On 12/21/2014 3:13 AM, Jonathan K. Tullett wrote: > Greetings, > > I've been using Postfix for many years - since about 2002 - and I've > finally come across a problem I've not been able to resolve by > searching online, or from tapping into my personal network. So I > have come to you all for help. > > I have two machines: > Machine A: My primary 8 core Xeon 2.27GHz, 24Gb RAM primarily > running Postfix 2.6.6 (SLES 6.6 distro) > Machine B: A test box 16 core Xeon 2.2Ghz w/16Gb RAM, primarily > running Apache, RabbitMQ, MemcacheD and finally Postfix 2.9 (Ubuntu > 12.04). > > Machine A is used to distribute to a couple of double-opt-in mailing > lists a week, total recipients between 30,000 and 180,000. The > 'sendmail' binary is used to inject messages into the queue from a > distribution manager. > > The setup on the server is simple: Postfix with two header checks to > prepend a List-Unsubscribe and a Precedence header, and using an > OpenDKIM milter for Domain Keys signing. Nothing other than that. > > Prior to the last week of October using the distribution manager, it > was possible on machine A to inject around 25 messages (full size - > about 70k each) a second into the maildrop queue. > > Since the end of October, that number has dropped to 16 a second on > a good day. > > I wrote a test script (basic for-loop which sent a 1 line, 500 byte > email) and disabled the milters (OpenDKIM and header_checks), it > took 12.75 seconds to inject 500 messages onto Machine A. > > As a test, I ran exactly the same script on Machine B. It injected > 1000 messages (about 500 bytes in size) into the maildrop queue in > 4.95 seconds. > > (I appreciate Machine B is slightly higher spec, but I wouldn't > expect such disparity!) > > I ran qshape during the last mailing on machine A, and the machine > was able to send mails out as fast as it received them; there was no > congestion in any of the queues (maildrop, incoming, outgoing, etc). > > I have no machine stats prior to October - I only came onto the > project last week - do I don't know what (if anything) changed on > that week to cause performance to drop so suddenly. > > I have run read/write tests on both disks - Machine A and B do about > 500Mb/second reads, and 380Mb/second writes; all looks OK. > > I'm not sure why SLES 6.6 was chosen as it was a new build in > August, but know only Postfix 2.6.6 is officially available in the > repo for that distribution. I have 2.11.3 built and ready to go on > that machine but would prefer not to just upgrade on the off-chance > it'll 'fix' the problem when there may be something I'm missing > entirely. > > Have there been huge improvements to the efficiency of the code base > between 2.6 and 2.9 (or 2.11)? Does anyone have suggestions on > where else I can look for the cause? > > Thank you in advance for any help you can provide. > > -- > Jonathan K. Tullett
Are you sure you're comparing apples to apples? The sendmail command is limited by disk performance, with fairly modest CPU requirements. Check your disk subsystem. -- Noel Jones