This is a really interesting thread.
>From my novice perspective, I wonder if the interrupt load actually makes a >difference on the performance of OpenBGPd on different hardware as bge or em. I always assumed different NICs and drivers had different "behaviours", characterized, for instance, by the features supported or interrupts per second handled. My question, if not too offtopic, is if we can use the interrupt load as a measure of NIC throughput. My uninformed guess is that its not. If its too offtopic, sorry for the noise. PS: nice to see this thread alive. OpenBGPd threads are always enriching --- -----Original Message----- From: Chris Cappuccio <ch...@nmedia.net> Sender: owner-m...@openbsd.orgdate: Fri, 16 Jul 2010 09:29:36 To: James Reid - McLean<james.r...@spacenet.com> Cc: <misc@openbsd.org> Subject: Re: Openbgpd Max Number of Neighbors per Instance FWIW, with whatever older chips I've tested with, the interrupt mitigation on the bge driver seems to be configured a bit more aggressive than on em..I see interrupt counts from bge that are 1/2 to 1/4th the count vs em for the same traffic. Both drivers support a broad range of features like hardware TCP/IPv4 checksumming, vlan tagging, ... James Reid - McLean [james.r...@spacenet.com] wrote: > I don't expect the traffic levels to reach Gigabit levels so I doubt I > will ever come close to hitting any sort of limit on the interfaces, but > would I have better support with Intel chipsets over Broadcom? > > Is there a preferred Ethernet chipset for this type of setup? > > James D. Reid > Spacenet Inc. > Network Engineer > Office: (703) 848 - 1266 > > -----Original Message----- > From: owner-m...@openbsd.org [mailto:owner-m...@openbsd.org] On Behalf > Of Claudio Jeker > Sent: Monday, July 12, 2010 5:30 PM > To: misc@openbsd.org > Subject: Re: Openbgpd Max Number of Neighbors per Instance > > On Mon, Jul 12, 2010 at 03:07:23PM -0400, James Reid - McLean wrote: > > Does anyone have information about the maximum number of BGP neighbors > a > > single instance of OpenBGPD could support assuming the following: > > > > 1. OpenBGPD would send only "Default Route" to each neighbor > > 2. Each neighbor would advertise only 1 subnet to OpenBGPD > > 3. OpenBGPD could run in passive mode for all of the connections > > 4. OpenBGPD running on new/current/modern fully supported hardware > with > > no other services running. > > > > I am looking to scale this configuration to support between 500 - > 10,000 > > peers and I need to know how much hardware I would need to purchase to > > support this. > > > > Nobody ever tested 10k peers but here are some tips. Get a box with > 3-4GB > of RAM. Do not run i386 (amd64 has less kvm restrictions and you will > need > a lot of kernel memory). Increase kern.maxclusters to 4-8 times the max > number of sockets you expect and don't forget to increase kern.nfiles. > > Expect to hit a few other issues as well. I know of people doing tests > with 500-1000 sessions that actually injected a few routes. But limiting > bgpd to only announce a default route should reduce the load on the RDE > massivly. > > good luck > -- > :wq Claudio > > > > ------------------------------------------------------------------------ > Spacenet Inc. Notice and Disclaimer > IMPORTANT: This e-mail along with any attachment(s) is intended for the > above named addressee(s) only, and may contain information which is > proprietary, confidential and/or privileged. If you are not the intended > recipient or if you received this e-mail transmittal in error, please be > advised that any review, copying, use, distribution or dissemination of > this e-mail and any attachment(s) is strictly prohibited. Please > immediately notify the sender by e-mail or to phone number 703-848-1000 > and delete this e-mail and any attachments. Thank you. > ------------------------------------------------------------------------ -- I know nothing except the fact of my ignorance -Socrates
