Grey Beards on Systems - 89: Keith & Ray show at Pure//Accelerate 2019

Episode Date: September 18, 2019

There were plenty of announcements at Pure//Accelerate in Austin this past week and we were given a preview of them at a StorageFieldDay Exclusive (SFDx), the day before the announcement. First up is Pure’s DirectMemory. They have added Optane SSDs to FlashArray//X to be used as a read cache for customer data. As you may … Continue reading "89: Keith & Ray show at Pure//Accelerate 2019"

Okay, I'm here at Pure Accelerate with Keith Townsend, CTO Advisor. How you doing Keith? Pretty good. How about you Ray? I'm doing fine. Thanks for being on the show today again. Pure has been announcing quite a few product lines and stuff like that. I think the biggest one of interest was the direct memory, storage class memory on their solution. So what do you think of what they're trying to do with that Keith? Well you know, we heard them in the Tech Field Day presentation,

they alluded to workloads like SAP, in-memory databases in general. And then on stage, they went right after it. You know, 65% of the cost, 90% of the performance. Which, like on paper, that sounds good. But in practice, that 10% is a big performance hit for the cost value. So what they compared it to was the cost of DRAM. Yeah, I got you.

So having 100% of the database in DRAM. Right, versus having a subset of the data in storage class memory. But if you took that same storage class memory and put it in the server that we've seen at the Intel events, et cetera, you get very similar cost savings and you get the 10% more performance. Interesting.

So, you know, it's- Well, it's not, I don't know if it's an exact, cause I mean, with the storage class memory, it's sort of a, it's a read cache. So they don't actually have, they don't have to have the whole database in there. They didn't give a lot of specifics on this. And Intel has done a lot with SAP to optimize the PCI path and the

memory caching specifically for SAP. And I don't know if that additional latency going to the storage array to access the same memory that just sounds like a big performance hit I would like to have challenged them maybe they can do a sponsored Greybeard or even a lab set or something like that or have somebody that could actually explain what they're trying to do and stuff like that well it's interesting. I mean, so they've got, what, a 3-terabyte and a 6-terabyte version of that solution that you can deploy.

I mean, what's a typical SAP HANA? It can't be more than 64 terabytes or something. Yeah, so the big HANAs would be, you know, you're looking at 12 terabytes. 12 terabytes, okay. 12 terabytes would be a pretty big one. Four terabytes is a typical deployment. Two to four is typical deployment. When you start to get big, you're talking about

12 to potentially 20 terabytes. Okay, okay. So, at the big end, it couldn't be 100% in storage class memory. It'd have to be some portion of it being flash and stuff like that, and there would be some sort of, you know, caching characteristics to that he did mention 90 of the performance what does that mean so if if if a workload takes 90 minutes it takes uh or no 100 minutes it takes 90 minutes

or 90 versus 100 and i think the whole 90 performance thing is one of those things that if it was that cut and dry, well you know what's 10 minutes? But I don't I suspect that the measurement I have to look at how they're measuring. I didn't see any little asterisks with a footnote on that slide at all. I'm assuming that it's 90 minutes it's when you get into the super large jobs and I just know if I was having a conversation a business outcome conversation with my executive and I said you know what you can get the answer 10% slower for 65% of the cost and or 65% cost savings the answer would probably be yeah that's not it's not really worth it. Yeah, I'm not interested in.

There was another guy I was talking today to the analyst session that said, our currency is time. And it was a Formula One guy. Yeah. I was pretty impressed with that. That was pretty, well, their whole setup was impressive. Oh, yeah. And the quote was actually really impressive. So if your currency is time, if you look at the scale thing, even the cost factor, you're not saving money in the real cost of SAP by putting the memory on DRAM versus storage class memory. It's the database licensing cost.

A 12-terabyte HANA instance is really expensive from a licensing perspective. Way more than the cost of the memory. Really? It's crazy. Okay, so alright so there's a whole licensing aspect of this. So in that case would, let's say if you split it up six terabytes in memory and six terabytes on the storage system. They're charging, SAP charges for the size of the HANA database. So regardless of where it's at. Yeah, you can take the performance hit if you want, but they charge for the size of it. So then they're only talking about the savings of hardware costs. Yeah, and hardware costs, when the last SAP HANA project that I did was somewhere in the order of 12

million dollars and the, from a hardware perspective. And the application team came to me and said, hey, can you shave $1 million off the price of the hardware? And I said, no. I mean, you guys are going to spend, what, $30, $40 million on this thing outside of hardware. I need all the hardware where I can get, you can find a million dollars in your own budget.

Some place else, some other budget or something like that. So it's not material. Okay, so we talked a lot about the SAP speed up and savings because they made a big point of that. But in reality, they started talking about response times at like 100 microsecond or 200 microsecond response time, and that's where the benefit starts to pay off. They had a couple of charts where, you know,

they've looked at their field base, and a significant portion of the field base will see, what, 40% performance improvements or something like that? That was the thing that I was mostly interested in because if you can take, if you can be elastic in your workloads and not be tied to that storage class memory that's inside a physical server

and that can spread that performance love across several different nodes as needed, now it gets pretty interesting. So that 65% savings I didn't care about is just that net new capability that I just can't practically do in the data center. Yeah, yeah, yeah. With this thing kind of opening up

a faster response time for applications, who knows where it plays out fairly well. Yeah, you know, you can have, the Formula One guy talked about his SCM, SCCM getting sped up, and who cares about SCCM? But it benefited, there a direct benefit to their business outcome so that was actually pretty cool yeah the Formula One guy was good session but it wasn't publicized so I don't know if we should talk about it or not oh yeah there you go all

right so another product they mentioned this week was the flash array C which is this capacity optimized storage what do you think of that? So, it seems like we've been chasing this promise of flash getting price parity with spinning disk. Forever. Forever, and you know what, you've been following the space way longer than me. I'm skeptical, I just am.

Well, you know, I had this discussion with Cos last night, the CTO, and you know, he's saying there's gonna be a point when there's no reason to buy disk. And I, God, I label into him big time. I mean, the 20 terabyte SATA drives. Yeah. So, I mean, who's buying disk today? I mean, besides the enterprises with hybrid and legacy concerns,

I mean, the boatloads of disk are going to the hyperscalers. They're buying it because it's cheap and deep storage. It may not be fast. It may not have all the ergonomics of Flash, but for cheap storage, it's not a bad way to go. Is that going to change? I think that, you know, Enrico mentioned last night, tried to hit me over the head with this set. You know, a disk drive can only do, let's say,

100 ops per second or, you know, 1,000 ops per second, whatever the number is. It's been flat for years, effectively. So back 20 years ago, you could probably still do that 100 ops per second per drive. Now you can do 100 ops per second, but it went from 64 gigabytes to 20 terabytes or something like that in the timeframe. So multiple scales of multiplication here. But here's the thing. There are ways to deal with that. I was talking with Matt today about there's these dual actuator drives now. It effectively doubles the number of ops per terabyte that they can do. So there's ways to deal with that over time. Is it going to be more costly?

Yes. Is it going to make it, you know, so that the cost of a disk drive that does that much IOPs is going to be on a dollar per gigabyte basis, much more expensive? I don't think so. Yeah, there's, you know, there's practical, there's the practical rules of space.

Then there's the trade-off of heat. I would love to be in a room with the AWS as they figure out the calculations, which is better to save the heat costs, to save the electronics or the physical space inside of a data center. Then you just look at the economics of a cheap and deep infrequent access S3.

That has to go somewhere, and it has to be cheap and deep. Is flash going to get to that point? I'm not smart enough to figure it out. I think, and I mentioned last night with KAWS, that tape was pronounced dead in 1967, and it's still alive today. And the people that are buying tape, the most part of these hyperscares,

they're using it for backup, which is archive, long-term archive. And for that purpose, it's fine. What has changed is tape in 1967 was primary storage. Right. And it got relegated to backup. It's almost gotten relegated now to archive-level storage. So those sorts of niche marketplaces,

our uses of that technology has changed. Will disk on grow the same change? Yeah. So primary storage, tier one, tier zero, it doesn't exist anymore in disk, unless you're doing hybrid. And so those sorts of things can change. And maybe at some point, maybe cause is right

that there'll be no reason to buy this. Not because it's cheaper per gigabyte, but because there are other aspects to make it usable and stuff like that. Just doesn't make sense. So what did they do with FlashArray C? We've been down this other path. a QLC, a quad-level cell storage device that's got the capacity of disk. It's got slower performance than Tier 1, so it's more like a Tier 2 storage, but it's an all-flash array. So it brings in all the benefits of all-flash, the economics, the power,

the vibration resistance resistance and all this stuff so it's a small flash array but big so they're talking five petabytes and nine you or something like that yeah so effective like 1.39 I think it was the number that they threw out with five terabyte five petabytes effective which is quite a bit of, that's a lot of capacity. I think that's world-changing capacity. You give most data centers five petabytes and a 9U form factor, they'll be okay for quite some time.

But, you know, QLC, you know, it's QLC. Yeah, so the QLC problem is that, number one, it's slower. Number two, it doesn't last as long. So you may be able to do 1,000 overwrites to a QLC SSD versus, you know, 2,000 or 4,000 or 5,000 to a TLC kind of numbers. But, I mean, and they did an okay job showing that even with 1,000 overwrites, I could still live, you know, for a lifetime of a storage system. And Pure has always had this evergreen thing for so long.

So if your storage runs bad, we'll replace it and stuff like that. Yeah, that subscription model that they talked about, which we didn't put on the rundown, was, I think, a big deal that they re-highlighted. A third of their revenue comes from a subscription model where if you service kind of thing. Yeah exactly. So if you start out with 50 terabytes you need 200 when you go to upgrade to 200 you're only paying for the 150. Yeah yeah it's an interesting play and they said that the current device will go out with TLC but next year they're going to release QLC and the intention is to keep the price the same and that sort of stuff. And really, they've modified the firmware of their flash drive, flash module I guess I'd call it, to be more

capacity optimized for TLC. So you know, we'll see if it makes its way into my Plex media server. There you go. If you could put a five petabyte makes its way into my Plex media server. There you go. If you could put one in a five petabyte media server, you got more media servers. I have a lot of friends. I will have a lot of friends. That's good, that's good.

Well, the other thing that was talked about a lot, Cloud Block Store has been in beta for a while, so it's been out there. They mentioned like 60 plus customers were using it, stuff like that. But now it's GA, I guess, and it's available on the AWS marketplace so you and I can go out and fire up a Pure storage array in AWS.

What do you think about that? You know what? I have mixed feelings about it. Pure is a hardware company. Of all the storage array companies, NetApp got kicked around in the news for being a storage dinosaur. And you look at Pure from a, their play has never really been software. But now they're making a play as having a consistent software experience, operations experience.

And that's what the game is. Yeah, it's what the game is. You look at VMC, VMware Cloud on AWS, that's exactly what VMware Cloud on AWS. So Pure is following that model, and I have to say, you know what? And of the storage rate companies, I think Dell EMC gave up that. I think you can still buy an appliance, a VNX appliance,

but they've gone with their cloud service, which is a software as a service solution versus this thing is, I'm warming up to the idea. I think, you know, and they gave a pretty good deep dive on the tech field day sessions the other day on the technology and the architecture. I was pretty impressed with what they were doing. And almost, you know, I would say say they took what AWS could supply and said what can we do that

gives you the best performance, the best reliability and availability in that sort of environment, which is quite frankly somewhat flaky occasionally. You know I was really surprised because most of the times I've seen this appliance model is kind of a controller with EBS storage attached which meant that making changes expanding it was disruptive it was very difficult and this virtual disk solution that they came up with was pretty clever it's pretty clever. You'd have to look. I'll see if I can dig up the diagram. I wrote on a bar the other day. Yeah, I saw that picture on Twitter. Oh, it's crazy. But any bet, they've got virtual drives on the back end, which are effectively EC2 instances with IO1

high-end storage, high-end flash storage, and EBS flash storage behind it it's sort of a read level cache and in front of that they got two high-end EC2 instances heavy networking heavy heavy processing power so those are effectively the controllers and the virtual drive instances are the drives and they got seven of these virtual drives it's like raid five kind of thing so yeah one dies that you don't lose performance. And the other one basically gets rebuilt from, as I understand it, from S3 storage. So behind the virtual drives is S3 cheap and deep storage, which of course is on disk.

I didn't want to say anything to Pure. But it's so they do reconstruction on the fly from the raid group which is all these virtual drives of the ec2 and and then uh over time they'll repopulate a new ec2 instance virtual drive instance rather with from s3 well it's an interesting architecture i mean i thought i give these guys kudos for taking a completely blank slate to things saying what does AWS do and how can I provide? High available storage nothing like any of what their competitors do so I bet I actually Absolutely tip my head to it like you know if you're gonna do something there if you're gonna do it do it differently It took him a while to get into this

To get into this market, but it is different and it is consistent operationally across the hybrid cloud. With high availability in AWS. I mean, these sorts of things just don't talk about in the same breath, quite frankly. No, you don't. Not high availability block storage in AWS. Yeah, it doesn't exist.

So, a couple other things they did which were kind of smallish but still significant. They've extended the number of flash blades that can be connected together in a cluster, which is kind of smallish but but still significant they they've extended the number of flash blades that can be connected together in a cluster which is kind of bizarre i can't so what was it before like 75 i didn't yeah i thought i heard the number which was already a pretty impressive number i'm like i don't know anyone i don't know that many pure customers yeah but uh in the flash blade the 75 systems and and and I think we both are kind of checking our math. I heard 150 blades. I think it's like double that.

So that's an insane amount of capacity. So per FlashBlade controller kind of thing, I think might be 15. And you stack these guys up, you get to 75 with five of them, I think. And so now it's 10 in a cluster. I can't tell you, managing data centers in the past, how many times I've been frustrated with bumping up against the theoretical limits, which are real limits, in the data center, and having to have to buy

a whole new group of controllers for a system for some of the other tier. Applications. Yeah, and it's not just a significant cost from a software licensing perspective and hardware licensing perspective, it's a significant cost from a data center footprint to have that overhead just because I've run out of a.

And not to mention the fact that you have to tech refresh and all that sort of stuff, which is, with the evergreen approach that these guys have, they pretty much, you know, provide that for you. So it's very interesting. It was some discussion today at the analyst sessions about, you know, all the AI stuff that's plugged into FlashBlade and how well it's playing.

I was impressed that even the Formula One guys had some FlashBlade stuff besides the FlashAr stuff that they had which was which is pretty impressive yeah and i'm i'm actually impressed across the market i know we did the the little short bit on computational computing a little bit ago i'm wondering when uh especially as as you look at someone like pure i would love to have a computational compute conversation with them. On the FlashBlade, they've got the intelligence. Yeah, but they don't use SSDs, so they don't have the, I guess, luxury of saying that they have spare compute inside of SSD these because they're using flash directly so yeah yeah but they've got some intelligence each one of those flash blades which which allows them to scale effectively with linear

performance and stuff like that so that that makes a lot of sense okay so moving on the last item of business was that they announced the file services on their flash array so both NFS and SM, and they mentioned some acquisition that they had that allowed, I can't remember what the name was. You know, I actually missed that acquisition. They named the company, I really hadn't heard of them. But I think functionally, this is a thing. Like, tier one storage arrays with flash services

are things that, when I was a customer, I asked for it. I hated that for the provider that I was going with, it was an expensive license to enable it. I thought it was ridiculous. I had to put a VM in front of my tier one storage array. We won't mention the vendor of that. I wasn't mentioning it, it was EMC, it was a VMAX. I had to, if I wanted to enable. They're not the only ones that do that. They're by far, they're the only ones that solution. I wasn't missing it. It was EMC. It was a VMAX. I had to, you know, if I wanted to enable...

They're not the only ones that do that. They're by far, they're not the only one. Hitachi does it. Everyone for their tier one storage, if you want file services, you're going to pay for it. And it's not necessarily easy to manage. So these guys embedded it kind of in their system, not unlike, you know, NetApp has done in the past and a couple of other select vendors.

And there was some discussion at Tech Field Day how you share the storage that's in the back with the file front end and the block front end. And they came up with a fairly good discussion. It was kind of done off camera, but in the end, it looks like it's completely shared. Everything is sent provisioned up front, and they both compete for whatever storage is behind it. So that's pretty impressive. Well, what else? Is there anything else you want to talk about at the Pure thing? No.

This is my first Pure event, Pure Accelerate. I just have not run into them as both a consultant or as a customer in the field. They had come and presented to me once. They told me their early SAP HANA story a couple of years ago. I didn't kind of buy it then, and now it's a little bit more viable. But one, it's a great event. And then two, these guys are laser focused on storage. They seem to be doing exceptionally well.

And they're growing. Their revenue is growing fairly decently. And they're obviously taking share from the other majors and any of the miners that still exist and stuff like that. And they are innovating in the technology. They're on the ball. Yeah, so I'm pretty impressed. I love Austin. They used to do, they did San Francisco, a couple of odd places in San Francisco, which were pretty interesting.

But I still, I like Austin as well. Yeah, Austin does good barbecue. And, you know, both of us like our barbecue and beer. So it worked out well. Yeah, it was a beer barbecue party last night. Yeah. Well, Keith, thank you very much for being on our show. And we'll try to get this out as soon as we can.

Ray, thanks a lot for having me again. All right, have a good day. Bye-bye.