply positive feedback unless
            concurrency < busy_refcount + init_dest_concurrency
        otherwise negative feedback effect could be delayed

After failure:
        if (concurrency > 0)
            fail_cohorts += 1.0 / concurrency
            if (fail_cohorts > cohort_failure_limit)
                concurrency = 0
        if (concurrency > 0)
            Be prepared for feedback > hysteresis, rounding errors
            failure -= f(concurrency)
            while (failure < 0)
                concurrency -= 1        Hysteresis 1
                failure += 1            Hysteresis 1
                success = 0
            Be prepared for overshoot
            if (concurrency < 1)
                concurrency = 1

Results for delivery to concurrency-limited servers

Discussions about the concurrency scheduler redesign started early 2004, when the primary goal was to find alternatives that did not exhibit exponential growth or rapid concurrency throttling. No code was implemented until late 2007, when the primary concern had shifted towards better handling of server concurrency limits. For this reason we measure how well the new scheduler does this job. The table below compares mail delivery performance of the old +/-1 feedback per delivery with several less-than-1 feedback functions, for different limited-concurrency server scenarios. Measurements were done with a FreeBSD 6.2 client and with FreeBSD 6.2 and various Linux servers.

Server configuration:

Client configuration:

Impact of the 30s SMTP connect timeout

The first results are for a FreeBSD 6.2 server, where our artificially low listen(2) backlog results in a very short kernel queue for established connections. The table shows that all deferred deliveries failed due to a 30s connection timeout, and none failed due to a server greeting timeout. This measurement simulates what happens when the server's connection queue is completely full under load, and the TCP engine drops new connections.

client
limit
server
limit
feedback
style
connection
caching
percentage
deferred
client concurrency
average/stddev
timed-out in
connect/greeting

20 5 1/N no 9.9 19.4 0.49 198 -
20 5 1/N yes 10.3 19.4 0.49 206 -
20 5 1/sqrt(N) no 10.4 19.6 0.59 208 -
20 5 1/sqrt(N) yes 10.6 19.6 0.61 212 -
20 5 1 no 10.1 19.5 1.29 202 -
20 5 1 yes 10.8 19.3 1.57 216 -

A busy server with a completely full connection queue. N is the client delivery concurrency. Failed deliveries time out after 30s without completing the TCP handshake. See text for a discussion of results.

Impact of the 300s SMTP greeting timeout

The next table shows results for a Fedora Core 8 server (results for RedHat 7.3 are identical). In this case, the artificially small listen(2) backlog argument does not impact our measurement. The table shows that practically all deferred deliveries fail after the 300s SMTP greeting timeout. As these timeouts were 10x longer than with the first measurement, we increased the recipient count (and thus the running time) by a factor of 10 to keep the results comparable. The deferred mail percentages are a factor 10 lower than with the first measurement, because the 1s per-recipient delay was 1/300th of the greeting timeout instead of 1/30th of the connection timeout.

client
limit
server
limit
feedback
style
connection
caching
percentage
deferred
client concurrency
average/stddev
timed-out in
connect/greeting

20 5 1/N no 1.16 19.8 0.37 - 230
20 5 1/N yes 1.36 19.8 0.36 - 272
20 5 1/sqrt(N) no 1.21 19.9 0.23 4 238
20 5 1/sqrt(N) yes 1.36 20.0 0.23 - 272
20 5 1 no 1.18 20.0 0.16 - 236
20 5 1 yes 1.39 20.0 0.16 - 278

A busy server with a non-full connection queue. N is the client delivery concurrency. Failed deliveries complete at the TCP level, but time out after 300s while waiting for the SMTP greeting. See text for a discussion of results.

Impact of active server concurrency limiter

The final concurrency-limited result shows what happens when SMTP connections don't time out, but are rejected immediately with the Postfix server's smtpd_client_connection_count_limit feature (the server replies with a 421 status and disconnects immediately). Similar results can be expected with concurrency limiting features built into other MTAs or firewalls. For this measurement we specified a server concurrency limit and a client initial destination concurrency of 5, and a server process limit of 10; all other conditions were the same as with the first measurement. The same result would be obtained with a FreeBSD or Linux server, because the "pushing back" is done entirely by the receiving side.

client
limit
server
limit
feedback
style
connection
caching
percentage
deferred
client concurrency
average/stddev
theoretical
defer rate

20 5 1/N no 16.5 5.17 0.38 1/6
20 5 1/N yes 16.5 5.17 0.38 1/6
20 5 1/sqrt(N) no 24.5 5.28 0.45 1/4
20 5 1/sqrt(N) yes 24.3 5.28 0.46 1/4
20 5 1 no 49.7 5.63 0.67 1/2
20 5 1 yes 49.7 5.68 0.70 1/2

A server with active per-client concurrency limiter that replies with 421 and disconnects. N is the client delivery concurrency. The theoretical defer rate is 1/(1+roundup(1/feedback)). This is always 1/2 with the fixed +/-1 feedback per delivery; with the concurrency-dependent feedback variants, the defer rate decreases with increasing concurrency. See text for a discussion of results.

Discussion of concurrency-limited server results

All results in the previous sections are based on the first delivery runs only; they do not include any second etc. delivery attempts. It's also worth noting that the measurements look at steady-state behavior only. They don't show what happens when the client starts sending at a much higher or lower concurrency.

The first two examples show that the effect of feedback is negligible when concurrency is limited due to congestion. This is because the initial concurrency is already at the client's concurrency maximum, and because there is 10-100 times more positive than negative feedback. Under these conditions, it is no surprise that the contribution from SMTP connection caching is also negligible.

In the last example, the old +/-1 feedback per delivery will defer 50% of the mail when confronted with an active (anvil-style) server concurrency limit, where the server hangs up immediately with a 421 status (a TCP-level RST would have the same result). Less aggressive feedback mechanisms fare better than more aggressive ones. Concurrency-dependent feedback fares even better at higher concurrencies than shown here, but has limitations as discussed in the next section.

Limitations of less-than-1 per delivery feedback

Less-than-1 feedback is of interest primarily when sending large amounts of mail to destinations with active concurrency limiters (servers that reply with 421, or firewalls that send RST). When sending small amounts of mail per destination, less-than-1 per-delivery feedback won't have a noticeable effect on the per-destination concurrency, because the number of deliveries to the same destination is too small. You might just as well use zero per-delivery feedback and stay with the initial per-destination concurrency. And when mail deliveries fail due to congestion instead of active concurrency limiters, the measurements above show that per-delivery feedback has no effect. With large amounts of mail you might just as well use zero per-delivery feedback and start with the maximal per-destination concurrency.

The scheduler with less-than-1 concurrency feedback per delivery solves a problem with servers that have active concurrency limiters. This works only because feedback is handled in a peculiar manner: positive feedback will increment the concurrency by 1 at the end of a sequence of events of length 1/feedback, while negative feedback will decrement concurrency by 1 at the beginning of such a sequence. This is how Postfix adjusts quickly for overshoot without causing lots of mail to be deferred. Without this difference in feedback treatment, less-than-1 feedback per delivery would defer 50% of the mail, and would be no better in this respect than the old +/-1 feedback per delivery.

Unfortunately, the same feature that corrects quickly for concurrency overshoot also makes the scheduler more sensitive for noisy negative feedback. The reason is that one lonely negative feedback event has the same effect as a complete sequence of length 1/feedback: in both cases delivery concurrency is dropped by 1 immediately. As a worst-case scenario, consider multiple servers behind a load balancer on a single IP address, and no backup MX address. When 1 out of K servers fails to complete the SMTP handshake or drops the connection, a scheduler with 1/N (N = concurrency) feedback stops increasing its concurrency once it reaches a concurrency level of about K, even though the good servers behind the load balancer are perfectly capable of handling more traffic.

This noise problem gets worse as the amount of positive feedback per delivery gets smaller. A compromise is to use fixed less-than-1 positive feedback values instead of concurrency-dependent positive feedback. For example, to tolerate 1 of 4 bad servers in the above load balancer scenario, use positive feedback of 1/4 per "good" delivery (no connect or handshake error), and use an equal or smaller amount of negative feedback per "bad" delivery. The downside of using concurrency-independent feedback is that some of the old +/-1 feedback problems will return at large concurrencies. Sites that must deliver mail at non-trivial per-destination concurrencies will require special configuration.

Concurrency configuration parameters

The Postfix 2.5 concurrency scheduler is controlled with the following configuration parameters, where "transport_foo" provides a transport-specific parameter override. All parameter default settings are compatible with earlier Postfix versions.

Parameter name Postfix version Description

initial_destination_concurrency
transport_initial_destination_concurrency
all
2.5
Initial per-destination delivery concurrency
default_destination_concurrency_limit
transport_destination_concurrency_limit
all
all
Maximum per-destination delivery concurrency
default_destination_concurrency_positive_feedback
transport_destination_concurrency_positive_feedback
2.5
2.5
Per-destination positive feedback amount, per delivery that does not fail with connection or handshake failure
default_destination_concurrency_negative_feedback
transport_destination_concurrency_negative_feedback
2.5
2.5
Per-destination negative feedback amount, per delivery that fails with connection or handshake failure
default_destination_concurrency_failed_cohort_limit
transport_destination_concurrency_failed_cohort_limit
2.5
2.5
Number of failed pseudo-cohorts after which a destination is declared "dead" and delivery is suspended
destination_concurrency_feedback_debug 2.5 Enable verbose logging of concurrency scheduler activity

Preemptive scheduling

The following sections describe the new queue manager and its preemptive scheduler algorithm. Note that the document was originally written to describe the changes between the new queue manager (in this text referred to as nqmgr, the name it was known by before it became the default queue manager) and the old queue manager (referred to as oqmgr). This is why it refers to oqmgr every so often.

This document is divided into sections as follows:

The structures used by nqmgr

Let's start by recapitulating the structures and terms used when referring to queue manager and how it operates. Many of these are partially described elsewhere, but it is nice to have a coherent overview in one place:

The first four structures are common to both nqmgr and oqmgr, the latter two were introduced by nqmgr.

These terms are used extensively in the text below, feel free to look up the description above anytime you'll feel you have lost a sense what is what.

What happens when nqmgr picks up the message

Whenever nqmgr moves a queue file into the active queue, the following happens: It reads all necessary information from the queue file as oqmgr does, and also reads as many recipients as possible - more on that later, for now let's just pretend it always reads all recipients.

Then it resolves the recipients as oqmgr does, which means obtaining (address, nexthop, transport) triple for each recipient. For each triple, it finds the transport; if it does not exist yet, it instantiates it (unless it's dead). Within the transport, it finds the destination queue for given nexthop; if it does not exist yet, it instantiates it (unless it's dead). The triple is then bound to given destination queue. This happens in qmgr_resolve() and is basically the same as in oqmgr.

Then for each triple which was bound to some queue (and thus transport), the program finds the job which represents the message within that transport's context; if it does not exist yet, it instantiates it. Within the job, it finds the peer which represents the bound destination queue within this jobs context; if it does not exist yet, it instantiates it. Finally, it stores the address from the resolved triple to the recipient entry which is appended to both the queue entry list and the peer entry list. The addresses for same nexthop are batched in the entries up to recipient_concurrency limit for that transport. This happens in qmgr_assign() and apart from that it operates with job and peer structures it is basically the same as in oqmgr.

When the job is instantiated, it is enqueued on the transport's job list based on the time its message was picked up by nqmgr. For first batch of recipients this means it is appended to the end of the job list, but the ordering of the job list by the enqueue time is important as we will see shortly.

[Now you should have pretty good idea what is the state of the nqmgr after couple of messages was picked up, what is the relation between all those job, peer, queue and entry structures.]

How the entry selection works

Having prepared all those above mentioned structures, the task of the nqmgr's scheduler is to choose the recipient entries one at a time and pass them to the delivery agent for corresponding transport. Now how does this work?

The first approximation of the new scheduling algorithm is like this:

foreach transport (round-robin-by-transport)
do
    if transport busy continue
    if transport process limit reached continue
    foreach transport's job (in the order of the transport's job list)
    do
	foreach job's peer (round-robin-by-destination)
	     if peer->queue->concurrency < peer->queue->window
		 return next peer entry.
	done
    done
done

Now what is the "order of the transport's job list"? As we know already, the job list is by default kept in the order the message was picked up by the nqmgr. So by default we get the top-level round-robin transport, and within each transport we get the FIFO message delivery. The round-robin of the peers by the destination is perhaps of little importance in most real-life cases (unless the recipient_concurrency limit is reached, in one job there is only one peer structure for each destination), but theoretically it makes sure that even within single jobs, destinations are treated fairly.

[By now you should have a feeling you really know how the scheduler works, except for the preemption, under ideal conditions - that is, no recipient resource limits and no destination concurrency problems.]

How the preemption works

As you might perhaps expect by now, the transport's job list does not remain sorted by the job's message enqueue time all the time. The most cool thing about nqmgr is not the simple FIFO delivery, but that it is able to slip mail with little recipients past the mailing-list bulk mail. This is what the job preemption is about - shuffling the jobs on the transport's job list to get the best message delivery rates. Now how is it achieved?

First I have to tell you that there are in fact two job lists in each transport. One is the scheduler's job list, which the scheduler is free to play with, while the other one keeps the jobs always listed in the order of the enqueue time and is used for recipient pool management we will discuss later. For now, we will deal with the scheduler's job list only.

So, we have the job list, which is first ordered by the time the jobs' messages were enqueued, oldest messages first, the most recently picked one at the end. For now, let's assume that there are no destination concurrency problems. Without preemption, we pick some entry of the first (oldest) job on the queue, assign it to delivery agent, pick another one from the same job, assign it again, and so on, until all the entries are used and the job is delivered. We would then move onto the next job and so on and on. Now how do we manage to sneak in some entries from the recently added jobs when the first job on the job list belongs to a message going to the mailing-list and has thousands of recipient entries?

The nqmgr's answer is that we can artificially "inflate" the delivery time of that first job by some constant for free - it is basically the same trick you might remember as "accumulation of potential" from the amortized complexity lessons. For example, instead of delivering the entries of the first job on the job list every time a delivery agent becomes available, we can do it only every second time. If you view the moments the delivery agent becomes available on a timeline as "delivery slots", then instead of using every delivery slot for the first job, we can use only every other slot, and still the overall delivery efficiency of the first job remains the same. So the delivery 11112222 becomes 1.1.1.1.2.2.2.2 (1 and 2 are the imaginary job numbers, . denotes the free slot). Now what do we do with free slots?

As you might have guessed, we will use them for sneaking the mail with little recipients in. For example, if we have one four-recipient mail followed by four one recipients mail, the delivery sequence (that is, the sequence in which the jobs are assigned to the delivery slots) might look like this: 12131415. Hmm, fine for sneaking in the single recipient mail, but how do we sneak in the mail with more than one recipient? Say if we have one four-recipient mail followed by two two-recipient mails?

The simple answer would be to use delivery sequence 12121313. But the problem is that this does not scale well. Imagine you have mail with thousand recipients followed by mail with hundred recipients. It is tempting to suggest the delivery sequence like 121212...., but alas! Imagine there arrives another mail with say ten recipients. But there are no free slots anymore, so it can't slip by, not even if it had just only one recipients. It will be stuck until the hundred-recipient mail is delivered, which really sucks.

So, it becomes obvious that while inflating the message to get free slots is great idea, one has to be really careful of how the free slots are assigned, otherwise one might corner himself. So, how does nqmgr really use the free slots?

The key idea is that one does not have to generate the free slots in a uniform way. The delivery sequence 111...1 is no worse than 1.1.1.1, in fact, it is even better as some entries are in the first case selected earlier than in the second case, and none is selected later! So it is possible to first "accumulate" the free delivery slots and then use them all at once. It is even possible to accumulate some, then use them, then accumulate some more and use them again, as in 11..1.1 .

Let's get back to the one hundred recipient example. We now know that we could first accumulate one hundred free slots, and only after then to preempt the first job and sneak the one hundred recipient mail in. Applying the algorithm recursively, we see the hundred recipient job can accumulate ten free delivery slots, and then we could preempt it and sneak in the ten-recipient mail... Wait wait wait! Could we? Aren't we overinflating the original one thousand recipient mail?

Well, despite it looks so at the first glance, another trick will allow us to answer "no, we are not!". If we had said that we will inflate the delivery time twice at maximum, and then we consider every other slot as a free slot, then we would overinflate in case of the recursive preemption. BUT! The trick is that if we use only every n-th slot as a free slot for n>2, there is always some worst inflation factor which we can guarantee not to be breached, even if we apply the algorithm recursively. To be precise, if for every k>1 normally used slots we accumulate one free delivery slot, than the inflation factor is not worse than k/(k-1) no matter how many recursive preemptions happen. And it's not worse than (k+1)/k if only non-recursive preemption happens. Now, having got through the theory and the related math, let's see how nqmgr implements this.

Each job has so called "available delivery slot" counter. Each transport has a transport_delivery_slot_cost parameter, which defaults to default_delivery_slot_cost parameter which is set to 5 by default. This is the k from the paragraph above. Each time k entries of the job are selected for delivery, this counter is incremented by one. Once there are some slots accumulated, job which requires no more than that number of slots to be fully delivered can preempt this job.

[Well, the truth is, the counter is incremented every time an entry is selected and it is divided by k when it is used. But for the understanding it's good enough to use the above approximation of the truth.]

OK, so now we know the conditions which must be satisfied so one job can preempt another one. But what job gets preempted, how do we choose what job preempts it if there are several valid candidates, and when does all this exactly happen?

The answer for the first part is simple. The job whose entry was selected the last time is so called current job. Normally, it is the first job on the scheduler's job list, but destination concurrency limits may change this as we will see later. It is always only the current job which may get preempted.

Now for the second part. The current job has certain amount of recipient entries, and as such may accumulate at maximum some amount of available delivery slots. It might have already accumulated some, and perhaps even already used some when it was preempted before (remember a job can be preempted several times). In either case, we know how many are accumulated and how many are left to deliver, so we know how many it may yet accumulate at maximum. Every other job which may be delivered by less than that number of slots is a valid candidate for preemption. How do we choose among them?

The answer is - the one with maximum enqueue_time/recipient_entry_count. That is, the older the job is, the more we should try to deliver it in order to get best message delivery rates. These rates are of course subject to how many recipients the message has, therefore the division by the recipient (entry) count. No one shall be surprised that message with n recipients takes n times longer to deliver than message with one recipient.

Now let's recap the previous two paragraphs. Isn't it too complicated? Why don't the candidates come only among the jobs which can be delivered within the number of slots the current job already accumulated? Why do we need to estimate how much it has yet to accumulate? If you found out the answer, congratulate yourself. If we did it this simple way, we would always choose the candidate with least recipient entries. If there were enough single recipient mails coming in, they would always slip by the bulk mail as soon as possible, and the two and more recipients mail would never get a chance, no matter how long they have been sitting around in the job list.

This candidate selection has interesting implication - that when we choose the best candidate for preemption (this is done in qmgr_choose_candidate()), it may happen that we may not use it for preemption immediately. This leads to an answer to the last part of the original question - when does the preemption happen?

The preemption attempt happens every time next transport's recipient entry is to be chosen for delivery. To avoid needless overhead, the preemption is not attempted if the current job could never accumulate more than transport_minimum_delivery_slots (defaults to default_minimum_delivery_slots which defaults to 3). If there is already enough accumulated slots to preempt the current job by the chosen best candidate, it is done immediately. This basically means that the candidate is moved in front of the current job on the scheduler's job list and decreasing the accumulated slot counter by the amount used by the candidate. If there is not enough slots... well, I could say that nothing happens and the another preemption is attempted the next time. But that's not the complete truth.

The truth is that it turns out that it is not really necessary to wait until the jobs counter accumulates all the delivery slots in advance. Say we have ten-recipient mail followed by two two-recipient mails. If the preemption happened when enough delivery slot accumulate (assuming slot cost 2), the delivery sequence becomes 11112211113311. Now what would we get if we would wait only for 50% of the necessary slots to accumulate and we promise we would wait for the remaining 50% later, after we get back to the preempted job? If we use such slot loan, the delivery sequence becomes 11221111331111. As we can see, it makes it no considerably worse for the delivery of the ten-recipient mail, but it allows the small messages to be delivered sooner.

The concept of these slot loans is where the transport_delivery_slot_discount and transport_delivery_slot_loan come from (they default to default_delivery_slot_discount and default_delivery_slot_loan, whose values are by default 50 and 3, respectively). The discount (resp. loan) specifies how many percent (resp. how many slots) one "gets in advance", when the number of slots required to deliver the best candidate is compared with the number of slots the current slot had accumulated so far.

And it pretty much concludes this chapter.

[Now you should have a feeling that you pretty much understand the scheduler and the preemption, or at least that you will have it after you read the last chapter couple more times. You shall clearly see the job list and the preemption happening at its head, in ideal delivery conditions. The feeling of understanding shall last until you start wondering what happens if some of the jobs are blocked, which you might eventually figure out correctly from what had been said already. But I would be surprised if your mental image of the scheduler's functionality is not completely shattered once you start wondering how it works when not all recipients may be read in-core. More on that later.]

How destination concurrency limits affect the scheduling algorithm

The nqmgr uses the same algorithm for destination concurrency control as oqmgr. Now what happens when the destination limits are reached and no more entries for that destination may be selected by the scheduler?

From user's point of view it is all simple. If some of the peers of a job can't be selected, those peers are simply skipped by the entry selection algorithm (the pseudo-code described before) and only the selectable ones are used. If none of the peers may be selected, the job is declared a "blocker job". Blocker jobs are skipped by the entry selection algorithm and they are also excluded from the candidates for preemption of current job. Thus the scheduler effectively behaves as if the blocker jobs didn't exist on the job list at all. As soon as at least one of the peers of a blocker job becomes unblocked (that is, the delivery agent handling the delivery of the recipient entry for given destination successfully finishes), the job's blocker status is removed and the job again participates in all further scheduler actions normally.

So the summary is that the users don't really have to be concerned about the interaction of the destination limits and scheduling algorithm. It works well on its own and there are no knobs they would need to control it.

From a programmer's point of view, the blocker jobs complicate the scheduler quite a lot. Without them, the jobs on the job list would be normally delivered in strict FIFO order. If the current job is preempted, the job preempting it is completely delivered unless it is preempted itself. Without blockers, the current job is thus always either the first job on the job list, or the top of the stack of jobs preempting the first job on the job list.

The visualization of the job list and the preemption stack without blockers would be like this:

first job->    1--2--3--5--6--8--...    <- job list
on job list    |
               4    <- preemption stack
               |
current job->  7

In the example above we see that job 1 was preempted by job 4 and then job 4 was preempted by job 7. After job 7 is completed, remaining entries of job 4 are selected, and once they are all selected, job 1 continues.

As we see, it's all very clean and straightforward. Now how does this change because of blockers?

The answer is: a lot. Any job may become blocker job at any time, and also become normal job again at any time. This has several important implications:

  1. The jobs may be completed in arbitrary order. For example, in the example above, if the current job 7 becomes blocked, the next job 4 may complete before the job 7 becomes unblocked again. Or if both 7 and 4 are blocked, then 1 is completed, then 7 becomes unblocked and is completed, then 2 is completed and only after that 4 becomes unblocked and is completed... You get the idea.

    [Interesting side note: even when jobs are delivered out of order, from single destination's point of view the jobs are still delivered in the expected order (that is, FIFO unless there was some preemption involved). This is because whenever a destination queue becomes unblocked (the destination limit allows selection of more recipient entries for that destination), all jobs which have peers for that destination are unblocked at once.]

  2. The idea of the preemption stack at the head of the job list is gone. That is, it must be possible to preempt any job on the job list. For example, if the jobs 7, 4, 1 and 2 in the example above become all blocked, job 3 becomes the current job. And of course we do not want the preemption to be affected by the fact that there are some blocked jobs or not. Therefore, if it turns out that job 3 might be preempted by job 6, the implementation shall make it possible.

  3. The idea of the linear preemption stack itself is gone. It's no longer true that one job is always preempted by only one job at one time (that is directly preempted, not counting the recursively nested jobs). For example, in the example above, job 1 is directly preempted by only job 4, and job 4 by job 7. Now assume job 7 becomes blocked, and job 4 is being delivered. If it accumulates enough delivery slots, it is natural that it might be preempted for example by job 8. Now job 4 is preempted by both job 7 AND job 8 at the same time.

Now combine the points 2) and 3) with point 1) again and you realize that the relations on the once linear job list became pretty complicated. If we extend the point 3) example: jobs 7 and 8 preempt job 4, now job 8 becomes blocked too, then job 4 completes. Tricky, huh?

If I illustrate the relations after the above mentioned examples (but those in point 1)), the situation would look like this:

                            v- parent

adoptive parent ->    1--2--3--5--...      <- "stack" level 0
                      |     |
parent gone ->        ?     6              <- "stack" level 1
                     / \
children ->         7   8   ^- child       <- "stack" level 2

                      ^- siblings

Now how does nqmgr deal with all these complicated relations?

Well, it maintains them all as described, but fortunately, all these relations are necessary only for purposes of proper counting of available delivery slots. For purposes of ordering the jobs for entry selection, the original rule still applies: "the job preempting the current job is moved in front of the current job on the job list". So for entry selection purposes, the job relations remain as simple as this:

7--8--1--2--6--3--5--..   <- scheduler's job list order

The job list order and the preemption parent/child/siblings relations are maintained separately. And because the selection works only with the job list, you can happily forget about those complicated relations unless you want to study the nqmgr sources. In that case the text above might provide some helpful introduction to the problem domain. Otherwise I suggest you just forget about all this and stick with the user's point of view: the blocker jobs are simply ignored.

[By now, you should have a feeling that there is more things going under the hood than you ever wanted to know. You decide that forgetting about this chapter is the best you can do for the sake of your mind's health and you basically stick with the idea how the scheduler works in ideal conditions, when there are no blockers, which is good enough.]

Dealing with memory resource limits

When discussing the nqmgr scheduler, we have so far assumed that all recipients of all messages in the active queue are completely read into the memory. This is simply not true. There is an upper bound on the amount of memory the nqmgr may use, and therefore it must impose some limits on the information it may store in the memory at any given time.

First of all, not all messages may be read in-core at once. At any time, only qmgr_message_active_limit messages may be read in-core at maximum. When read into memory, the messages are picked from the incoming and deferred message queues and moved to the active queue (incoming having priority), so if there is more than qmgr_message_active_limit messages queued in the active queue, the rest will have to wait until (some of) the messages in the active queue are completely delivered (or deferred).

Even with the limited amount of in-core messages, there is another limit which must be imposed in order to avoid memory exhaustion. Each message may contain huge amount of recipients (tens or hundreds of thousands are not uncommon), so if nqmgr read all recipients of all messages in the active queue, it may easily run out of memory. Therefore there must be some upper bound on the amount of message recipients which are read into the memory at the same time.

Before discussing how exactly nqmgr implements the recipient limits, let's see how the sole existence of the limits themselves affects the nqmgr and its scheduler.

The message limit is straightforward - it just limits the size of the lookahead the nqmgr's scheduler has when choosing which message can preempt the current one. Messages not in the active queue simply are not considered at all.

The recipient limit complicates more things. First of all, the message reading code must support reading the recipients in batches, which among other things means accessing the queue file several times and continuing where the last recipient batch ended. This is invoked by the scheduler whenever the current job has space for more recipients, subject to transport's refill_limit and refill_delay parameters. It is also done any time when all in-core recipients of the message are dealt with (which may also mean they were deferred) but there are still more in the queue file.

The second complication is that with some recipients left unread in the queue file, the scheduler can't operate with exact counts of recipient entries. With unread recipients, it is not clear how many recipient entries there will be, as they are subject to per-destination grouping. It is not even clear to what transports (and thus jobs) the recipients will be assigned. And with messages coming from the deferred queue, it is not even clear how many unread recipients are still to be delivered. This all means that the scheduler must use only estimates of how many recipients entries there will be. Fortunately, it is possible to estimate the minimum and maximum correctly, so the scheduler can always err on the safe side. Obviously, the better the estimates, the better results, so it is best when we are able to read all recipients in-core and turn the estimates into exact counts, or at least try to read as many as possible to make the estimates as accurate as possible.

The third complication is that it is no longer true that the scheduler is done with a job once all of its in-core recipients are delivered. It is possible that the job will be revived later, when another batch of recipients is read in core. It is also possible that some jobs will be created for the first time long after the first batch of recipients was read in core. The nqmgr code must be ready to handle all such situations.

And finally, the fourth complication is that the nqmgr code must somehow impose the recipient limit itself. Now how does it achieve it?

Perhaps the easiest solution would be to say that each message may have at maximum X recipients stored in-core, but such solution would be poor for several reasons. With reasonable qmgr_message_active_limit values, the X would have to be quite low to maintain reasonable memory footprint. And with low X lots of things would not work well. The nqmgr would have problems to use the transport_destination_recipient_limit efficiently. The scheduler's preemption would be suboptimal as the recipient count estimates would be inaccurate. The message queue file would have to be accessed many times to read in more recipients again and again.

Therefore it seems reasonable to have a solution which does not use a limit imposed on per-message basis, but which maintains a pool of available recipient slots, which can be shared among all messages in the most efficient manner. And as we do not want separate transports to compete for resources whenever possible, it seems appropriate to maintain such recipient pool for each transport separately. This is the general idea, now how does it work in practice?

First we have to solve little chicken-and-egg problem. If we want to use the per-transport recipient pools, we first need to know to what transport(s) is the message assigned. But we will find that out only after we read in the recipients first. So it is obvious that we first have to read in some recipients, use them to find out to what transports is the message to be assigned, and only after that we can use the per-transport recipient pools.

Now how many recipients shall we read for the first time? This is what qmgr_message_recipient_minimum and qmgr_message_recipient_limit values control. The qmgr_message_recipient_minimum value specifies how many recipients of each message we will read for the first time, no matter what. It is necessary to read at least one recipient before we can assign the message to a transport and create the first job. However, reading only qmgr_message_recipient_minimum recipients even if there are only few messages with few recipients in-core would be wasteful. Therefore if there is less than qmgr_message_recipient_limit recipients in-core so far, the first batch of recipients may be larger than qmgr_message_recipient_minimum - as large as is required to reach the qmgr_message_recipient_limit limit.

Once the first batch of recipients was read in core and the message jobs were created, the size of the subsequent recipient batches (if any - of course it's best when all recipients are read in one batch) is based solely on the position of the message jobs on their corresponding transports' job lists. Each transport has a pool of transport_recipient_limit recipient slots which it can distribute among its jobs (how this is done is described later). The subsequent recipient batch may be as large as the sum of all recipient slots of all jobs of the message permits (plus the qmgr_message_recipient_minimum amount which always applies).

For example, if a message has three jobs, first with 1 recipient still in-core and 4 recipient slots, second with 5 recipient in-core and 5 recipient slots, and third with 2 recipients in-core and 0 recipient slots, it has 1+5+2=7 recipients in-core and 4+5+0=9 jobs' recipients slots in total. This means that we could immediately read 2+qmgr_message_recipient_minimum more recipients of that message in core.

The above example illustrates several things which might be worth mentioning explicitly: first, note that although the per-transport slots are assigned to particular jobs, we can't guarantee that once the next batch of recipients is read in core, that the corresponding amounts of recipients will be assigned to those jobs. The jobs lend its slots to the message as a whole, so it is possible that some jobs end up sponsoring other jobs of their message. For example, if in the example above the 2 newly read recipients were assigned to the second job, the first job sponsored the second job with 2 slots. The second notable thing is the third job, which has more recipients in-core than it has slots. Apart from sponsoring by other job we just saw it can be result of the first recipient batch, which is sponsored from global recipient pool of qmgr_message_recipient_limit recipients. It can be also sponsored from the message recipient pool of qmgr_message_recipient_minimum recipients.

Now how does each transport distribute the recipient slots among its jobs? The strategy is quite simple. As most scheduler activity happens on the head of the job list, it is our intention to make sure that the scheduler has the best estimates of the recipient counts for those jobs. As we mentioned above, this means that we want to try to make sure that the messages of those jobs have all recipients read in-core. Therefore the transport distributes the slots "along" the job list from start to end. In this case the job list sorted by message enqueue time is used, because it doesn't change over time as the scheduler's job list does.

More specifically, each time a job is created and appended to the job list, it gets all unused recipient slots from its transport's pool. It keeps them until all recipients of its message are read. When this happens, all unused recipient slots are transferred to the next job (which is now in fact now first such job) on the job list which still has some recipients unread, or eventually back to the transport pool if there is no such job. Such transfer then also happens whenever a recipient entry of that job is delivered.

There is also a scenario when a job is not appended to the end of the job list (for example it was created as a result of second or later recipient batch). Then it works exactly as above, except that if it was put in front of the first unread job (that is, the job of a message which still has some unread recipients in queue file), that job is first forced to return all of its unused recipient slots to the transport pool.

The algorithm just described leads to the following state: The first unread job on the job list always gets all the remaining recipient slots of that transport (if there are any). The jobs queued before this job are completely read (that is, all recipients of their message were already read in core) and have at maximum as many slots as they still have recipients in-core (the maximum is there because of the sponsoring mentioned before) and the jobs after this job get nothing from the transport recipient pool (unless they got something before and then the first unread job was created and enqueued in front of them later - in such case the also get at maximum as many slots as they have recipients in-core).

Things work fine in such state for most of the time, because the current job is either completely read in-core or has as much recipient slots as there are, but there is one situation which we still have to take care of specially. Imagine if the current job is preempted by some unread job from the job list and there are no more recipient slots available, so this new current job could read only batches of qmgr_message_recipient_minimum recipients at a time. This would really degrade performance. For this reason, each transport has extra pool of transport_extra_recipient_limit recipient slots, dedicated exactly for this situation. Each time an unread job preempts the current job, it gets half of the remaining recipient slots from the normal pool and this extra pool.

And that's it. It sure does sound pretty complicated, but fortunately most people don't really have to care how exactly it works as long as it works. Perhaps the only important things to know for most people are the following upper bound formulas:

Each transport has at maximum

max(
qmgr_message_recipient_minimum * qmgr_message_active_limit
+ *_recipient_limit + *_extra_recipient_limit,
qmgr_message_recipient_limit
)

recipients in core.

The total amount of recipients in core is

max(
qmgr_message_recipient_minimum * qmgr_message_active_limit
+ sum( *_recipient_limit + *_extra_recipient_limit ),
qmgr_message_recipient_limit
)

where the sum is over all used transports.

And this terribly complicated chapter concludes the documentation of nqmgr scheduler.

[By now you should theoretically know the nqmgr scheduler inside out. In practice, you still hope that you will never have to really understand the last or last two chapters completely, and fortunately most people really won't. Understanding how the scheduler works in ideal conditions is more than good enough for vast majority of users.]

Credits