Grey Beards on Systems - 78: GreyBeards YE2018 IT industry wrap-up podcast

Episode Date: January 9, 2019

In this, our yearend industry wrap up episode, we discuss trends and technology impacting the IT industry in 2018 and what we can see ahead for 2019 and first up is NVMeoF NVMeoF has matured In the prior years, NVMeoF was coming from startups, but last year it’s major vendors like IBM FlashSystem, Dell EMC … Continue reading "78: GreyBeards YE2018 IT industry wrap-up podcast"

Hey everybody, Ray Lucchese here with Howard Marks here. Welcome to the next episode of the GrayBirds on Storage podcast, a show where we get GrayBirds storage bloggers to talk to system vendors, discuss upcoming products, technologies, and trends affecting the data center today. This GrayBirds on Storage podcast was recorded January 2nd, 2019. It's just Howard and I for our annual year-end podcast. How are you doing, Howard?

I'm doing pretty well, Ray. It's been snowing pretty good here in Santa Fe. It's pretty dry where I'm at at the moment, so thank goodness. So what's new in the storage and systems industry? Well, I think the biggest story is still

NVMe over fabrics. God, I'd have to agree. In 2017, 2018, we had E8 and Accelero and Aperion going, here's this really good idea for a new protocol. It'll have lower latency than SCSI

and we can move it over the network too. Yeah, that was last year or that was the year before. That was the last year and the year before. That was the last year and the year before. But then this year, we've got NetApp and Dell EMC not just saying, oh, yeah, we support this as an idea, but integrating it into their flagship product. And for Dell EMC, turning VMAX into PowerMax and going a whole hog into NVMe over Fabrics. Right, right, right. Well, it's not that. We also did a three-separate podcast on NVMe over Fabric this year. So it's also got some technology advancement going on as well.

And from the technology advancement front, at the end of the year, the TCP version was approved. So that's now part of it. Oh, God. So Solar Flare is locking and loading now. Yeah, Solar Flare and even more LightBits Labs, who are the guys who wrote this back. I got you. And it looks to me like in the short term, NVMe over TCP is going to be big in the hyperscaler end of things.

Yeah, because of the cost of the switching for the other one, I guess. Well, and just the scale that they work at. Right. Building lossless Ethernet over 10 switch hops is hard. 10? 10 switch hops? Well, if you've got a data center that's a million servers how many switch hops is it from

one end to the other it's a lot it's 10 yeah doing that in tc doing that in tcp makes a lot more sense for the short term in the corporate data center it's going to be nvme over fiber channel because i have fiber channel i know fiber channel i don't have to argue with the network guys when i deal with fiber channel and i'm using it as a lower latency transport to talk to a traditional array because that's the next step. And the enterprise guys are very incremental. Yeah, yeah. Well, it is a different protocol. It's obviously a different transport layer.

So, yeah, there's some subtleties there. And, you know, it's not a g gimme but obviously the dell emc guys and the netapp guys took that on and made it happen and ibm as well yeah well and i used i used dell emc and netapp more at the i would expect those guys to trail somewhat right more you know more i'm surprised they did it i mean pure Pure Storage made their announcements months before Dell EMC and NetApp. Right. But I was less surprised to see Pure try and front run. Yeah. And then Flash Systems has it as well.

So, yeah, the IBM guys. But they're also, you know, sort of high end stuff, too. So it's interesting well today envy me over fabrics is viewed as a high-end thing yeah i mean who's left hatachi and ventara and hp i guess are the big majors that haven't made the the switch yet right yeah i haven't seen announcements from either one of those but you know but i fully but but it's come down to i fully expect it yeah very soon i mean i talked to the guys at itachi and obviously they're working on it and stuff like that we haven't talked to the hp guys in a while but that's a different story. Yeah. But next year is when people are going to start to discover that NVMe over Fabrics isn't just about my highest end workloads.

Right. I love NVMe over – I love NVMe SSDs. Oh, God, that's a whole different discussion. But part of why I love NVMe SSDs is there's no HBA or RAID controller. Yeah, yeah, yeah. Well, you know, there's obviously RAID stuff going on. It's just not at the controller level. Right, but the driver talks directly to the SSD controller.

Yes. There isn't that intermediate layer. Hmm, hmm. And that intermediate layer, whether it's a Fiber channel HBA or a SAS HBA or a RAID controller, adds queue limitations. Queue depths are, you know, today if I have 64 VMware servers attached to my SAN. Well, wait, wait, wait, wait. So you're saying NVMe over Fabric Fiber Channel doesn't require an HBA?

It does. Because you still have to talk Fiber Channel protocol. But the queue management is NVMe queue management, where instead of having one queue per target LUN, there's one queue per CPU core. So it's much, it's even over fiber channel, it's much more parallel. Based on compute cores. Yeah, yeah. So the queue depth management problems of both local storage and SANS become much less significant because it's software, not that HBA device.

Anymore managing this stuff. Yeah. And if we're doing NVMe over Ethernet, it's one less driver to worry about. Right. Although, you know, although then we'll get into the, it's the ethernet offload. Yeah. Well, it's a different ways of doing it too, right? So the, the Mellanox ethernet offload might work better or worse than the Chelsea ethernet offload.

Right. Right. And, you know, from my conversations with folks inside both Microsoft and VMware, many of the offload features built into today's Nix that less than 20% of customers use aren't that good. Which is why 20%, less than 20% of the customers are using it. I'm sure there's a chicken and egg problem. It's, you know, do you use the iSCSI offload? Well, no, because I never figured out how.

Well, that's good because if you did figure out how, it doesn't work so well either. Yeah, wouldn't make that beneficial. Yeah. Hmm. Interesting. Well, yeah, the other side of this is, you know, we've seen a lot more usage of NVMe SSDs, even in, I'll say it's software defined storage as well as the big, big players now. I mean, just about everyone has NVMe SSDs in their configuration at the enterprise

level, I would say. The one, the one I, the one I didn't, I still don't see NVMe configurations from is Nutanix. Which, you know, and of course they yell about other people's architectures not being ready for NVMe, but that's a whole other story. That is a different discussion. If your VSA runs under Linux and there's a standard NVMe driver in Linux for NVMe SSDs, all you have to do is test. Right. And so why not? Yeah.

And the NVMe SSDs aren't necessarily more expensive anymore. Right. They just happen to be faster. And the latest servers have enough U.2 slots. And then it becomes, why do I want to run SAS and put an HBA in the server to manage them? Anymore, or RAID controller. Yeah, yeah, yeah. I have had so many bad experiences with so many HBAs and RAID controllers

over the years. Howard, you've got a lot of scars here coming out. Well, you know, I bought a set of servers when vSAN was in beta and did some betaing of vSAN was in beta and did some beta-ing of vSAN on those servers using the ACHI controller on the motherboard

to run my SSDs and HDDs. vSAN came out and they said, we're not going to support ACHI. It doesn't have a deep enough queue depth. Ouch. And just over and over again, problems have turned out to be in this intermediate layer that we can get rid of.

Yeah. That serves to be a bottleneck that complicates the cabling inside the server. Because running new cables inside a one-use server that set up for H that's set up for ACHI when you're, when you put the SAS mother, the SAS HBA in. Right. Right. And it's like, Oh yeah. Fishing those cables. That was fun. Yeah.

So, you know, and so at the laptop end i think all laptops have switched to m.2 because the space savings is so big right and once you went to m.2 you could run m.2 sata but why bother yeah it doesn't cost anymore right right right and the difference is significant yeah so today you know the only downside to nvme ssds is in servers if you want a large number that the plx pcie doesn't support switch chips yeah yeah are still expensive yeah yeah yeah and so the when you go i want to go from the chassis that supports 24 sas drives to the chassis that supports 25 nvme drives you know on the Dell site, it goes, yeah, plus $800. Yeah, yeah, yeah. Mine are a little weak.

But that's about what the HBA would have cost, so. Right. So you're pretty much, you know, it's a wash, I guess. Oh, it's even better, though. Now you're going NVMe, though, so. Right. Yeah. Right.

For the same amount of money, I would much rather have NVMe. Right, right. And considering the fact that the SSDs don't cost that much different anymore, they're better served. You know, it's $100 on a $2,000 SSD. Yeah, yeah, yeah. Well, and the other thing, we were talking to somebody, it was this Brian, right, last podcast about the fact that NVMe SSDs

are just configured're just they're just configured faster anymore they've got all the the right electronics and buffering and all that other mechanisms in place to just make it a faster ssd and the sas and sata ssds if anything are kind of becoming almost archive like or tiered not quite archive obviously but i mean they are we're moving into the point where the standard architecture for the next few years is going to be some 10 15 of high performance high endurance flash and some 90 of of read intensive lower performance flash. And Archive is a little bit big of a step,

but it's definitely the capacity, not the performance. Yeah, yeah. So it's almost like the high-capacity slow performance versus low-capacity high-performance disks that we've known and loved for years here. Well, it's interesting that we're going from three or four years ago where we had a performance SAS or SATA SSD and a capacity spinning disk and a performance difference between the two of hundreds of times.

Yeah, or magnitude. Yeah, yeah, magnitude yeah yeah yeah yeah and we're going back to fast flash and slow flash which is going to have a performance difference of two to four times kind of like the performance difference between a 15k rpm drive and a 7200 yeah it's interesting isn't it that the the ranges are shrinking back to being really close to what we used to do. But the software for managing data placement has gotten much smarter than when we had two kinds of hard drives. You didn't mention that. You said it was the high performance were also high endurance. Are you seeing that?

I just don't know. Well, to get high write performance from Flash. Yeah, got to have sophistication. Well, you need to have more over-provisioning so that your garbage collection can ensure that you always have clean pages and so and so the same things you do to an ssd to make it fast are the things you do to make it higher endurance okay yeah and you're just suffering the the capacity costs to some extent. Yeah. Right. And so they track together. You know, you use more of the flash in SLC mode to absorb the rights. And then when you do garbage collection, you rotate it down into.

MLC, TLC kind of stuff. And in a high-capacity, low-endurance drive, you run very little in that mode, and you garbage collect more often, and you have large space of less over-provisioned flash. And it doesn't matter that writing to the TLC takes longer, because it's only part of the garbage. Yeah, it's masked from an online perspective. Yeah, it's post-act latency doesn't matter. Yeah, yeah, yeah. Yeah, so I talk about high performance, high endurance because they go together. Right, right, right.

And when we talk about lower performance, the difference is only maybe 50% slower in terms of read latency, but four or five times slower in terms of write latency. Yeah, no, I agree. I agree. So you look and you end up with a model like vSAN. Yeah, multi-tier SSD. If you're in all flash mode, writes go to the performance tier, but they don't promote on read because the back end is flash already and fast enough. Yeah, I agree. Speaking of vSAN, what do you think of HCI these days? Well, I saw some data, and I forget from where, that HCI has surpassed converged infrastructure. As a market revenue generator.

Yeah, yeah, yeah. Well, I mean, it's been on a tear since VMware vSAN started coming out here. Yeah, I think it was like vSAN 6.2. Yeah, that was really the driver. Where you start going, okay, that's good enough for my production stuff now. And it's taking off. Just about everybody nowadays has got vSAN ready nodes. And gosh, obviously Dell EMC has made a lot of hay with VxRail,

but there are other players out in that space as well coming out. Yeah, the HCI market hasn't completely settled yet. Yeah, yeah. Nutanix and VMware Dell EMC are clearly the volume leaders. Right. But HPE could make a run with SimpliVity. I don't think the door's closed on that yet. And Microsoft could decide to get serious.

Yeah. Hardware-wise, yeah. Well, I mean, you look at Microsoft and even in Windows 2016, they have storage spaces direct, which is the distributed storage layer. So there's an HCI system in there. If they wanted to. That they're just haven't spent any time and money promoting. It appears they're going to be much more serious when Windows 2019 comes out any day now. Well, you know, they've, well,

they've had this Hyper-V stuff for quite a while and that, you know, it's just, it's sort of sitting there on the side. I think that, you know that from an Azure stack perspective, that they would start pushing that side of the game. Yeah, and they have a significant cost advantage because you buy Windows Data Center Edition if you run Windows VMs because that's how you license the Windows VMs.

And Hyper-V, the hypervisor, and Storage Spaces Direct, the storage layer, are just built in. They don't cost anything extra. Right. So if you run Hyper-V with Storage Spaces Direct to build a Windows hyper-converged environment, you didn't pay for Nutanix, you didn't pay for ESXi, you didn't pay for vSAN, you didn't pay for SimpliVity. All that stuff just came with. Now, you do have to buy System Center because if you have more than one or two nodes, managing it without System Center isn't worth it.

Yeah. But there's still $10,000 per node of software. You don't have to buy it to go the Windows way. Right. So it'll be interesting to see if Microsoft tries and if they get any traction with it. And Azure Stack might be an interesting way for them to get into the data centers and get people to see, well, look.

There's some options here. We have this whole, we have this, you know, you've been running our backend for a year on the stuff that you're doing that looks like Azure, and it's time to do a server refresh on the other stuff. Why don't you think about changing hypervisors? You know Hyper-V works. You've been using it.

Yeah. I think what they've missed from a perspective is is is uh you know i mean obviously the dell emc vmware tie-up has been significant but vmware has kind of offered uh this plethora of vsan ready nodes almost from the get-go well i mean first first there was evo rail which did die in a miserable death well everybody offered ev offered EvoRail choices, and EvoRail just was badly packaged. Yeah, yeah, yeah. But VxRail, which is the Dell prepackaged version, and then ReadyNodes, which are basically reference designs. It's order this server bill of materials.

The combination has been certified to run with vsan right um which for all practical purposes means it's a meat in the channel and your var will bring appliances right right and i i have a blog post that i'm just about to put up. Actually, there's a blog post I put up a couple weeks ago where I talked about HCI as an architecture and the precepts of HCI design. And the one I didn't include that I kind of am making, it's not really required, but for all practical purposes it is, is that it has to be delivered as an appliance. Because the promise of HCI is simplicity. Yes, just like CI actually, right? Simplicity. And while you could buy a server from Lenovo and SSDs from Micron and an ethernet card from Mellanox and an HBA from LSI. That sounds like a paper I wrote a couple of weeks ago. And software from Maxta and make it all work.

And it would be architecturally HCI. Right. It wouldn't be simple. There's nothing simple at all about that. Right. It wouldn't be simple. There's nothing simple at all about that. Right. And so I think that the delivery to the end user has to be an appliance. Right. I don't care.

If it's done by the channel or a vendor, but it's got to be an appliance. Yeah. And so Nutanix has changed their go-to-market, and they now call themselves a software company. Right. And so you can order Nutanix on Lenovo. And there's a, like ReadyNotes, there's three models and these options, and you can only order these options. Yeah. And they've all been verified.

Mostly the options are SSD and DIM size. So kind of doesn't matter. And they arrive on my loading dock with the software on it. And I don't care if the VAR did it or if the distributor Arrow or Avnet did it or if Nutanix did it in a factory that has a Nutanix sign on it. Right.

I want to order something simple to order, you know, a small number of options, and I want it to arrive complete. You and I are of an age when you could go to the Chevy dealer and custom order a car that would be made at the factory and delivered for you. Yeah, yeah, yeah. And the options list was like 600 options. Yeah, yeah, yeah. Nobody wants that.

Anymore. No. Well, nobody wanted it. Well, a couple of people wanted it then. But it wasn't worth it. Yeah. Today, you go to the car dealer and you go, I want that model car. And they go, do you want package A, package B, and package C?

Right. And here are the colors that are on our floor that you can take home today. Right. And HCI should be that simple. Well, and it has. Obviously, for VMware, it has. To some extent, for Nutanix, it has. And whether Microsoft can make it there is really one of the questions here.

Yeah, I think that HP and Microsoft, and to a lesser extent Cisco, have the opportunity to challenge VMware and Nutanix. With a Microsoft solution, yeah. Or with their own. Right. Those are the three guys who are in AAA and could go up to the majors. Yeah. Okay. I got you. Much as I love solutions like scale computing for the use cases they're really good at, I wouldn't want to have a data center with a thousand nodes of scale computing because it's not what they do well. Yeah, I guess. I guess you're right. They're a different scale. Scale computing is on a different scale.

Yes. Ditto for so many other players out there, too. I mean, obviously, there are other software-defined solutions out there. There are niche players who know their niches. Right. There are players who thought they could do everything

and haven't gotten enough traction yet that the amount of traction the leaders have means it's just not going to work. Yeah. You know,

but that's the way the market goes. Right. Right. Right. Well, I think it's, I think to some extent, with the possible exception of Dell EMC

and VMware, this HCI market taking off like it is has been a surprise. I mean, a lot of players, a lot of these big players invest a lot of money in CI, right? Yes. And Dell EMC is a big one

in that space. Well, I mean, they created that whole VCE infrastructure. Everybody else had a reference design or sold their stacks. They took it to a whole other level. Yeah. And then everybody else followed. I mean, just about every vendor down the block, just one after another, came up with their own solution in that space. Yes. And most of them had Cisco UCS servers because Cisco didn't have a storage division

to compete with. If you're nimble, promoting HPE or Dell servers wasn't a good idea. At the time. Yeah, I agree. Oh no, if you're nimble now, you promote HPE servers. But I mean, it's the same discussion for, you know, yeah, I agree. I agree. Oh, no, if you're nimble now, you promote HPE servers. But, I mean, it's the same discussion for, you know, yeah, I agree. I agree. Every one of the vendors came out with a CI solution. But I think Dell EMC has been somewhat prescient with respect to the VXRail thing. I mean, they were first out of the gate with that.

They've done real, real well with it and continue to do well. Only now are we starting to see other players in that space. I mean, I would argue that's where Nutanix always was. Yeah, I agree. I agree. I mean, they had appliance, but it was a little bit different game because, I mean, they came at it from the software side and built an appliance surrounding that with the hardware sort of hidden, whereas Dell EMC and VMware sort of split it between the two. Yeah, I agree.

I agree. Yeah, you're right about Nutanix. They've always had an appliance model. Yeah. And they're now changed to we don't sell appliances, we just sell the software because it makes our margins look so much better on Wall Street. Right. The low margin server is on somebody else's books.

Right. But that's all finances and channel. It's got nothing to do with the user experience. But the interesting thing, so, you know, we talked about NVMe over Fabrics, and we talked about HCI. And a lot of the HCI solutions have NVMe SSDs. Already built in, yeah.

As that performance layer. The interesting thing is when we start looking at the software that drives HCI, that SDS layer, and NVMe over fabrics, so that vSAN, instead of strictly consuming local resources for a server, can consume resources from an NVMe JBoff too. Don't you think that might be sacrilegious on their behalf? They had a demo at VMworld. Okay.

So they already made the statement. That's interesting. And I've had some chats with Mr. Christos. Yes, of course. Because one of the problems with HCI is it's a shared nothing model. And in small clusters, that means it's storage inefficient. Yeah. clusters that means it's storage inefficient yeah yeah i i read a paper by somebody at vmware a while ago and i of course lost it about building application centric clusters in with vsan

so you'd have a vsan cluster of of three or four or five or six nodes that runs all your sql servers and then the dba can be involved in only you know but if you've got a cluster of four or five servers and they're your sql servers so you're running n plus two data protection because it's the most important thing you do you have to triple. You have to leave enough room for it to rebuild when a node goes offline. That's really inefficient in terms of storage. Do you think the storage inefficiency is because of replication or RAID solution that they use? Well, I mean, it starts with in any HCI cluster, your usable capacity is the capacity of N-1 nodes,

so that you always have to have space to rebuild. So if your cluster is four nodes, then you gave up 25% just to have rebuild space. If your cluster's 20 nodes, you give up 5%, you don't care. The next problem is that the data protection mechanisms have to protect against the loss of a whole node. So if you have three nodes

and you want to lose a node and still be protected, the only way to do that is to replicate three ways. And in fact, you need four nodes to replicate three ways and be able to rebuild. So it's 33% times 75%. So you're roughly 50% kind of numbers. No, no, no. Worse than that, 25%. So if you've got four nodes, each of which has four terabytes of capacity. You've got four terabytes of capacity.

You've got four terabytes of capacity. Yeah, although terabytes of capacity. You got four terabytes of capacity. Yeah. Although you've got 16 here, right? It takes 16 to give you four. Yeah. Now, you know, so by the time you get up to, and I can do four plus two data protection across 10 nodes. So that's about 50% efficient.

Actually, 59% efficient. So by going from a four-node cluster to an eight-node cluster, you can more than double your storage efficiency. Right, right, right. But if the media is shared so the ssds are in a jbof and all four of the sds processes can access it well there's no reason they can't write to four plus two ss. Yeah. And when a node goes offline, you don't have

to start copying data to rebuild it onto another node. You just have the other node software take responsibility for the LUNs or VVOLs or whatever that it was managing. NVMe space, namespace kind of thing.

Published NVMe namespaces most likely. Yeah, yeah. So, I mean, here's the question though. You know, a regular small or smaller clusters are probably not going to be able to afford NVMe over fabric JBOFs. Or are they? I mean, so is there an expense to this? There's the SOHO robo use cases where in that data center there is one small cluster. Right.

And those guys, it doesn't matter. Yeah. But if you're building application clusters you say here's my four node vdi cluster here's my four node sql server cluster here's my six node exchange cluster and they can all use ssds from the same 24 ssd jbop yeah yeah now it's working pretty well. Then that composability means that we don't have the inefficiency, but we do have the noisy neighbor protection and other advantages of having an application level of functionality. And you could actually, if you wanted to, share that JBoff across multiple application clusters, which makes it even just like shared storage, only without the… It's shared storage without the… well, with the intelligence still in the compute hosts, not in separate controllers. Right, right, right. Right, because if we look at storage, there's two pieces. There's media and intelligence.

Yeah. And we can aggregate or disaggregate those pieces as it serves the purpose of the problem we're trying to solve. Yeah, yeah. problem we're trying to solve yeah yeah so aggregating media can give us efficiency continuing to disaggregate the intelligence means that now that intelligence is managed by the vm guys so i don't need dedicated storage guys nearly as much yeah um and that i might just be buying 28 core instead of 24 core processors just to provide the support but but that difference on each node might be less than the cost of the controllers and software that turn oh god those jbofs. Yeah, yeah, yeah. And so there's lots of interesting, you know, by the time we get to the second generation of this

where we can say the SDS layer runs in a container. Oh, yeah, yeah, yeah. And you can run it with a controller or you could run it on the server node. It's your choice. Right. You can have dedicated servers to run those containers or you can have those containers distributed across your Kubernetes clusters with any other containers. Yeah. Yeah. When the right latency of a virtual array made up of some number of these controllers, controller containers. Right. When the right latency exceeds some trigger and the CPU utilization exceeds that trigger, fire up more containers because that means we're recording a lot of data and the dedupe is taking a lot of CPU

and having more containers will distribute it and reduce the latency. And you can get that to be completely dynamic and automatic. This is 2121, 2122. Excuse me. No, no, this is 2019. Excuse me. No, no. This is 2019. We can do this next week if you want. Well, as commercial products, it's 2020.

Okay. Maybe. Yeah, I agree. I agree. I agree. It's not next year. It's the year after. They're certainly containerizing a lot of this functionality

already. You can look at, you know, God, the EMC guys have been talking about container, you know, container functionality for their storage system for quite a while. So it's not like they haven't made that step. And the Caminario guys have some of this functionality. It's just, you know, nobody's quite made it all the way there yet it would be nice to have the next version of the nvme over fabrics management interface spec so that well so that the command to a jbof to create a child namespace was standard and not every vendor came up with their own way to do that.

You know, so there's some maturity pieces, but it looks very interesting to me. Speaking of containers, Howard, we talked to one guy this year, NGD Systems, about how they were putting containers on their flash drives, running containers on their flash drives, running containers on their flash drives, computational storage game starting to come out of the woodwork here. Yeah.

Well, for a lot of applications, storage to CPU bandwidth is the bottleneck. NVMe was an attempt to solve that. It moved the storage closer to the CPU. There's another way to do it. Well, what NGD Systems and a couple of other players whose names escape me are doing. ScaleSoft, I think, was one of them. But yeah.

Is saying, well, instead of moving the storage closer to the CPU, why don't we move the CPU closer to the storage? Or the application, in this case. Well, the CPU cycles need it. Yeah, I agree. Agree. cycles yes yeah i agree agree and one of the dirty little secrets of the ssd business for the past few years is that flash controllers have arm processors uh-huh yeah yeah and ngd took one of the development chips that are on the market that have an fGA or an ASIC, and I forget which one NGD used, so I'll leave it at that. and an ARM processor on the one chip and you program the ASIC or the FPGA

to do the things that you want to do in a hardware because it's faster and you do it all the time. Didn't VMware just announce ARM support for ESX? Yes, they did. We will come back to that. Let us not go down that rabbit hole quite yet. All right. Okay.

I'm sorry. Sorry for introducing that. And so what NGD did is they said, okay, so here's an ARM processor. It runs Linux. And you can download your application into the ARM processor and run your application in the SSD. Yeah, yeah, yeah. There's some API stuff that surrounds all this. But yeah, once it's done, you got to get it there and you got to use their service to.

Yeah, I mean, there's development kits. You have to write your code for this environment. But basically, you write an application for ARM Linux. Right. Basically, you write an application for ARM Linux. You test it on a Raspberry Pi or whatever other ARM Linux box you have. Then you go to the SDK and you go, how do I get it in? How do I do the couple of things that are different? But this makes a huge amount of sense for things like search.

Well, lots of database stuff, joins and stuff. I've got two petabytes of image data. And I want to find the pictures of Ray dressed like Santa Claus in that image. Good luck with that. Go ahead. Okay, or a Howard in an Aloha shirt. That was easy. There's more, right? I got you.

Okay. Well, clearly there's pictures of me in Aloha shirts in this database. Absolutely. But if I just have enough code to recognize what an Aloha shirt is and right and to have each ssd search all of the data on that ssd and just return the pictures of people in aloha right that would be a major step then the next layer up has a lot less data to search through and the fact that the SSD is further from the big CPU doesn't matter. Yeah. Yeah. So,

so even if you're just doing that first pass of this pattern recognition stuff, being able to distribute across hundreds or thousands of arm cores. Talking hundreds of thousands of NGD SSDs out there. Yes. Yes. Well, one, one SSD is four or six available cores.

Yeah, which some have to be done for IO. But yeah, okay, let's say, you know, 30% or something like that. Yeah, but if we call it four, then a cabinet of 24 SSDs is roughly 100. Yeah, you're right, you're right. So thousands is easy. That's a rack. Yeah, yeah, yeah.

And we started going, yes, and there's thousands of cores doing this search, and it's just one rack. Yeah. I can see a lot of applications where that makes sense. And when I left the casino industry, which, of course, we're supposed to call the gaming industry. Anymore, yeah. Facial recognition was just being pitched to them and casinos were very interested in facial recognition um because of two things the obvious one is cheats come back to the casino right and i'd rather know that this

guy tried to cheat before and we've thrown him out at the door than at the table. Yeah. But there are also people who, let's see, this is a family podcast, right? This is a family podcast. Okay. There are other undesirables. Yes. Yes. Who tend drunkenly to come back to the same casino over and over again. And they need to be filtered as well. And they need to be filtered as well. Yeah. Yeah. And, you know, we got a demo of this stuff and it was going to be millions and millions of dollars. That was in the old days. Nowadays, you can do it forever.

Oh, no. It was in the old days. And this stuff today cloud ai all that stuff well i mean you wouldn't want to go cloud because it's continuous thing um but storing the images of all your undesirables in ngd ssds and sending down every image of somebody who walked in the door going is this guy an undesirable would be a good idea yeah yeah huh that's interesting stuff yeah we we were doing this quite frankly for database stuff back in the late 90s on an enterprise class

storage system. And, you know, it was all based, you know, because at the time you had to have special interfaces and stuff like that. It was even worse. SDKs and APIs didn't really exist in those days, or at best were rudimentary. So it never took off because, quite frankly, there was one database that would use it and the rest of the guys wouldn't. So unless you were using that database, you didn't see the speed up yeah it was it was too complicated and too expensive with today's with containers and linux and arm processors

and you can run just about anything you want in the container and now you can just about run it on an ssd out there at the at nvme or fabric you're talking serious stuff here. Well, I mean, by the time you get to be a gray beard, you see ideas come around for the third or fourth time. Yeah, more or less, yeah. And the third or fourth time, the market's ready for the idea, and it gets traction. Or the technology is better.

Right. It gets traction. Yeah, no, I agree. Yeah, we tried I agree. Yeah. It's, you know, we tried that before and it was too slow and too expensive. And then everybody goes, that was a bad idea. And I was like, well, it wasn't a bad idea. It was just premature.

So we had talked about a couple of things that's up and coming that you might see in the new year. I think one of the things was the Optane DIMMs was, we're starting to see some of that come out, but. I went to the Intel Optane DC DIMM announcement. Right. And I'm very bullish on this technology. Optane SSDs have been a disappointment because the price performance ratio just didn't work out yeah for the same money you could buy a 400 gigabyte optane ssd or 1.6 terabyte nvme

flash flash ssd and you know five times the performance at five times the cost per gigabyte wasn't a big enough thing. Yeah, it wasn't an advantage. But if we move that optane to the dim slot, instead of being five times faster, it should be 20 or 50 times faster. Right. And 20 or 50 times faster at five times more expensive, that's economics that makes sense, especially as seasoning.

Seasoning? Well, as opposed to the steak. Okay. I got you. I got you. So, I mean, we talked about the performance tier, high performance, high endurance. Yes. Well, obtain dim as that landing zone when your rights come in. It's even better.

Right to memory, replicate to the other controller, act when he answers. Right. Yeah. Yeah. That makes a huge, you know, 1% of your capacity in something that fast is a really good idea. Right. Just to absorb the right activity. Yeah.

And then there's, you know, all the stuff we talked to Brian Bulkowski from Aerospike about, about how having that just as addressable memory changes application development. All right. Well, Howard, you know, I think we need to leave it at there. Sounds good to me, Ray. Is there anything else you'd like to say to our listeners? Well, I mean, we talked about HCI. One of the Network Field Day delegates, Ivan Peponyak, runs online courses, and I'm doing six hours in three two-hour pieces for him on HCI that

first one was December 11th the next one is January 21st but you subscribe to his class you get access to all this stuff online and a slack channel for asking questions if you get them from the recording um it it's slide webinar. I didn't have the camera on. So actually my bathrobe, but that's different. We don't want to talk about that. Well, this has been great. Thank you all for listening in on our podcast. Next time, we'll talk to another system storage technology persons. Any question you want us to ask, please let us know. If you enjoy our podcast, tell your friends about it and please review us on iTunes and Google Play as this will help get the word out. That's it for now. Bye, Howard. Bye, Ray. And Happy New Year, everyone.