Re: Question about nmbclusters

[David Young <dyoung%pobox.com@localhost>]

On Thu, May 23, 2013 at 11:11:04PM +0200, Konrad Neuwirth wrote:
> Okay, this is getting more puzzling. 
> Am 23.05.2013 um 03:50 schrieb yancm:
> 
> > On 2013-05-22 19:07, Konrad Neuwirth wrote:
> >> Hello everyone,
> > I am pretty sure you will need to compile this into a new kernel.
> > I have used:
> > options         NKMEMPAGES=98352
> > options         NMBCLUSTERS=65568
> 
> I did. We are running the new kernel now, and it contains the following: 
> 
> options         NKMEMPAGES=98352
> options         NMBCLUSTERS=262140

This probably won't help with the particular problem at hand, but I will
make a few general observations about clusters.

262140 clusters is about 512 megabytes.  Typically no less than a
quarter of it is wasted because the kernel loads the 2kB buffers with
1500-byte ethernet frames.  I figure there's a lot more waste than
that because you may have hundreds or thousands of clusters dormant on
receive rings, clusters containing runty packets, etc.

Ignoring waste and overhead for a minute, 512 megs is enough memory to
buffer more than 4 seconds of packets if you're receiving them at 1
gigabit/second.  If your box is just a router and it's not forwarding
more than 1 Gb/s, you don't ordinarily need or want to buffer so many
packets.

For a box with many active socket servers, it may be reasonable to
buffer >4 seconds of packets, however, it seems to me that the kernel
should shift packets from the kernel's maps & limits and onto the user
servers' maps & limits much more aggressively than it does, today.

So you see there is lots of room for improvement.

Dave

-- 
David Young
dyoung%pobox.com@localhost    Urbana, IL    (217) 721-9981