18 Replies Latest reply on Mar 4, 2013 1:50 PM by spoitras

    Which Compute Architecture Will Reign Supreme?


      I would imagine that most folks who have gone either all rack servers or blade servers for the data center will lean a bit towards their chosen solution, but I’m curious about other options that are beginning to permeate the stack. More blended ideas, also called “hyper converged infrastructure”, such as example offerings from Nutanix, Scale Computing, and Pivot3, are entering the market and gaining traction.


      The idea with this compute shift is to combine all three food groups into one physical server box: the storage layer (often a mix of PCIe flash, SSD, and SAS / SATA), network layer (iSCSI or NFS front end, with perhaps Infiniband on the back end), and compute (Intel or AMD chips in a “mini blade” format).


      The end result is the removal of the traditional stack model. Compute now has direct access, or zero hops, to the storage. There is also a significant reduction in cables snaking back through the data center cabinets. It also removes some barriers to entry; traditional skill sets around concepts such as SAN design are much less relied upon.


      Regardless of product, do you feel the hyper converged infrastructure model is right for your data center servers? Or is the traditional monolithic model of racks or blades connecting to a SAN and LAN the way to go?

      What do you think some drawbacks and benefits of both models are? I’d be especially curious to hear from folks who have tried the hyper converged model and have some war stories (positive and negative) to share.


      Reply to this post and earn 1 entry to win an iPod Nano!

        • Re: Which Compute Architecture Will Reign Supreme?

          I like the idea as it becomes a cleaner and quicker data center. On the other hand I don't like having to rely so heavily on the same hardware, single point of failure.

          • Re: Which Compute Architecture Will Reign Supreme?

            We are going the hyper converged route.  It is doing wonders for a cleaner install, noticeable improvements in data performance, but I do agree about the single point of failure concern.

            • Re: Which Compute Architecture Will Reign Supreme?

              We still utilize the SANs in our datacenter.  While this does increase latency to some degree it also comes with the additional benefit of easy replication between datacenters along with data deduplication that allows for less storage to be utilized.

              • Re: Which Compute Architecture Will Reign Supreme?

                  We use the SAN model here too. None of the vendors we've been working with have seriously pitched hyper-converged to us and a little latency isn't an issue.

                • Re: Which Compute Architecture Will Reign Supreme?

                  We're currently utilizing rack servers simply because a 1u rack server is more cost-effective than a blade with the same specs, but I would prefer the blades from a concept standpoint.


                  I agree with freid.42 though, on both accounts, there is a single point of failure, but it can also be offset in cost. You're just pushing the bottle-neck though with this concept right now, and I'm not saying that it won't work itself out in the future, but it's kind of like saying do you really need a 50mb wireless connection when your gateway is a 10mb or less connection?


                  It depends on what you're going to do with it. If the slowest piece of your puzzle is from the SAN to the LAN, or from the LAN to the APP, and they can all be part of the same appliance without a degredation in performance, then a converged appliance is probably the way to go, but only if it's cost-effective.



                  • Re: Which Compute Architecture Will Reign Supreme?

                    As a cloud services provider we use the rack server model with large back end storage systems.  This model allows us to be very flexible with both capacity as well as different hardware types.


                    I personally don't have any experience with the hyper-converged model though I find it very interesting and would love to have an opportunity to compare it against what we already have.  I did take a look at a few of the vendors that you mention and what stood out to me is that they also have different boxes to add both more compute and more storage so in the end, isn't it effectively the same thing as the rack server or blade model, just a different hardware vendor?

                    • Re: Which Compute Architecture Will Reign Supreme?
                      Carlo Costanzo

                      As a consultant, I don't really have much of a set opinion on the types of hardware the client's choose.  The converged models work and fit the bill but so do the stand alone components.  For the smaller environments, it usually comes down to price points.  Depending on the season and market share of the provider, converged can sometimes come in cheaper (think Nutanix).  The nicest part of all of this is the VIRTUALIZATION though.  Moving systems from hardware vendor to hardware vendor has never been easier.  I think it is one of the primary reasons that these new guys even get a shot at running in the datacenters.  Purchasers know that if it doesn't work out, you can just easily migrate your systems over to a more stable/better preforming hardware platform.  That wasn't the case when you where running physicals and had significant investments in application and software setups and configurations that would be redone if porting over to new hardware.


                      I think Virtualization has definitely given rise to a new set of risk takers.  


                      • Re: Which Compute Architecture Will Reign Supreme?

                        I've been working in semi-converged (UCS) environments for a year or so. The network is virtualized to some extent, and if your systems are standardized enough, your servers are just actors playing a role. So maybe halfway there, more when the storage appliances based on UCS platforms start kicking in (Nexenta comes to mind).


                        But I would feel some fear of vendor lock-in if I go with just one provider's all-in-one solution. There are some nice people at a lot of these companies, and I obviously wish them absolutely no harm, but I'm aware of at least one Redwood City company that's been known to buy great companies and dilapidate them beyond recognition.


                        I'm not trying to play the oracle of Silicon Valley, but what if one of these big companies buys and buries my solution provider, or stops giving me updates? Or what if your company buys the competitor and bans the product you chose (more likely in big companies, I'd guess, but you never know). And how practical is it to move the environment (or even parts of it) between solutions? With a rack environment (or blade environment) with distinct network and storage, if (for absurd example) HP stops selling servers, there's a good chance that my software platform, and possible a fair bit of the hardware, will work on, say, Dell or Supermicro hardware.


                        If you're virtualizing on top of this, then you have a step toward portability, as Carlo mentioned this morning. And if your platform is standard (hadoop, for example), you can move it out of a custom walled garden to a standard one (or a different walled garden).


                        So I guess, it depends. Do you trust the converger and the converged platform, and/or do you feel you could handle it if the platform had to go away for some reason?

                          • Re: Which Compute Architecture Will Reign Supreme?

                            Interesting point. Sounds kind of like a "eggs in one basket" type of deal. Although, to play the other angle, even if your hyper converged player gets bought would there not be a parent company selling the rebranded version? Or you could slip in another product when you're due for a refresh (probably 3-5 years anyway). The only major difficulty would be growing the infrastructure, but if you're designing for a pod like structure it wouldn't be too painful.


                            The killer would be if hyper converged completely died and you had no way of getting back into that model (highly doubt this).

                              • Re: Which Compute Architecture Will Reign Supreme?

                                Good point, and I'm kinda devil's advocate in the first place as you might guess.


                                If the new parent company feels it's in their interest to continue supporting the product, rather than just swiping the underlying technology (or staff or patents or warehouse of cool swag), you may actually be in better shape. Better supply vendor relationships, benefits of scale, larger support and development/enhancement resources... all good things.


                                I've usually seen rolling refreshes... to meet capacity or replace failed hardware. Admittedly I've never been at a company for five years, and many companies I've worked for hadn't been around that long yet anyway. And there's always something you just can't move... So I'm pessimistic about refreshes.


                                If you're modular enough, and retain the resources to support the environment regardless of what product it lives on, this sort of solution can be a godsend. And I like the ideas and implementations for Pivot3 and Nutanix and others (I know a couple of the cool Scale guys but don't know anything useful about their products). I just lean toward providing CaaS (Caveats As A Service).

                            • Re: Which Compute Architecture Will Reign Supreme?

                              I'm pro anything that makes the data center less cluthered with cables. I'm really getting fed up with all the spaghetti at the back of the too narrow racks.


                              More on topic, however, a colleague of mine was talking about the future being putting a Cisco UCS server inside a Cisco router NM-slot with all storage (SSD, SAS) and CPU's  fitting inside a "little" router. I think this sure sounds plausible if storage keeps getting smaller and CPU's getting more and more cores.

                              • Re: Which Compute Architecture Will Reign Supreme?

                                *WARNING* - I work for a "hyper-converged" company...  but I've also had/deployed UCS/Vblock/FlexPod and HP CSM (heck UCSM is an awesome product)


                                I think it's simple: one size doesn't fit all, but it can get pretty close.


                                Coming from the tech consulting industry and working primarily with f500's "convergence" has been a great thing.  In the past years "cloud" was all the buzz, but when it comes down to reality - it was about enabling IT to be agile and flexible.  Semi-converged platforms like UCS/Vblock/FlexPod offered the bundling of components into a single combined appliance, making it easy to standardize and scale.


                                I've worked hands-on with Vblock/FlexPod/HP CSM and there are definitely benefits to these platforms and the scalable convergence they provide.  However, unless you're a f500 or service provider the unit of scale might not necessarily be something applicable.


                                Convergence runs in multiple areas: the network (10GbE/FCoE), the hardware (storage/compute), and the roles.  For example, with UIM a single admin can provision "Vblocks" which handles the full stack: network/compute/storage without necessarily needing to know WWNNs/WWPNs, NX-OS and the like.  Think about DevOps and CloudOps, though they are more the concept of an ideal environment vs. true reality, do we see these being anywhere near possible without converged platforms?


                                If we look at the trends for application architectures, they are moving towards the distributed model (eg. Hadoop, Splunk, etc.) for the large scale cases.  Why?  It can increase performance, provide against SPOF, and make things more flexible and efficient.  It takes the logic out of the hardware and distributes it into the software, SDDC anyone?


                                We likely won't be seeing SAP running on a hyper-converged platform anytime soon, but that's an example of an application built for legacy "mainframe-style" infrastructure.


                                If we look at what UIM does for Vblock, it abstracts the hardware from the services, making it easy to quickly provision Vblocks (IaaS).  Sure you can do this with traditional infra + BladeLogic or an orchestration tool, but why make things difficult when they can be easy?  That's a big driver for hyper-converged platforms, you shouldn't need to have unnecessary complexity to deliver simple IaaS services, it should just work.


                                I'm not a fan of "buzz words" and SDDC is definitely a buzz word...  However, the concepts behind it are simple: hardware gets dumber, software gets smarter -> actually enabling true scalability, agility, and things like CloudOps.   Is there any way this could be possible without hyper-converged platforms?