CUPS job processing time

[John A. Murdie]

I've been doing some timing of jobs being processed by our CUPS 1.3.7 installation on a QEmu (0.9.0) virtual machine which uses one core of a dual quad-core Intel Xeon application server. The virtual machine has 2Gb of the available 16Gb physical memory. Other virtual machines on this hardware run MySQL, LDAP, Kerberos servers. A Netapp Filer is used for the spool directory, though we'd normally use a faster though less robust filestore, except that the machine serving that was out of commission for a time, so the Filer was used.

The motive here is to speed up perceived print times from Windows PC clients via the Linux based CUPS system - overall print time for a single page is about 70-75 seconds. We'd prefer 30-40 seconds for one page (job setup time in all the nodes along the print path adds 20-30 seconds or so to all jobs, and many, many, of the files our admin staff print are just one page).

So that we have a single queue for each printer (the Linux and Windows users all look at the CUPS web page), we direct jobs from our Windows Server 2003 to cups-lpd (command 'cups-lpd -o document-format=application/vnd.cups-raw'). We'd like to have the Windows clients have IPP print queues published to them from the Windows server, but apparently Windows Server 2003 is not capable of that.

Incidentally, even with 'raw', I note that page_log still contains page counts for Windows -submitted files. Is there a 'raw raw' which will prevent even this examination of the input files? (We do page counting with an SNMP-based backend wrapper somewhat like PyKota.)

We also have Linux PC clients, whose print time I have not yet measured.

Typically, a one-page job takes about 18 seconds to get from a Windows PC client to CUPS, spends about 35 seconds in CUPS, and takes 20 seconds in the (in this case, slow HP LaserJet P2015) printer.

It may be that much of the contribution to the CUPS processing time is from the Netapp filer - but I'm wondering why CUPS should take 35 seconds to process a single page job - is this expected with our overall hardware/software? - or could we knock 30 seconds off this time with a little care in configuration? It doesn't seem very much absolute time, but would be a large percentage, and would improve matters psychologically for the users of our small desktop printers.

John A. Murdie