INN with hyperthreading/SMP

[Mike Zanker]

On 11 May 2004 08:04 +0000 Miquel van Smoorenburg 
<list-inn-workers at news.cistron.nl> wrote:

> It's not the kernel. You built INN with largefile support. Lots of
> internal (in-memory) and external (on-disk: history file hashtable)
> structures now double in size because of 32/64 bit issues.

Aha, that makes sense. I should have mentioned that it was built with 
largefiles on the reader box, too. I ran pmap on both boxes and it 
seems that the difference lies in the size of the history hash and 
index.

On the reader box the hash and index come to around 50MB - on the new 
feeder hash+index is 140MB.

> Another thing is that you rebuilt INN and it may now use mmap() in
> places where it used read()/write() before. That shows up under VSZ
> in the ps aux output, but that's quite misleading. mmap()ed memory
> can be compared with disk cache - if the memory is needed for
> other purposes, the mmap() dirty pages will be written to disk.
> That's the difference between VSZ and RSS.

VSZ is about the same size as hash+index, RSS is around 70MB at the 
moment but has been as large as 350MB. Does memory used also depend on 
the size of CNFS buffers? At the moment I'm using two 8GB raw 
partitions as buffers (SEQUENTIAL because they're on the same physical 
disk).

Mike.