Datrium Architectural Overview - Open Converged
Transcript
[Music] hi welcome to the CTO of Weiser day needles you know what a little bit different format I'm on-site at de triomphe interviewing de triomphe CTO Hugo Patterson which we've had on the podcast before you go welcome back to the CTO adviser audience and also Devon Hamilton director of engineering I'm sorry systems engineering Ford atrium we're here with the V Brown Bag crew doing a build a super excited to just talk about the build a quickly Devon you've been helping the crew set up for build
day ten quick buy what is build day and what are you guys doing exactly sure so really the intent of our building exercise is to take a set of day dream assets which would be our compute nodes and our data nodes and actually put those forth into an install process it would be just like when we deployed these at a customer site so basically getting the whole hands-on experience soup-to-nuts from how you set things up initially out of the box working through all the steps that
are there kind of seeing what's simple what takes some attention and the network and things like that and bringing it up online so that the customer can just kind of see through this process of the build a what it would be like if they had it on-site alright and you can actually watch the build day event online actually YouTube is one negative you can follow the brown bag or on YouTube or you can go to the V Brown Bag comm website and there'll be a link
there but I wanted to talk to you too specifically about this concept of open converged citizens now that you go we've talked about this in the past just the date room platform in general but I'm starting to see the concept of open converged and other vendor literature so I wanted to at least spend some time talking through what is open converged how is it different from hyper converters so let's start there so thanks Keith and I'm very glad to be here the the open converged is
leverages has a lot of commonality with hyper-converged in that it scales the performance as you add compute resources as you add hosts the performance increases and that's the same as hyper-converged but the difference is that in open converged we don't turn the servers into persistent storage we have a separate set of data nodes where all the persistent data resides and though they scale independently of the compute so you can scale performance by adding compute or you can scale capacity by adding data nodes so Devin help
me out here how is that different from my basic three-tier architecture from a customer perspective you talk a lot of customers and try ok I can go out and buy my storage rate I have my compute I have my management layer I can scale those independently now why am I going to date you to get what I basically already have that's a great question Keith right off the bat the first thing to understand is that we're server powered storage so you want to kind of take
on the mindset of really software-defined and yet we've added to that the components of a full turnkey solution that's enterprise grade that's all supported by one vendor and so while you have a lot of choice and capability you still have a cohesive model to take advantage of if you look at traditional architecture with a SAN network environment and compute environment and move that forward through HCI where you've kind of unified those and you have caching and compute up there in the host side what you find
is that in both of those camps there is rigidity and cost associated with how those infrastructures are laid out in our model you have a completely open architecture that in fact focuses on what we call split provisioning you can basically put out as many data nodes as you need for capacity as many hosts compute nodes as you need to do that hi oh and that that compute side those are dissimilar resources potentially you're saying that you can use basically any x86 architecture those can be servers
that they are buying from day trium in the form of what we call compute nodes or intermixing those with servers that they already have where we're making a betterment to the infrastructure that they already own in fact I'll even take that a step further and basically say that because of how we are deploying our technology as software on those compute nodes we also can come into a customer environment non-disruptive li and run right alongside legacy concerns that they already have that are advertising that make sense
makes sense in a traditional array that comes with a controller that has a certain I ops capability and every host that you add to that environment gets to ever you know it slices that I ops capability ever more thinly in a server powered architecture the hosts are responsible for the performance every host you add adds more performance capability to the overall system so you're not like subdividing a fixed pie you're growing the pie in in a open convergence architecture I have to push back a little
bit on this what when I hear open converged the first thing that comes to my mind is OpenStack and I go to a talk to a ton of vendors and the concept of oh okay I can take what I have install of control plane aka OpenStack on there and then I can you know go to my Linux distribution of choice and do some type of classic file system across some standard x86 holes and basically get the same thing in my mind where does that break in
practice and where does the atrium comments have filled a gap and say in the truly end value so daydream does support Linux and KVM environments is where as well as docker bare-metal persistent volumes but we also work very well with VMware and virtualized environments and that's very different from OpenStack which is kind of more of a open source kind of a path but the the core of the data center really most commonly is running VMware and so integrating and supporting that environment is in a very
important part of of what des trium is all about so let's talk about support and growing is this a how is this sold is just a because you definitely said I can put it on existing yeah piece my saw some amazing benchmarks and I don't know if that was one day true branded heart we're like house did said is true soul so those benchmarks were developed in partnership with Dell so they were a huge help in providing a hundred and twenty eight servers because we scale
up to supporting a hundred and two friends I do and also ten of our new flash based data nodes so that's a pretty big environment and yes eighteen million I ops you know supporting eight thousand I'll mark VMs it was amazing but that we worked with Dell to help deliver that so it's really pretty standard x86 servers so a hundred and twenty notes that gives me in a mindset when I think of data center still a lot a lot of the marquee in this industry has
been around at least in the hyper-converged space web scale data center scale all kind of I can replace my data center with this architectural where some of that falls down practically speaking is in this unevenness of distributing two cute resources and storage resources you know I can't that's not how my data center work sure how how does open converged and date really help with that problem can I build an entire data center architecture around this will be converging concept yes and and you know Devin touched
on this a little bit before but because our hosts are isolated from each other they're not a lot of crosstalk each host is responsible for its own performance and it interacts really just with the pool of data nodes and what that means is those hosts don't have to be all the same the larger data center environments they tend to be heterogenous there's a variety of different server types maybe supporting different kinds of workloads and data centers need the flexibility to support all of that those make
workloads in the environment and we can support third party so we offer compute notes but we can also support your own compute nodes that you bring your own x86 servers and you can mix and match and they don't have to be all the same you can have quad socket servers with lots of flash for a data warehouse and you can put it next to you know a few servers running VDI and that can all be part of the same dvx and the secret to this and
the secret to supporting the you know really the range of servers is that isolation of one server from another the elimination of the crosstalk from one server to another and just the conversation that happens between the servers and the data pool which is storing those persistent snapshots so Keith can I bring a field perspective to that sure so it's really important that our customers kind of recognize right off the bat that because we're not distributing i/o across the host nodes that the host can be dissimilar
and that's kind of the first blush comprehension of des trium what we add to that is really a new modality referred to a split provisioning and to get back to the root of your question a minute ago we can have a diverse amount of data nodes and a diverse amount of host compute nodes that there is no tie between them there is nothing that relegate that I have to have one for one or or two to one or anything like that that split provisioning concept means
that if they have a capacity heavy workload they can add more data nodes and grow up that cluster scale if they have a more i/o intensive work load they can branch out and have more of those compute nodes and drive to tremendous i/o numbers so really you can start from the customers perspective with a single a one and one a data node and a host compute and then scale about all the way out to ten of these data nodes and 128 of these compute nodes flash
can scale literally up to 16 terabytes per each host compute node factoring deduplication you're talking pushing almost 7 petabytes of active data in flash in a fully blown out cluster so we're going to get into this conversation on data management toppy data and a couple of other videos that we're going to shoot part of this dvx building I'm really intrigued on how customers can potentially save money we appreciate nature and for sponsoring the CTO advisor daily dose on-site for the day trium build day on the
brown bag we'll talk to you next CTO daily doles you go Devin thanks a lot my pleasure Thank You Keith
