[cups] tuning CUPS for centralized printing

[Tim Mooney]

All-

I'm looking for tuning suggestions for what other sites tune when running
a central CUPS server, especially universities that use CUPS for printing
in their computer labs.

The tl;dr version:

What are other universities using for MaxClients and MaxClientsPerHost?
What other tunables have you found important to set for your environment?

The full version:

Our print server is a dual hexacore (+hyperthreading) Xeon E5-2620, with
32 GiB RAM.  It's running RHEL 6.x, which means the 2.6.32 kernel and
CUPS 1.4.2 (+ vendor sauce, as Michael would say).

We have about 130 queues defined within CUPS.  All job submission is done
via ipp.  Michael's advice last year steered me away even considering
cups-lpd.  All queues are set to hold the print job once it is submitted.
We use additional print management software (in this case, GoPrint)
that integrates with our campus card system that allows students and
faculty to select their held print jobs.  It's that software that
issues the Move-Job/Release-Job commands to CUPS, to cause the job
to be sent (socket to port 9100) to one of our 60+ HP printers.

The print management software also automatically discards
(Delete-Job) print jobs that aren't released within a period of time,
so that we don't have jobs sitting on the system for more than a few hours.

We have close to 600 Windows PCs (mainly Windows 10 Pro) and about 100
Macs (OS X 10.11, El Capitan) in those labs.  In addition, as much as
possible, we try support "bring your own device" (BYOD), which amounts to
a lot of Windows 7, 8, 8.1, and 10 laptops and all of the versions of OS X
that are still receiving updates from Apple.  Obviously, tablets and
phones are also hugely popular, but for these purposes I'm mainly
interested in tuning for traditional desktop & laptop use.

During peak periods of the day and peaks in the academic cycle, most of
our lab systems will be in use, though obviously not everyone is printing
at once.  In the past, before we switched to using CUPS as our print
spooler, we had seen peaks of between 35K and 40K print jobs submitted
to the print server per day, though not all of those jobs were actually
selected and released to a printer.

With that lengthy and hopefully useful information about the
environment, now the questions.

1) Does anyone have a rough idea how much RAM is used for each
concurrent client connection?  I'm trying to get an idea of how high
I can safely set MaxClients while still leaving plenty of RAM free
for filesystem cache and other processes, including the print management
software.

2) What's a reasonable setting for MaxClientsPerHost?  I currently have it
set to 16, but we're seeing log entries indicating that some clients are
attempting to initiate more than 16 concurrent connections.  Under what
conditions is that "normal"?

3) I have

 	MaxJobs 0
 	PreserveJobHistory Off
 	PreserveJobFiles Off
 	ErrorPolicy retry-job
 	JobRetryInterval 60
 	JobRetryLimit 60

set.  What other tunables, beyond MaxClients/MaxClientsPerHost, should I
be looking at (and why, if it's not obvious from the CUPS docs)?

4) Any other system tuneables (things like net.core.rmem_max,
net.core.somaxconn, etc.) that you've had to tune for a busy CUPS server?

5) Any client-side settings that you've found that make Windows or Mac
clients work more reliably with a CUPS server?

I appreciate any suggestions people might have, especially people that are
running a busy, centralized CUPS server in a large campus environment.

Thanks,

Tim
-- 
Tim Mooney                                             Tim.Mooney at ndsu.edu
Enterprise Computing & Infrastructure                  701-231-1076 (Voice)
Room 242-J6, Quentin Burdick Building                  701-231-8541 (Fax)
North Dakota State University, Fargo, ND 58105-5164