nanog mailing list archives
Re: Open source hardware
From: Saku Ytti <saku () ytti fi>
Date: Fri, 3 Jan 2014 15:33:56 +0200
On (2014-01-03 07:48 -0500), Ray Soucy wrote:
Juniper is a FreeBSD shop, and Cisco's new OS lines are based on Linux. Ciena is largely based on Linux as well. In poking around at these platforms recently one of the big things I'm noticing is that there is a lot less done in hardware than we traditionally saw, especially from Cisco.
I'm not sure which platforms you refer to. But if we look at SP segment we're talking about JNPR M, MX, T, PTX or Cisco ASR9k, NCS6k, CRS-1. JNPR is indeed FreeBSD, but FreeBSD is used very sparsely, to boot box up and to run RPD, which is essentially router-control-plane-in-a-process, it runs all routing protocols and configures hardware. ASR9k, CRS-1 run IOS-XR on QNX and NCS6k on Linux and there at least Cisco capitalizes on OS scheduling, it's not single fat process on top of OS. All of these boxes do all packet pushing in NPU (ezchip, trio, ichip...) For IOS XE boxes, it's almost same as JNPR, except instead of single process single threaded RPD, IOSd is actually running several threads.
by under-sizing and over-pricing their CPUs for years. But when you have a modern server-grade platform, multi-Gigabit performance, even with significant levels of packet processing and small packet sizes, is a joke. So at least for the low end of the spectrum there is a huge savings for equal (often better) performance.
Low end has always been using COTS CPU, RISC, PPC etc, so not much has changed there. For low end, linux pc can be competitive in some applications.
With the new Intel DPDK stuff, Intel is claiming 80M PPS performance on a standard Xeon platform: http://www.intel.com/content/www/us/en/intelligent-systems/intel-technology/packet-processing-is-enhanced-with-software-from-intel-dpdk.html
DPDK is super interesting and it shows Intel is looking at the NPU market, unfortunately these numbers have nothing to do with real-life application, lookup against million+ routes, ACLs, QoS etc. But maybe not in too distant future x86 Intel is usable as NPU, Intel seems to be looking NPU market demands when designing new x86 chips. Right now, if you need perfomance, you're going to have to buy something like bcom chip and then cumulusnetworks linux on top of it, it's as close to 'open source' as you're going to get with good performance. And this is more or less DC stuff, SP market needs more intelligent chips than those ASICs, and I don't think there anything 'open source' in the market place for NPU stuff. -- ++ytti
Current thread:
- Re: Open source hardware, (continued)
- Re: Open source hardware Jimmy Hess (Jan 02)
- Re: Open source hardware Daniël W . Crompton (Jan 03)
- Re: Open source hardware Darren Pilgrim (Jan 03)
- Re: Open source hardware Arnd Vehling (Jan 03)
- Re: Open source hardware Daniël W . Crompton (Jan 04)
- Re: Open source hardware Arnd Vehling (Jan 05)
- Re: Open source hardware TGLASSEY (Jan 06)
- Re: Open source hardware Daniël W . Crompton (Jan 04)
- Re: Open source hardware Ray Soucy (Jan 03)
- RE: Open source hardware Raymond Burkholder (Jan 03)
- Re: Open source hardware Saku Ytti (Jan 03)
- Re: Open source hardware Benno Overeinder (Jan 04)
- Re: Open source hardware Saku Ytti (Jan 04)
- Re: Open source hardware Nick Hilliard (Jan 04)
- Re: Open source hardware Mark Tinka (Jan 04)
- Re: Open source hardware Vlade Ristevski (Jan 07)
- Re: Open source hardware Ray Soucy (Jan 07)
- Re: Open source hardware Ray Soucy (Jan 08)
- Re: Open source hardware Saku Ytti (Jan 08)
- Re: Open source hardware Brandon Ross (Jan 08)
- Re: Open source hardware Saku Ytti (Jan 09)