Personal computing discussed

Moderators: renee, Steel, notfred

 
Hz so good
Gerbil Elite
Topic Author
Posts: 768
Joined: Wed Dec 04, 2013 5:08 pm

Cisco "Merchant Silicon"

Thu Feb 20, 2014 3:04 pm

I was looking thru this Gizmodo article about folks starting to use "military-grade" rooftop lasers as a backhaul solution, in order to shave milliseconds off transfer times in trading environments.

That reminded me of something I'd heard in one of the CCNA Data Center courses, about the Nexus 3000 series using "merchant silicon" to achieve ultra-low latency for those types of trading scenarios.

Anybody aware of a book or training manual that goes into detail (ala Odom's books) about how they achieve those nanosecond levels of latency? I looked at the data sheets on Cisco's site, and they mention Algo Boost, Warp mode, Hitless Network Address Translation, etc..., but they don't really go into the level of detail on exactly how those technologies achieve it that i'd like.
 
notfred
Maximum Gerbil
Posts: 4610
Joined: Tue Aug 10, 2004 10:10 am
Location: Ottawa, Canada

Re: Cisco "Merchant Silicon"

Fri Feb 21, 2014 9:02 am

The high end routers and switches (Cisco or any other manufacturer) use custom ASICs. You aren't going to get much beyond marketing fluff terms unless you are working directly on them as these things are competitive advantages between manufacturers. They all know roughly what each other is doing anyway from getting hold of the other ones equipment, beating up on it in testing and reverse engineering, but the details are still secret.
 
Hz so good
Gerbil Elite
Topic Author
Posts: 768
Joined: Wed Dec 04, 2013 5:08 pm

Re: Cisco "Merchant Silicon"

Fri Feb 21, 2014 12:42 pm

notfred wrote:
The high end routers and switches (Cisco or any other manufacturer) use custom ASICs. You aren't going to get much beyond marketing fluff terms unless you are working directly on them as these things are competitive advantages between manufacturers. They all know roughly what each other is doing anyway from getting hold of the other ones equipment, beating up on it in testing and reverse engineering, but the details are still secret.



I'm trying to get as many cisco certs as possible, and I'd imagine that have that info would help with me getting my CCNA Datacenter, since those switches tend to be ToR. Shame Cisco is trying to hold this close to it's vest, since I'm curious if any of those latency reducing techniques would be useful in other DC operations.
 
UberGerbil
Grand Admiral Gerbil
Posts: 10368
Joined: Thu Jun 19, 2003 3:11 pm

Re: Cisco "Merchant Silicon"

Fri Feb 21, 2014 2:56 pm

You might try posting this question over at the RWT forums. Kanter attracts a pretty high-end range of viewers, and a sizable number of the people who post on there do (or have done) actual commercial chip design. Of course they be restricted in how detailed they're allowed to get, but you're more likely to stumble across somebody who knows something concrete.
 
Hz so good
Gerbil Elite
Topic Author
Posts: 768
Joined: Wed Dec 04, 2013 5:08 pm

Re: Cisco "Merchant Silicon"

Fri Feb 21, 2014 3:43 pm

UberGerbil wrote:
You might try posting this question over at the RWT forums. Kanter attracts a pretty high-end range of viewers, and a sizable number of the people who post on there do (or have done) actual commercial chip design. Of course they be restricted in how detailed they're allowed to get, but you're more likely to stumble across somebody who knows something concrete.



Thanks much! I hadn't thought of heading over to RWT. The articles get updated sporadically. so they'd almost dropped off my radar... Was unaware they had a vibrant forum community!


EDIT - Left a post there. Fingers crossed someone can help me, since this could be something beneficial, if I find fulltime work at a datacenter or major Telco providing services to the high-speed trading ecosphere.

2nd EDIT - I'd also left this with the networking gurus at SA. No dice there, either.
 
Flatland_Spider
Graphmaster Gerbil
Posts: 1324
Joined: Mon Sep 13, 2004 8:33 pm

Re: Cisco "Merchant Silicon"

Fri Feb 21, 2014 5:08 pm

Hz so good wrote:
I'm trying to get as many cisco certs as possible, and I'd imagine that have that info would help with me getting my CCNA Datacenter,


The system and hardware engineers at Cisco are going to do all of the heavy lifting with tuning the hardware and software (IOS). People with Cisco certs are going to be plugging boxes together and tuning packet size and network protocols. How the ASICs work is going to be irrelevant. You just have to know Cisco gear has the lowest latency.

Systems programming is where the operation of the chip is really relevant.

Shame Cisco is trying to hold this close to it's vest, since I'm curious if any of those latency reducing techniques would be useful in other DC operations.


Not really and probably not.

This is really specialized hardware, and generally really specialized hardware is useless for anything outside of its design scope.
 
SecretSquirrel
Minister of Gerbil Affairs
Posts: 2726
Joined: Tue Jan 01, 2002 7:00 pm
Location: North DFW suburb...
Contact:

Re: Cisco "Merchant Silicon"

Sat Feb 22, 2014 11:46 am

Hz so good wrote:
notfred wrote:
The high end routers and switches (Cisco or any other manufacturer) use custom ASICs. You aren't going to get much beyond marketing fluff terms unless you are working directly on them as these things are competitive advantages between manufacturers. They all know roughly what each other is doing anyway from getting hold of the other ones equipment, beating up on it in testing and reverse engineering, but the details are still secret.



I'm trying to get as many cisco certs as possible, and I'd imagine that have that info would help with me getting my CCNA Datacenter, since those switches tend to be ToR. Shame Cisco is trying to hold this close to it's vest, since I'm curious if any of those latency reducing techniques would be useful in other DC operations.


As someone else pointed out, you will never be exposed to the secret sauce that makes the ASICs so fast. You're left at the implementation level where most of it is common sense for someone who knows how a big switch works: no L3 routing on the switch, set QOS appropriately, no filtering, all ports at the same speed, etc. When you get down to micosecond latencys, it does help a little to stop and think about how the switch is architected. You don't want to cross line cards. Going from the top line card to the bottom line card in a large switch could add 5ns, just in physics. In fact, if you can stay within the same switching ASIC (usually a group of four ports), even better. You can get really carried away if you feel like it, but even in an high perfomance datacenter you are still only plugging things and twiddling some IOS settings. Unless you are doing HFT, nobody cares because the difference between 150us and 350us latency across a switch doesn't matter anywhere else.

You mentioned the freespace laser link in the beginning. As an aside, that laser link (when it is working) would be about 50% lower latency than a fibre optic link of the same distance and something like 30% lower latency than a copper link of the same distance. It can also be disrupted by a bird....

--SS

Who is online

Users browsing this forum: No registered users and 1 guest
GZIP: On