em forwarding performance (was Proposed 6.2 em RELEASE patch

[Mike Tancsa]

At 12:51 AM 11/30/2006, Nick Pavlica wrote:

>>Did a quick default install. Results are not so interesting since one
>>stream livelocks the box. Basic stats at http://www.tancsa.com/blast.html
>>
>>If there are some OpenSolaris wizards out there who want me to tune,
>>I am happy to retest...
>
>Mike,
>  I'm not an OpenSolaris/Solaris expert, but was curious which build
>you were testing with.

Hi,
         I grabbed the latest DVD bits that were available at the time.
# uname -a
SunOS interlope 5.11 snv_52 i86pc i386 i86pc



>SolarisExpress CE or B52 at the time of this writing.  Of course I
>patched all of these boxes before I did my testing which was mostly
>centered around disk I/O performance on UFS and ZFS, and some
>experimentation with Zones/Containers.

Didnt do any patches. The only thing I did was kill off X and disable 
and enable ipfilter.  Its quite possible there was other cruft 
running that I didnt know about, but like I said, this was my first 
exposure to OpenSolaris so I have no idea if there are things I 
should have set.


>  I'm surprised that the console
>locked up during your tests.

>My limited experience with Solaris 10+
>thus far has been positive in terms of performance and stability.

It does recover afterwards, but pretty well all other processes stop 
as the CPU I guess is pegged dealing with all the interrupts.

Thinking further about my tests, it doesnt really do that great of a 
job of simulating normal real world conditions.  In the real world, 
the packet sizes will vary and the speeds will be all over the 
place.  I am wondering if some of these modern nics have that in mind 
with their design.    But then again, this is sort of the scenario 
when a firewal gets blasted by a high PPS attack :(

>When I have stressed my test systems, they remained responsive and
>seemed to have better  performance than FC6 and Ubuntu6.10 when
>copying large files across my network.

But thats pretty different then my test setup. All the OSes I tested 
can do that no problem :)


>Thanks for digging in with this testing, I hope you keep at it.


Yeah I inadvertently slighted the NetBSD folks by leaving them 
out.  So I guess I better give them a try as well.

The part that really surprises me is the drop in performance as 
firewall rules are added to RELENG_6 and above.  Both LINUX and 
RELENG_4 seem to scale well with the number of rules added but 
RELENG_6 takes a big drop.

         ---Mike