Re: Max throughput ?
* Michael Gale [EMAIL PROTECTED] [2007-09-05 00:16]: Hey, It was suggested that we create an OpenBSD server with 9GB interfaces to start. 7 Will be used right off the bat. This would function as a core router brining 7 GB networks together on the inside of a main firewall. I suggested that maybe we would have some bandwidth issues with trying to push that much traffic through a single server. you might have thruput issues, you might not. depends on the traffic characteristics and hardware you choose. Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. sure it is better, assuming you call I paid $100,000 for a $5 CPU that falls over at 5000pps* better. *when the packets are just a tiny bit different from what cisco expects and can handle in the fast path, they go to the main cpu, which is incredibly slow on pretty much any cisco you can buy -- Henning Brauer, [EMAIL PROTECTED], [EMAIL PROTECTED] BS Web Services, http://bsws.de Full-Service ISP - Secure Hosting, Mail and DNS Services Dedicated Servers, Rootservers, Application Hosting - Hamburg Amsterdam
Re: switch or server? (was Re: Max throughput ?)
* David Newman [EMAIL PROTECTED] [2007-09-05 00:59]: Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. Most el cheapo gig switches will do the job without packet loss. you are beeing tricked by marketing terminology. layer 3 switches are routers. vendors use the term to.. well I dunno :) most so-called layer3 swicthes are regular layer 2 switches with a little extra logic to be able to inspect IP headers and take the switching (it is routing of course) decision based on that. Rule of thumb: they all suck. -- Henning Brauer, [EMAIL PROTECTED], [EMAIL PROTECTED] BS Web Services, http://bsws.de Full-Service ISP - Secure Hosting, Mail and DNS Services Dedicated Servers, Rootservers, Application Hosting - Hamburg Amsterdam
Re: switch or server? (was Re: Max throughput ?)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 9/5/07 2:01 AM, Henning Brauer wrote: * David Newman [EMAIL PROTECTED] [2007-09-05 00:59]: Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. Most el cheapo gig switches will do the job without packet loss. you are beeing tricked by marketing terminology. layer 3 switches are routers. vendors use the term to.. well I dunno :) most so-called layer3 swicthes are regular layer 2 switches with a little extra logic to be able to inspect IP headers and take the switching (it is routing of course) decision based on that. Rule of thumb: they all suck. That's a statement of value, not of fact. The OP asked about switch throughput. Even the el cheapo ones you describe as sucky can forward packets at line rate with zero loss. They have many other problems -- execrable routing code, CLIs and GUIs written by idiots, and horrible hashing algorithms, to name a few -- but basic packet forwarding isn't one of them. That said, I share your allergy to the term layer-3 switch. I don't use this meaningless marketing term. Switches switch; routers route. dn iD8DBQFG3swDyPxGVjntI4IRAkqkAJ93LmSLnpTft6j/sOZ/0bbdeBuSdQCfWENS gEH1SSQe1g0dxOaYp/+p+68= =loeJ -END PGP SIGNATURE-
Re: Max throughput ?
2007/9/5, David Newman [EMAIL PROTECTED]: the case. Spending US$100k on a switch from Cisco, Foundry, or Force10 will get you fast-path processing in the tens of millions of pps or more (which AFAIK even the studliest of server hardware doesn't do today) and Which reminds me: Is there a real chance that we can expect 4.4 to run good on a SUN T2 with support for the 10G NICs? Best Martin
Re: Max throughput ?
Which reminds me: Is there a real chance that we can expect 4.4 to run good on a SUN T2 with support for the 10G NICs? Well, kind of difficult since we don't have any.
Re: switch or server? (was Re: Max throughput ?)
* David Newman [EMAIL PROTECTED] [2007-09-05 17:51]: -BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 9/5/07 2:01 AM, Henning Brauer wrote: * David Newman [EMAIL PROTECTED] [2007-09-05 00:59]: Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. Most el cheapo gig switches will do the job without packet loss. you are beeing tricked by marketing terminology. layer 3 switches are routers. vendors use the term to.. well I dunno :) most so-called layer3 swicthes are regular layer 2 switches with a little extra logic to be able to inspect IP headers and take the switching (it is routing of course) decision based on that. Rule of thumb: they all suck. That's a statement of value, not of fact. The OP asked about switch throughput. Even the el cheapo ones you describe as sucky can forward packets at line rate with zero loss. switch, aka layer 2, yes. route, aka layer 3, no. not even under perfect conditions in case of teh small ones. -- Henning Brauer, [EMAIL PROTECTED], [EMAIL PROTECTED] BS Web Services, http://bsws.de Full-Service ISP - Secure Hosting, Mail and DNS Services Dedicated Servers, Rootservers, Application Hosting - Hamburg Amsterdam
Re: Max throughput ?
* David Newman [EMAIL PROTECTED] [2007-09-05 17:40]: -BEGIN PGP SIGNED MESSAGE- Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. sure it is better, assuming you call I paid $100,000 for a $5 CPU that falls over at 5000pps* better. *when the packets are just a tiny bit different from what cisco expects and can handle in the fast path, they go to the main cpu, which is incredibly slow on pretty much any cisco you can buy Here you are referring to slow-path processing for packets with IP options set. That's normal with all switches, not just Cisco's. yep. but basicaly everybody else has faster host CPUs - so they still suffer, but they don't go down as badly. This also suggests 5000 pps is the expected performance, which is not the case. Spending US$100k on a switch from Cisco, Foundry, or Force10 will get you fast-path processing in the tens of millions of pps or more (which AFAIK even the studliest of server hardware doesn't do today) and slow-path processing in the 1s of pps or more. no, I have fixed networks by removing $100k cisco gear that was falling over under way less than 5k pps. OTOH I fully agree that lower end boxes (and even some higher ones such as older Sup cards on Cat 65xxs) have relatively slow CPUs. i have yet to see a cisco box where the host CPU is not pathetically slow. The key question is whether you have slow-path traffic to begin with. your slow-path traffic is a perfect attack vector... and some stuff goes slow-path that you totally would not expect to. anyway, this is not a cisco list, so no point in discussing their design fuckups here. -- Henning Brauer, [EMAIL PROTECTED], [EMAIL PROTECTED] BS Web Services, http://bsws.de Full-Service ISP - Secure Hosting, Mail and DNS Services Dedicated Servers, Rootservers, Application Hosting - Hamburg Amsterdam
Max throughput ?
Hey, It was suggested that we create an OpenBSD server with 9GB interfaces to start. 7 Will be used right off the bat. This would function as a core router brining 7 GB networks together on the inside of a main firewall. I suggested that maybe we would have some bandwidth issues with trying to push that much traffic through a single server. Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. -- Michael Gale Nothing is impossible to a willing mind. - Monk Hae Chang
switch or server? (was Re: Max throughput ?)
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 9/4/07 3:03 PM, Michael Gale wrote: Hey, It was suggested that we create an OpenBSD server with 9GB interfaces to start. I think here you mean 9 1-Gbit/s interfaces 7 Will be used right off the bat. This would function as a core router brining 7 GB networks together on the inside of a main firewall. I suggested that maybe we would have some bandwidth issues with trying to push that much traffic through a single server. RFCs 2544 and 2889 define router and switch test methodologies. A related document, RFC 1242, defines throughput as the maximum zero-loss rate. Note that throughput is a single rate. Ergo, there's no such thing as max or min or any other kind of throughput. There's just throughput. Can any one comment on this ? Would it not be better to use some think like a Cisco layer 3 GB switch. Most el cheapo gig switches will do the job without packet loss. Manageability, routing, an sshd server, redundant power, support, etc., cost extra. Commercial switches achieved line-rate, zero-loss performance around a decade ago, with small-frame latency and jitter in the tens of microseconds. These use ASICs or FPGAs or NPs to get there. Big studly servers equipped with 10G interfaces currently achieve goodput somewhere north of 1G but south of 10G with higher latency and jitter than switches. I'm not aware of anyone getting loss-free performance at N-Gbit/s (where N 7) using server hardware alone. dn iD8DBQFG3eCTyPxGVjntI4IRAqu8AKDotF/6ReuA+V/L2Z6Ng7f8tbCpQgCg1YR4 4g+vFsK6cmph88YQGnrXl54= =0N3R -END PGP SIGNATURE-
