Hyper Converged Considerations - Sydney 2016 VMUG Keynote
Transcript
all right so I think this is a fun topic because we can get into what we call a little bit of a religious conversation which is when to select hyper conversed infrastructure it's almost like saying when to select virtual virtualization for a solution is there really a bad time to select a virtualized environment or a hyper conversed environment you know what the answer is it depends a lot of the time not all the time but a lot of the time the argument is not necessarily technical
I think I'm probably a good selection to talk about a discussion that dips in and out of the technology a little bit and we'll get into why I think uh this is a topic that's near to what I do date today so unlike Scott low I haven't written seven books so you don't know who I am so we'll spend a couple of minutes talking about who I am uh why I'm up here talking to you about hyper conversed infrastructure if uh you might have known me
in a previous Twitter life or previous blogger life I'm uh traditionally virtualized geek so a lot of the traditional Community content was around how to you know you run into a problem or you want to do nested NSX in a lab that's the type of stuff that I would have done on the virtualized GE Channel over the past few years I've moved out of a pure technical role into more of a business focused role that began a couple of well maybe about 5 years ago now
I became the Chief Architect for one of the programs L Martin had providing to the US government in that role I was kind of abstracted away from the day-to-day technology while I was responsible for doing traditional solution architect stuff much more customer focus a lot of project management a lot of uh business case development uh trade study analysis getting away from the day-to-day keyboarding that progressed I went to to PWC if you attended the noon session with uh me and Scott L you will have heard
that at PWC the argument and for those of you who don't know accur on PWC is price water house Coopers they changed their branding uh in the US and globally to PWC a few years ago I walked away a little bit even further from the data center in that day-to-day job and the the typical conversations were no longer should I go with NSF NSF I scuzzy versus fiber channel those arguments were no longer the arguments we had with clients at PWC the conversation was should I
Outsource all of my it to a managed server service provider and let them worry about the NFS ice gy and fiber channel conversations so gave me a unique view from the business side of how decisions are made so for this conversation we're going to dip in and out of technology and business to understand what are some of the drivers behind making a decision on a hyper conversed infrastructure and then at the bottom is some housekeeping where you can find more of my content virtualize keep.com for
I still blog over there whenever I get the urge to scratch the the the geek part of my DNA I I can't get away from it I still enjoy just doing random stuff like trying to run aor stack and and vcloud air why would anybody want to do do that type of pain to themselves on that guy Tech republic.com which is a larger uh for-profit Endeavor run by big us Broadcasting Company the CTO advisor which is more around the topic that I'm talking today in Tech
Target I think we all know Tech Target fairly well they're a pretty good resource within the communicate and within the community so I don't know if this is the case here it see I I think I get a sense that it is one of the interesting things that I've learned in conversations with you guys over the past day or so is that virtualization is a given most environments are 80 to 100% virtualized if you can't virtualize it why would you want it and that changes the
argument for hyper converged infrastructure a little bit but I think from a business perspective the thing that I think is common throughout Western culture is that application teams are having a bigger voice in the infrastructure kind of outsides than what we've traditionally expected in this x86 based world it's been for the most part I think uh the new tant folks did a great job in the presentation talking about monster VMS Mike and Josh where they talked about the application owner saying you know what I need
a 96 gig VM with 20 cores and the application probably only needs 16 gig of RAM with two cores and we've fought that battle in the past and to this point I think we won the we' we've proved virtualization out that virtualization in a shared infrastructure most applications can live in that well but what I'm starting to see is application providers such as Oracle sap going straight to the application owner and especially in Oracle case is selling a full stack so storage compute networking as a
single appli what they call an appliance and I see parallels from this Appliance approach in the hyperverse market so we'll talk about my experience dealing with application owners from the sap side of the house specifically sap Hana the again uh Josh Mike did a really great job giving an example of a sap Hana workload sap Hana workload at the minimum I think the slide they had a uh 20 20 to 40 cores and 128 gig I think uh my experience has been even in in that
might be a good test environment but production 256 gig you're talking about a scale out environment 256 gig and the number of cores debatable uh 20 to 40 is probably a pretty good number so we'll talk about the parallels and the business side of making the decision go with a Hana Appliance or a by o or bring your own data center approach by the way I mentioned this in the afternoon session I can talk I not only am I loud I enjoy talking uh if you
don't interrupt me and ask questions I'll just go right through so please feel free raise your hand if you have any questions I really like this slide this is provided by the folks at wik Bond they spent a lot of time looking at the x86 what they initially called the server sand market the server sand uh taxonomy never caught on seran is basically x86 storage scale out storage what we've seen in different solutions we I have a what we call NASCAR Slide with different options but
the interesting point that I wanted to show on this slide is that somewhere Circa 2020 2021 server sand or x86 based storage will start to outsell traditional storage arrays so there's a lot as think these are these revenues are in the millions so you know we're talking about 4 40 billion a $40 billion market and server s taking half of that software taking half of the market from the traditional storage guys so slide for an effect so let's start off the conversation with level setting on
what's hyperon converged if you blog and you've ever written about hyperon converged chances are the vendor has a vendor has emailed you in response some vendor you've never heard of and said you know what our solution is hyper converged too and you're like wait you're uh isn't that like open source server software server sand software from 1999 you guys how you hyper converged all of a sudden so let's level set on the definition of of hyper converge a virtual storage array so you know your typical
three node minimum storage uh x86 cluster with hypervisor software running on those units and management software to manage the deployment and administration of the the hardware cluster so the Storage level the hypervisor potentially and then obviously adding to adding on to that so every time I want to add a node i' add a node I go into some type of administrative console click next next next whatever and I have I've expanded out my hyper conversed infrastructure any debate or objection to that definition you guys are
great this is going to be a great dictatorship I I I I like Australian culture all right hyg versus converged so what's hyper about converged it's hard for me to to get hyper I'm from Chicago it's 11 o' I got in yesterday and it feels like 11:00 my bedtime is about 10:30 so I'm going to do my best whenever I say hyper to you know let you feel it so a convert system has similarities the difference in a convered system again your definition may vary but
this is a great dictatorship and I'm enjoy being the dictator has compute Network and storage that's engineered tightly together with a management overlay so normally you'd have some type of centralized managed solution so these are completely independent systems we'll pick on one of uh the V communities or Federation companies examples for example vce the V block is a great example of a convered system you have csto for your network in compute then you you have VMware for your hypervisor level and layer somewhere in a factory
someone has engineered they spec out their solution and they've put it together and they've engineered it in a way that it should run extremely well and we'll get into to the advantages of that from a support model so examples of that from an industry perspective are or rack sap Hana we call them in the sap world we call them appliances they're not really hyperconverged a lot of I'm sorry converged a lot of vendors will sell their converged solution as a sap Hana product an example is
HP their converge system 7,000 uh the uh which is basically or their I forget the the names I think this the convert system 9,000 is a Superdome X some three-part storage and a uh some HP switches they engineer together they go to sap say test this configuration sap tests the configuration and says you are blessed to run sap Hana on this platform and then the third option which is bring your own what we've been doing in the data center for years obvious question is can I
build my own hyper Converse solution if you went to Chris Wall's power shell demonstration you'll see and even in the nxs session in this room a little bit before that the scripting tools that we have available are incredible for mean if you're a Windows administrator as I am typically I cut my server Administration teeth in the windows roll on on Windows NT 4.0 and the concept of what you can do I can I can write a heck of a batch file I mean it you would
be amazed at the stuff that I can do with the batch file but nothing near the level automation I can achieve with power shell power CLI that we use in the Community quite a bit I would guess that if you worked hard enough you could create a product that look a lot like a hyperon converted solution but for the most part it's really tough and it's probably not worth the effort to to try and emulate the stuff that the uh vendors have done from a management
level so from a boo bring your own hyper conver solution for the most part the big difference is is the lack of the management capability so we have the virtual storage we can go out and get a scale IO which is completely free build a server sand pretty capable server sand on top of vmw vpar hyperv and I don't know if they I don't know if skill iio supports KVM and then still run our hypervisor workloads on top of those same x86 storage platforms and it
looks a lot like hyper converged we can pick on vsan is vsan hyper convered VMware has started to to Market vsan as a hyperon convered solution for the purposes of this definition is we're managing tightly the hardware and hypervisor level so vsan didn't make the cut in my definition sorry before it g any you guys are really not working hard to win these uh these Apple TVs this is an exciting topic I'm not doing it justice and the Apple TVs so I'll go back and look
at the video and see what I did wrong but we we'll get all right so who are the hyper Converse players where can I get it I don't have a I don't have a slide or a projector big enough to show all the players who can't who claim to be hyper converters but these are off the top of my head some of the more obvious names in the community couple of the vendors are sponsoring today's vmug user con uh It's Curious you know I put HP
e they beat me up if I say HP Enterprise HP actually has a hyper convers solution I'm told scale Computing we'll get in Dale and we'll get into the details a little bit of these different vendors and then VMware has their Evo rail their evil rack reference platform you know I think most of us in the room are engineers and it's all about requirements if application owners and business people could always just give us the right requirements our Solutions will be dead on every time it's
all about not getting the right requirements and choosing hyper converged over one of the other two options is again about requirements I put performance scale and support in the primary require requirements cost we're Engineers we don't care about cost we we care about delivering the right solution unfortunately we live in a world that requires cost so our will cover cost at the end so let's talk about performance one of the things that I loved when I get off the airplane was that I saw I'm not
a car guy so to speak but this holding brand from GM I was really excited to see the different variations of the cars I love in the US here one you guys drive on the wrong side of the world which is confusing I'm a runner so I have to really think about not getting hit when I go run in the morning but the other thing is that the variations on different cars uh are amazing I got off the off the and there's a UT what we
call a utility bill vehicle if you guys call it youtil you and we call I would have called it a El Camino SS and it was very similar I thought man that's really cool and really throwback then one of your V mug uh leaders Ryan gave me a a ride back in from the airport in his car and it's really fast and the sedan that's just really exciting so from a performance perspective hyper Converse infrastructures can rock I mean these things can perform they are built
on the same technologies that we've loved for years in VMware I'll hearken back to the new tanic presentation and I have to stop talk calling it new tanic because the Josh and team did a great job Josh and Mike did a great job of not selling new tantics they sold the concept of virtualization that if you want you can scale up your storage array to provide one million iOS to a single VM take it takes two flash arrays but you can do it so you can
build anything you want from a performance per performance perspective and Hyper converged a lot of the advantages from a technology perspective for hyper conversion performance is the disc is on is tied with the computer one of the advantages of sap Hana is that you've combined the database with the analytics and CPU so there is no latency going from disc to CPU and Hana goes One Step forth it's in memory so you even get great uh greater performance from a latency perspective because you don't have the
latency of the flash bu you you're processing from memory so the closer you can get you know it's basic computer science the closer you can get comp the storage to the CPU the better performing the application will be hyper conversed is no difference so when hyper Converse was first introduced to the industry a lot of the initial use cases was vdi because vdi is read heavy uh when you know the I remember a lot of the first white papers were were was around combating boot storms
read boot storms when when you're booting up you know 800 VMS first thing in the morning these things haven't changed on the flip side one size doesn't fit fit all if you've ever tried to solve a problem just by adding flash to an example just by adding flash to A system that doesn't necessarily solve the problem adding a higher speed uh IO doesn't necessarily uh reduce right latency there are what I call nerve knobs and my good friend over at the packet pushers calls it a
little something a little he wrote a blog post that called it something a little bit different than nerd knobs I won't repeat it on stage because I'm reporting myself but basically when you need to tune for performance most and this is the generalization most hyper pervert Solutions don't give you the knobs to to make those adjustments and obviously one of the big things is lack of support for inmemory database it's not from a tech and this I'll uh I might hit on it yet again it's
not uh you do disqualify yourself if you're phone go off for the Apple TVs just so you know the problem isn't a technology problem it is a perception problem you know we you guys have done this region have done a great job of eliminating the perception problems when it comes to support of VMS for your mission critical apps unfortunately in memory databases specifically sap Hana from a certain workload perspective has not so the while we can virtualize those workloads to get support on that from uh
specifically sap you the the hardware ACL is very strict and they don't support hyper Converse I don't think there's a hyper Converse vendor yet that has gained sap Hana certification let's talk about scale snapshots just to give you guys an idea of how big some these Solutions can scale new tanic announced one of their largest deployments 1,500 nodes I don't know the details around it I don't know if that's 1,500 no I don't know the architecture I don't know if that's 1,500 Noles in a single
cluster I don't think it's really important just ideal that a uh nanx customer has 1,500 nodes in a single environment Simplicity they recently announced a deal in October I don't know if it's deployed that's you know or for months plus of 200 armist stack systems and then on the SMB side scale largest customer or largest configurable cluster is about four nodes or roughly 200 250 VMS one of the considerations when it comes the hyper conversed architecture is scale out versus scale up most hyper conversed Solutions
very modular we can take a brick what I call a hyper conver brick add it to your existing infrastructure continue to scale that infrastructure out so if you're in an environment that can leverage scaling out from a performance perspective that performance is usually pretty linear I can keep adding bricks and keep increasing my uh ability to increase performance So you you're in a very parallel environment scaling out and Hyper conversed very good solution versus scale up again this is the generalization once you buy it there's
very few systems that allow you to go back and add individual components and this is a hard computer science problem you know if we think about it from a uh one of my favorite industry terms to think about or talk about is eraser coding I hate the term but basically this ideal that I add par to protect data when I think about it from a brick level if I add more bricks eraser coating is pretty simple I just spread the parody across the the I rebalance
the my my cluster and the parody bricks are ridden across or the parody bits are written across the new noes that I add to the cluster but what happens if I take a single node and I add disc to a single node how does that impact eras coding and parody skew on yourcity it's it's it's a hard computer science problem or if I change memory in a single node in a in a cluster these things are highly engineered to work together well when I change one
part of that system that wasn't meant to be changed or customizable post Factory config add complexity to the environment this is one of the nerd knobs we talked about some of the advantages to the software only solution I had a great conversation with uh rollinson Rivera from VMware he works in the office of the CTO he was on the podcast and we talked and I asked him this question you know why the static configurations with vsan when it first came out and who walk me through
kind of these hard computer science problems but the advantage of software is that you can fix these problems and this is not unique to VW vent new TS simpli scale all these vendors will as the products mature they get more data from customers they'll also be able to solve some of these challenges but generally speaking scale up is a challenge in hyper conversed infrastructures any questions on scale this is usually the Hot Topic so just isn't this just basically the nature of Intel you might as
well go back to the main frame so got this big lump of compute and the system has to deal with it this is so I'm a sap guy now I'm a sap guy I'm a sap guy because that that's who that's I got a great question of why did I choose sap as a career sap infrastructure because the sap team has a lot of money just simp it's a simple thing follow the money it's and uh actually posed that question to the team which was sa
if I was not going the Hana ra runs on AI a why not just buy a z13 this a nice system and it's super reliable and you can't get much more converged than that so yes good point so this is going to be a competition this is unlik this is not a $35 chomecast he's leading and he brought up Mainframe so the bar is low all right support and I think support is one of the big big are is when we talk about hyper convers versus
the other two options or mainly B the ability to pick up the phone and say vendor [Music] my I would like to say network but hyper Converse doesn't really include Network yet but my storage is performing poorly or my hypervisor isn't doing what it needs to do or I'm getting uh Hardware errors from the disc you have one we like to call it thr the choke or one single point of contact with conversed systems and Hyper conversed systems the burden of support is a huge huge
relief this is why sap Hana today doesn't run on hyper conversed platforms today if I bought a uh sap Hana Appliance and again I use air quotes when I say Appliance because it's it looks very much like a convered system excuse me if I bought a sap Hond solution I had it a problem and I think Hana and Oracle are great examples of kind of the ultimate level support when it comes to to this model let's say that the application team made a CH I'm sorry
back okay if if one throat to choke is a big benefit of hyper converged systems does that change when you buy um mics on Dell or Simplicity on Lenovo because you're system from One V then you are actually still buying one system most of those mods are let the the those specific vendors talk to themselves but when I talk to Dell about ask Dell the same question Dell's the throw the trouble okay's not the expert inic software the software that is the secret source and the
hardware that's commodity so this is where Ching the right thr so that's a valid question and I think that the sap Hana example is a great example so if I let's say the application team by the way we have two people leading the way for the Apple TVs one asks sap Hana question and I want to ask the main frame we're getting better but come on guys if I bought sap Hana Appliance from let's say Dale it's a series of say it's a Cisco or Dell
switch uh dell storage and Dell compute with sap Hana running on top of redhead uh real or Susi it doesn't matter the application team makes a change over the weekend changes a cury and a cury that was taking 13 seconds is now taking 13 minutes they pick up the phone they call sap not sure if the problem is the problem with Fury if there or if there's a problem with the underlying storage or if we're having a configuration problem and and rail you know patches we
patched the system the same weekend as well we just have four Change Control we made two major changes in in the same weekend the this support model sap owns the entire stack they have Dale sitting in the next room and they will have the sap guys go and look at the the application guys look at the security changes that were made and then concurrently they have the Dale guys looking at the hardware I'm not privy to how Dale and nanic set up their relationship or inity
in Cisco but my understanding it's a similar relationship that the support teams are tightly integrated so very good question yeah that's exactly to how BCE does they don't the there's not many Cisco guys Cisco anymore Cisco employees working at BCE is there's a very tight support Arrangement support yeah our sponsors their sponsors here for for Dale uh new tanic Simplicity they can most definitely answer that that level of question but that was a very good question so let's get to one of my leg slides let's
talk about cost so I told you guys I spent two years as a management consulting at the end of the day you know what I learned there's no such thing as saving money it's it's a Fool's earring the you you move it from one bucket to another bucket you're either going to spend it on Opex or you're going to spend it on capex vendors will disagree with me but we've done this analysis time and time again you can make these numbers say whatever you want them
to say but generally speaking you're not going to save money everything is relative what I have here is a relative chart so I I built kind of this model that said 050 100 so your scale is your small organization this is the caveat if you're small organization and you go out and buy a converge system you're buying you're probably buying the wrong system there's a minimum there's a minimum footprint for a conversed system so if you go out and buy a conversed system versus a hyper
convert system in most cases or a BYO a bring your own you spent more money than you probably had to to get the solution now requirements can drive that to be different if if you're in an organization that says Erp cannot go down and you and you spend $500,000 to get into and I'm making this number up you spend $500,000 to get into a convert system versus a $100,000 to get into a hyperon convert system and there's virtually no difference in uptime between the two systems
from a technical perspective it's been my experience Executives feel more comfortable when they spent the 500 it's you know when you put the weighted calculator in front of somebody oh this calculator feels it's it's it's heavy it must it must be good same concept if you were to rightsize the solution and bought a BYO system versus a hyper convert system you probably spend roughly the same amount of you might spend less in uh C but you'll end up spending a little bit more in Opex and
the difference becomes negligible there's a reason why I showed the application slide at the beginning sap Hana deployments for a large organization can be anywhere from 20 to $70 million infrastructure cost in that will be anywhere from 5 to $20 million that's not you're not going to squeeze savings out of the infrastructure for that type of project so when the application team is driving the project they're going to look at infrastructure and say you know what the cost isn't this is why I didn't even have
cost on the title slide cost isn't the factor for infrastructure I'm trying to figure out how I can get everything off of Oracle and on to Hana so I'm not paying two database charges that's where the real cost savings is at so when we talk about infrastructure costs especially if you're looking at some of these higher-end hyper convert systems it's being my experience that there's not an awful lot of cost savings any questions on doesn't this come down to the nature of workflows no one is
writing software that's designed to really scale out Amazon is allowing you to spin up a lot of things UC for the way people have always done workload but no one is saying reinvent the way we do work so you're solidifying your Apple TV here so I I like that conversation one one of the one of the advantages of hyper conversed over something like a scale out a true scale out or web scale or Cloud native application is that we can buy a hybrid conver infrastructure and
not change the way that we deliver applications today it's invisible to the application Level later so in an earlier and I'll ask the question here how many people in here have containers in production that's telling the answer is none and that's telling in a part of the world that really embrac virtualization hyper Converse is just another way to do what we've always done the biggest difference is in I think the big biggest difference is in support and simp and [Music] simplicity I really like the fact
that imity named name VC I have enough problems talking I can say those two things in the same conversation I think the one of the biggest values of hyper conversed and conversed infrastructure is the support model is the ability to call a single organization to get support uh the scaleout piece of the architecture you know what the simply adding infrastructure is nice I I I don't buy into that as much because I don't have environments that that that need to scale from an infrastructure perspective that
often and that uh so taking 15 minutes to install uh to expand the infrastructure versus the weight time working with my MSP to uh have it done in one month I have bigger problems than that that's not that's not solving a problem for my environment so for me that's that's not a big deal but the great thing about hyper ConEd it it doesn't change what how we do or what we do today in the data center as far as application delivery all right so last slide
in question so when did you select hyp conver you guys are in a dayto day when do you do it money why not now why not why not so what what what why wouldn't you let's reverse the question why wouldn't you select hyper converters if the cost is if I made the argument that that the cost is cost neutral that you're either going to spend the money in Opex or you're going to spend the money you know the these hyper coverts Hardware vendors are pretty smart
they know how much you spend Opex and they know what the tolerance is of what they can charge uh before they start to eat up all of your Opex savings so when do you when is it not a great [Music] time or great use case to buy Hyper conver depends on the depends on the workload depends on the licensing model that you have if you've got something that's locked in big process driven then uh it's probably not a good time it depends on the way again
in the pocket so you know a great example of this you guys remember when blade systems first hit the scene great ideal and concept but to make it economically feasible yeah have almost populate the whole thing out the gate so there was this minimum buyin to get into blade systems if you didn't you know if let's say something modern like a c c7000 I think holds I mean I came ium guy I know a c7000 holds eight itanium blades I don't know how many x86 blades
it holds I think it's 16 I like spend a lot of money on Legacy Harware the uh there was a point that if you didn't at least have eight blades in there that it just wasn't cost effective I think hyper Converse is pretty much at that same point when it comes to uh performance or or or value per per dollar spent which is one of the points you haven't touched on and from a small Enterprise point of view it's not about necessarily how cost neutral is
it's about exporing the resources so if you've got a team of 10 people looking after your Enterprise you've got to have and you've got a oldfashioned sand with BL wires everywhere and brade switches and stuff you got to have two dudes who know how to drive that thing all of a sudden the tyer converged you can repurpose those two BLS and have them do something else you have but again it depends on your word on your circumstances whether or not you need those so the question
has been you know if you go with Hyper Converse and you're SMB and you've dedicated resources to let's say sand Administration versus server you know you have these traditional silos and you go to hyper convers you're able to get rid of the silos good point you be your fun L able to change your it organization my argument and this would bear out and I haven't done it for SMB but it would be an interesting exercise to do if I went with let's say a scale solution
which is at a certain price point and gets rid of my and I have the right workloads gets rid of my sand gets rid of my uh Network complexity and then simplif and potentially gets rid of uh to make this argument to you too gets rid of my V licenses have I saved enough to repurpose those guys I think that's a that's something to consider I think that's worth uh having an extended conversation over once we're we're done because I'm I think I'm pretty much out
of time I think I have time for one or two questions any any observations on more exotic typ of converged architectures like um D core I mean it's it's it's a nice story that you kind of you know use all the CES and treat them as memory um but that's kind of ex and I the whole H conver play was and I I don't think the hyd converge play is as much about money as it is and a lot of the vendors would tell you it's
about future capability and going Beyond just the infrastructure and as you start to provide apis and assuming that you know this goes into a whole another talk and conversation but when you look as at the data center as a apid driven thing and the underlying and HP calls it composable infrastructure and I you know I take something like the machine and have a data center that in a computer and a data center is just one big computer you know the the kubernetes approach and all that
that becomes an interesting thing so I think it that goes beyond hyper converg and I think even from an infrastructure perspective we're not quite ready for what that looks like from an operational perspective there that that's really cutting edge and on the Leading Edge of what we're I think accustomed to managing but I like the I like the the few the you know I'm a architect so I like looking five six years out and I think that's a a great Lookout on what happens when the
whole data center is a single cluster or a single resource any question so I get two winners for the yeah it's called the main frame I just want to bring that as you can the the that's yes it is called and we went back again to the main frame uh guys it's been a wonderful vug thank you very much for hosting me [Applause]
