[openbeos] Re: Quick query about network benchmarks

  • From: Brian Verre <bverre@xxxxxxxxxxxxxx>
  • To: <openbeos@xxxxxxxxxxxxx>
  • Date: Sat, 17 Mar 2007 10:43:59 -0500

Zenja-

For your first run through on a by the book network stack, it seems you are 
doing great. Further optimizations may reveal themselves over time; creating 
more user space drivers is surely moving the ball forward for Haiku. You will 
likely discover implementations in the current kernel that could be improved 
upon in the process!

Keep on hacking!

Regards,
Brian



From: Zenja Solaja
Sent: Sat 3/17/2007 4:02 AM
To: openbeos@xxxxxxxxxxxxx
Subject: [openbeos] Re: Quick query about network benchmarks


Sorry for responding to my own post, but when I remove all debug output from 
the code, I get a much more reasonable ping response time of 0.18ms (matches 
Zeta stack, which runs in kernel space).  Running TOP shows a CPU user space 
code at 4.6 %, with a Zeta reference of 3.0, so I guess that that is penalty of 
running a stack from user space.

Now to chapter 12 - UDP.





On 3/16/07, Zenja Solaja <solaja@xxxxxxxxx> wrote: Hi everyone.
More for experimental purposes than anything else, I've picked up Douglas 
Comers books on networking last week, and I'm playing with a from scratch 
net-stack, running in user space under Zeta.  I'm trying to create a 
single-memcpy stack (from driver to application buffer in user space), even for 
fragmented IP packets.  So in a week since I've started, I've got to chapter 9 
which explains ICMP.  For a user-land stack, I get a ping of 0.54ms.  Just for 
reference, the Zeta stack in kernel space gives me 0.18ms for the same 
connection.

Out of curiosity, does anyone remember the ping speed to a machine on the local network using user space networking. Sadly I still cannot get my USB keyboard working under Haiku so I cannot perform the experiment myself.
Cheers everyone.











Other related posts: