[Beowulf] Intel’s 800Gbps cables headed to cloud data centers and supercomputers

Prentice Bisbal prentice.bisbal at rutgers.edu
Wed Mar 12 09:51:56 PDT 2014


On 03/12/2014 12:22 PM, Mark Hahn wrote:
>> "Intel and several of its partners said they will make 800Gbps cables 
>> available in the second half of this year, bringing big speed 
>> increases to supercomputers and data centers."
>
> this seems very niche to me.  here's a very practical question: how 
> many of your cluster's links are saturated, and which of them
> are both saturated and currently composed of trunks or wide FDR?

I think one benefit  (could be the only benefit) for us would be further 
reducing latency over IB, if it does that. I did call out that there was 
no mention of latency in the article. That could be by design: latency 
isn't really improved, or the author thinks readers only care about 
bandwidth and latency is not important. The latter is probably accurate 
in many IT realms.

As you know, I've been investigating dual-rail IB for an upcoming 
purchase, and from talking to others it seems that saturated IB links 
are very rare, especially with FDR. Many people want big bandwidth for 
'big data', whether the bandwidth is really needed or not.*

I agree with your point, but I still think this is an interesting 
development, since we need to keep pushing the limits to move forward 
over the long-run.

* I put 'big data' in quotes to show my contempt for the term.

>
> I speculate that although it would be natural for larger clusters
> to need higher-bandwidth backbones, the majority of large clusters
> are, in fact, not used for giant runs of a single job, and therefore
> mostly experience more locality-friendly BW patterns.  (a counter
> to this would be workloads that spend much time doing all-to-all,
> but I've never been clear on whether that's laziness or necessary.)
>
> but basically, this is just a very dense cable.  what would make more
> of a difference is a breakthrough in the electro-optics to make a 25 
> Gb bi-directional link cheap.  today, even 10G optics run about 
> $1k/link, which is a big problem if the device at the endpoint is 
> likely only worth $1-2k.  (ie, optics are a non-starter for nodes that 
> are not pretty fat, and most of the action seems to be either at the 
> traditional 2s level of plumpness or lower.)
>
>> "US Conec established an MXC certification program to help other 
>> companies sell the cables. Tyco Electronics and Molex are the first 
>> besides Corning to announce that they will build and sell MXC cable 
>> assemblies."
>
> which is, of course, the least interesting thing :(
>
>> So it sounds like there will be competition for the cables, but what 
>> about the NICs and switches? Will Intel have a monopoly on that, or 
>> will this be a standardized technology that will allow other 
>> manufacturers to make their own silicon/complete products?
>
> IMO, there's a strong smell of "build it and they will come" to this.
>
> OTOH, optical usually gets away without dramatic serialization/EC 
> overheads.
>
>> Years ago (the late 90s?) I read an interesting magazine article 
>> about Intel and why they started making their own NICs, graphics 
>> processors, etc. According to the article, Intel was content to let 
>> 3Com and others make networking gear, but when network speeds weren't 
>> increasing fast enough,
>
> usually, I think of most Intel moves as promoting a lower-friction 
> industry.
> standards they create/sponsor/push, like power supplies, MB specs, IPMI
> tend to be good for everyone, and pull vendors away from the kind of
> customer-hostile lockin they (vendors) love so much.  it's hard to tell
> how much is self-interest, of course, since Intel manages to take a 
> pretty
> big bite of everything.
>
> my memory, though, is that Intel didn't make much of a difference to 
> the Gb transition.  most of the hardware I experienced from that era 
> was from
> Broadcom, with a few trivial vendors (some of which are still around,
> though still insignificant).
>
>> Intel got into the game because without increasing network speeds, 
>> there wasn't much of a need for faster processors. We all know that 
>> Intel has bought QLogic and is spending a lot money on high-speed 
>> interconnects.
>
> are they?  I see the occasional little-ish product-ish thing pop up,
> but not much vision.  thunderbolt appears to be a cul-de-sac to make 
> Apple
> happy (they get off on being "special"...)  IB doesn't appear to be going
> anywhere, really (it'll always be a weird little non-IP universe.)
> this new connector (just a connector) looks exactly like "big version 
> of the original optical thunderbolt".  how about pcie-over-whatever - 
> is that going anywhere?  then there's all that blather about 
> dis-aggregated servers (what about latency?!?).
>
>> Following the logic of that article, I guess Intel realized you can't 
>> sell truckloads of processors if your don't have an interconnect that 
>> makes it worthwhile.
>
> I think this connector is a solution in search of a problem.
>
> if they can make a 800Gb connector that adds $100 to a motherboard, 
> and plugs into switch ports that cost $100/port, they could move some 
> product.
> even 80Gb would be epic.  even 10Gb would shake up the market bigtime...
>
> regards, mark hahn.




More information about the Beowulf mailing list