[Mailman-Users] remove

Iain Dooley mail at iaindooley.com
Tue May 10 20:13:22 CEST 2005

Brad Knowles wrote:
> At 3:27 PM -0700 2005-05-09, Tyler Strickland wrote:
>> That's entirely possible - the system is an old sun running sendmail.
> 	Older versions of sendmail can certainly have some issues in this 
> respect.  More modern versions work much better, but still require 
> some additional configuration changes in order to get them to work 
> better with large mailing lists.
>> I'm currently building a new mail server to replace it that should be
>> able to handle the load much better by running postfix on a
>> hyperthreaded Pentium 4.
> 	See the stuff in the FAQ about "performance".
> 	For one, HyperThreading almost always hurts performance and does 
> not help.  This is a seriously dain-bramaged idea that Intel had that 
> never panned out, although they did get enough people interested in 
> the concept that once AMD came out with a real dual-core chip, 
> performance really has improved significantly.
> 	For two, you don't need CPU.  You need RAM and disk I/O capacity. 
> Not disk space, but I/O capacity.  When an MTA receives a mail 
> message, it creates one or more temporary files to store that 
> message.  When the message has been delivered, the temporary files 
> are deleted.  Between those two points, there may be many more file 
> creations, deletions, and renaming operations just for that one 
> message.
> 	Each of these types of operations are known as Synchronous 
> Meta-Data Operations, and they require that the entire directory be 
> locked against all other changes during the process of that 
> operation.  You would think that this would not hurt performance very 
> much, but think of a turnstile going into a football stadium -- no 
> matter how fast it operates, the simple fact that it only allows 
> through one person at a time will mean that the total throughput is 
> greatly reduced.
> 	A fast intelligent hardware RAID array with a large quantity of 
> battery-backed write-back RAM cache and high-speed internal 
> interfaces to high-speed drives with intelligent disk queueing (i.e., 
> SCSI with Tagged Command Queueing, or good SATA drives that actually 
> properly implement a similar feature) can really help.  A fast 
> filesystem which optimizes synchronous meta-data operations can also 
> help.
> 	But all this is covered in the FAQ, and you should read about it 
> when you search for "performance".
> 	I will tell you that if you're trying to build a high-performance 
> mail/mailing list server, about the worst possible mistake you could 
> make would be to try to build the machine without first reading the 
> information on this subject in the FAQ, and reading the various 
> pieces of documentation that are referenced.

More information about the Mailman-Users mailing list