HPC @ Cisco?

Cisco announced a new product today, code named California. You can read some of the info here. This system, also known more formally as a “Unified Computing System” aims to integrate computing, networking and storage into a single managed system. Cisco appears to be aiming for what it believes to be a sweet spot in virtualized infrastructures.
Their play appears to be focused upon virtualization. They tied in a slew of players on the software stack side, and Accenture on the services side.
Ok, but is it an HPC system or an HPC play? Could it be?

This I don’t know yet. Cisco’s pages are short on gory details. You can see elements of it on some of their pages.
It seems they are missing storage. Hey Cisco, we have a great product here, and it would integrate real well there …
(yeah, I know, wishful thinking …)
Is it an HPC play? First guess is no, as its costs will be out of line with traditional HPC costing. So there would be a sizable premium to use this over traditional HPC gear.
Is this an original play? Not in HPC, but apparently it is in the non-HPC world. The large IT vendors, Dell, IBM, HP, etc all have elements of this they tie together with services. So you can buy your Dell blades, your EMC (or Scalable Informatics!) storage, and other things … and pay for integration. Or you buy a pre-integrated “push this button” thingy and off you go.
But could it be an HPC play?
A few months ago I wrote:

But they need special sauce as well. Which is why the Cisco non-announcement, and my speculation about someone grabbing ScaleMP is so interesting. ScaleMP is differentiation. In the era of virtualization-gone-wild, ScaleMP is the counter to this ??? SMP built on demand.

Imagine you have one of these nice California UCS sitting around. And you get a project that comes in, that needs, oh, I dunno, a 32-way server and huge ram to complete. Fire up vSMP, and off you go.
Done with that project, and now time to do something else … tear it down (virtually) and do something else.
This of course depends upon Shai and team getting vSMP going over 10 GbE and the Cisco fabric (no I don’t know if they have, though I’d be willing to bet that Cisco is talking … or should be talking … to Shai about this).
Think about this … lighting up as many small -> large SMPs as you need, in a unified way, quickly and easily.
Yeah, UCS could be an interesting HPC play. Add vSMP to the stack, add some fast storage somewhere into the stack …
I think ScaleMP’s value just doubled. Pure speculation of course.
Make as many small VMs as you want. Make as many large ones as you want as well.
Maybe I am wrong about ScaleMP suitors. I wonder (aloud) if the right match would be VMware. Scale down, scale up.
VMware fundamentally changed the economics of servers, allowing for massive consolidation onto smaller numbers of systems. ScaleMP has (IMO) altered the economics of the large scale servers, by creating the ability to “stitch together” servers to build a large single system image.
Interesting ….

5 thoughts on “HPC @ Cisco?”

  1. I’m not so sure about running vSMP over 10GigE (feel free to chime on Shai!). The problem is that the latency on 10GigE is just so damn high compared to IB. It’s about a factor of 10 larger. One of the pieces of true innovation in vSMP is the guaranteed latency of the algorithms. You might be able to do that with 10GigE but I’m not sure (it would be kind of like guaranteed arrival on US Airways – it’s always late and that’s your guarantee).
    Back to HPC and Cisco California. Everything I’ve seen on the web is focused on virtualization. It’s where the money is (HPC is still niche compared to enterprise but it is a large niche). So I don’t think California is _targeted_ at HPC but if someone comes along and says, “hey can I run an MPI application on the servers,” Cisco will do back flips to make sure it happens.
    There are some suitable gotchas though – the biggest one is that it only works with Cisco’s Nexus switches. So you have to buy one of the big 10GigE switches to run the California servers. For some parts of HPC that are still GigE only, this might be a big deal, but for those that need more, it’s not a big deal (if you can accept 10GigE).
    So my summary is that California will work for a very small set of HPC customers who love 10GigE and insist on TCP only (“non of that stinkin’ IB for us!”) and are a dedicated Cisco shop and who have a Nexus switch or are willing to buy one. But I think the main target is the virtualization space.
    One last thought. The Register and the Inquirer, while providing some fun rumors, didn’t kill Cisco over this one. The crucified Dell when it was rumored that Dell might be developing a Smart Phone (“What does Dell want to be when it grows up?”) but when Cisco, the main networking company on the planet wants to get into server, these rags didn’t bat an eye. But then again they didn’t use profane language as they have been apt to do lately 🙂 I guess when no one takes you seriously you have to do _something_ to get people to read your site.

  2. @Jeff
    True, latency on 10 GbE isn’t quite there yet (the ConnectX EN latencies are pretty good though, relative to “normal” 10GbE).
    I was (honestly) speculating … reaching effectively. Infiniband gives you many nice things (including the latency aspect). One of the things it gives you is the ability to route signals over it at a very low level, before you load a TCP/IP stack. Which you don’t quite get with 10GbE. Of course you could go the Chelsio RNIC/iWarp/RDMA route (often disparaged in HPC circles, though done so in a competitive vein, not out of anything inherently wrong with the approach).
    I was simply saying that I think UCS “could” be an HPC play, and speculated that ScaleMP could be an interesting direction for this play to go.
    [begin dripping pre-caffiene sarcasm, not directed at Jeff[
    As for the register/inquirer picking favorites and bashing others … well … yes …. didn’t you know that the role of the (mainstream) media is not to objectively and intelligently report, but to tell us rumors and innuendo masquerading as fact, and tell us right from wrong, good from bad, and winners from losers? You must not have gotten the memo on that.
    [end of sarcasm, you will now be placed back in the universe of this blog post … have a nice day]
    Dell has to branch out. Not necessarily finding a niche, but something more.
    In computing Dell has carved out a nice region for it. There are (quite a few) things missing, but it has lots of stuff covered on the business side. The consumer side is a problem (as usual). Consumers are somewhat more … fickle … more prone to google for a particular vendor and good/bad reviews for the level of research they do.

  3. Jeff,
    you are completely correct. IB latency is dramatically lower than 10GbE, and I’ll be happy to be proven wrong. Although vSMP Foundation is interconnect agnostic (and in-fact we demonstrated vSMP Foundation on Ethernet some time back), IB has several advantages over it.
    To sum it up, if there will be a significant demand for vSMP Foundation using 10GbE (instead of IB) – then we will get it working, but I estimate that it will always be slower than IB.

  4. @Jeff
    non of that stinkin??? IB for us!
    …that???s not a valid ranting point for users of ScaleMP???s vSMP Foundation. vSMP Foundation hides the InfiniBand infrastructure, which results in dramatic simplification. There is no need to install any software/firmware or use InfiniBand directly; it just magically works (appearing as the SMP???s machine’s memory bus/backplane). So I expect non-IB fans to be happy with vSMP Foundation even today!
    The icing on the cake is that it even manages dual-rail-IB transparently. Yep, without any configuration or setup. Just buy that second IB switch, and make sure you got two HCA ports on each system :-). High Availability and increase in bandwidth, for the price of the hardware alone.

  5. Pardon my acknowledged ignorance, but since this thing is a virtualization play, and “unified” computing, are we sure it doesn’t already come with a layer that stitches together the blades into one monstrous hypervisor platform? That’s what I’m trying to figure out about the thing.

Comments are closed.