Current state of NVMe over Fabrics- VMware VMorld 2019 Announcements
Transcript
[Music] all right welcome back to the CTO does my name is Yuki scar I'm joined by Howard Marks and we're we're gonna talk about you know some of the developments here at the at vmworld you know what's going on what are we seeing what's making our hearts think you know what do you eat what we see that makes us happy so no I want to first you know kind of lead with you know if there anything in particular they kind of made you happy this week
then you saw well there's the there's the things that make me happier they're I understand yet kubernetes is the hot new hotness just like it and I just AM your storage guy so you know it's interesting to see how the evolution is going one of the things I'm really happy to see is VMware re-engaging on V vaults right that you know evolves came out five or six years ago you know I spoke five times at that VM world about how wonderful V vaults were going to
be everybody got really excited and then two or three years later everybody that really ups really disappointed that people hadn't adopted it right but it should have been obvious in order to get V Vols working you had to replace your hypervisor with the latest version that supported meatballs you know and then you had to also replace your storage system with the latest version that supported the meatballs and that's at least three years and two years after meatballs came out people were calling it dead well now
we're five or six years after V balls came out and VMware has gotten pure and HPE and to a lesser extent Dell EMC because they don't have the best implementation but they've gotten all the vendors in a row so now they move back to users and go okay you have featured array you just bought you have this feature the hypervisor you're just bought turn them off start using array snapshots instead of VMware snapshots which apologies to my friends such an Pagani stink base tonight and so
you know I'm excited that VMware hasn't decided well the sand is the answer or VMware storage and that's the way we're going to do it because that would win but the storage renders out of business which to many companies but decision to many companies were met yeah so I'm interested in that right I'm no longer an independent observer of the storage industry I'm now a participant I work at vast data we make a very large storage system and we are not quite busy in the VMware
space yet because the number of customers who need many petabytes of VMware storage are small that's a large yeah it's a small number of console yeah it's a small on American stores but the technologies that underlies what we do envy me over fabrics yeah it's just on the cusp of coming to be a more yeah I mean I haven't heard much about it this week I mean there's you know tidbits of information yeah a session hero there but today what we're seeing is storage systems using
nvme over fabrics on the back end so our system or pure flash array X instead of using SAS uses Ethernet and nvm you for fabrics to connect the engine chokes in their model it's interesting because it reduces latency in our model that's more interesting because nvme over fabrics breaks the two-two-one connections to a divisive limit that's asshat so in SAS you have two controllers and you have a dual port drive but you can't connect three controllers or five controllers or acquainting controllers to an SSD or
arc yeah in a fabric Ethernet sit in between it so you can have an arbitrary number of controllers access an arbitrary number of SS that's going to be many many right and so that many to many starts to become very interesting for you know what we do we fronting that with software and build a very large storage system or even with a system way these fan if instead of having the SSDs and visa and in the server they were in an external cab it's still owned
by one server hmm when a server wannafly you wouldn't have to rebuild the data and have all that traffic and load you just have to have another server attached to the SSDs that used to be attached to the first server sort of recovery can be much faster and things the complexity that HCI adds to the maintaining a cluster whatever goes away yeah because you know when you install these in all of a sudden what used to be a stateless compute server becomes a stateful server storage
circle and you have to treat all of your servers like storage yeah yeah it's a track and as an in-flight storage guy that means I'm a fairly person because when things go wrong with storage things stay wrong with starting so managing servers like storage means what we don't don't do vMotion even though the day we don't destroy your upgrades and the lid we declare it how an edge because something could go wrong and I'd much rather run my vSphere as stateless and be able to do
that work in the middle of the day than have to come in are we gentlemen do notice things so I'm really the the nvme River fabrics is just getting to that stage of maturity the next step is for VMware to put the driver into the hypervisor which is probably the next version yeah it's talk to it I've talked to enough people who cited VMware that it's like yeah it's working inside VMware we're not really ready you know not quite ready to release yet so I'm looking
forward to it is it gonna be a same cycle it just has with visa and you know people are going to talk about it you know probably next year and then it will take a couple of years before adoption I think it'll be faster because the storage system you're already ready yeah so there so the problem with V valves was there was one starting point in both sides had to start at the same point the storage guys have already started so now it's waiting for VMware
to catch up as opposed to waiting for both of them to reach some both state of maturity and having been around long enough that people have already bought it yeah you know because people replace their are gonna three four or five year cycle and so a new feature to get to ninety percent penetration that's five years absolute minimal it salts what's that I what's the use guys of this I mean I get it in in like a large media company but why would you want to
use this in any environment well there's there's the cleaning up of the how HCI works up pieces and there's just it's going to be the future that SAS is to use a term from men and black old and busted well it's it's a point-to-point technology it doesn't get faster as fast as Ethernet gets faster it gets more expensive every time it gets faster because the number of customers who care is much smaller and so being able to use Ethernet to connect SSDs the server's is just
over time when it become a better and better idea because the old versions don't get updated there are latency and performance advantages but frankly few viens need more performance than a fiber channel also flash of a provides yes the new systems are faster but that's like you went from the Dodge Challenger with the 600 horsepower engine to the Hellcat with the 800 horsepower engine but you're driving in San Francisco where you can't go uphill faster than 20 miles an hour anyway yeah so the need just
isn't there you know a couple of years ago away send out a request for help to the day field a group saying how many AI ops does a server use does it be an amuse and that group sent me about four yeah it's data on about 14,000 VMs yeah well I remember sending my and the average was just over a hundred hours yeah that's yeah most people have most people have made a lot of not very active via so the performance isn't really the answer it's
the flexibility days yeah that's kind of where you know I'm still I'm not entirely seeing the use case in terms of you know your bog-standard traditional fees for a customer I mean they only have a couple houses maybe a couple of dozen hosts it's not complex I mean whatever I'm moving to the cloud how relevant is this technology gonna be I don't believe everybody's moving to the cloud right you know there's I have dealt been dealt with too many customers where the workloads just don't fit
the cloud where the locations don't fit the cloud and I used to used to have a claim the central soya company they processed soybeans at the soybean oil and soybean meal and animal feed their plants are literally in the middle of nowhere yeah and if you say we're gonna take all the applications that run this plant we're going to put them in Amazon miles away well where there's only one ISP that serves that plant someday in the fall when farmers would bring trucks of soybeans to
that plant that link is going to fail yeah and the plant has to shut down so there are you know if you are in a major city where there's multiple providers will give you gigabit connections then and then you can put everything in Amazon if you have to do plant automation and things have to run even when the internet goes down you have to be on premises and at the large scale you know Amazon is more efficient than the average IT department Amazon's not that much
more efficient than a than the leading well-run IT departments and Amazon has a profit margin yeah and so if you're JP Morgan or Bank of America then you're gonna keep a lot of things on premises and then there's the dead rat problem because if you literally are JP Morgan or Bank of America the core application runs on the mainframe yeah there is no mainframe as a service in AWS or GCP but there are 4,000 VMs that run on x86 is that access same data sense as
the mainframe application that is two thousand joules yeah and the latency of moving those servers to AWS so that they have to do database queries against db2 and the mainframe in my data center it's gonna make those things bright yeah that sounds like a good story and so you know move everything to the cloud is easy why don't we refactor all of our applications well it took 12 years to write this application the first place and everybody who wrote it retired to refactor it just to
understand what it does is gonna take 12 years and it works now is that yeah if I move things to the cloud do they necessarily work better well sometimes they do no question but being in the cloud isn't necessarily better sometimes it's better it is yeah yeah no but that makes it harder I mean you have it working now you know don't touch it even well you know that creates additional technical debt right so and I find it amusing when young people hmm talk about eliminating
a technical debt technical debt is like the federal deficit yeah you don't eliminate technical debt you refinance technical debt you take your old technical that you're paying twelve percent interest on and you roll it over to a new node you can make four percent interest on and so your old system was you know and so the system's reach the point where they get really creaky and you fix the ones that are really creaky and making them okay just good anymore you know for you or you
know if you're a well-run organization you perfect those but there's always another one you know ten years from now what His perfect will be creepy yeah there's always technical debt you don't eliminate it you just roll it over and and so the whole let's clean everything all at once it's kind of like okay I was involved in that I was a senior manager at Deloitte and a large food company hired us to replace their entire IT infrastructure because y2k was coming yeah that company spent thirty
eight point four million dollars with Deloitte and was sold for twenty four million dollars nineteen months later sometimes oh I didn't clean up all this technical debt and you'll have real dollar debt and won't be able to keep the business going yeah it's either way yeah there's no winning all right so I mean back to victim of you mean for a little bit are there any sessions on at vmworld where you think okay this is a good explanation people should go here I the sad thing
is I have not spent any time with the schedule but I'm here to evangelize vast data and to talk to my friends in the community who I've been lucky enough to make over the years who are in positions to tell me what their needs are as a customers so that we can refine the product and so I'm really here to talk more than to go to the sessions when I get home the sessions will be on video and I can watch them I saw Craig waters
from pure on the Brownback stage this morning doing something on NVB but that's because I happen to be in the neighborhood so all right so I'm not the person to talk to about what sessions are really interesting sorry so the good tip is after vmworld you know everything's online look it up if you're a if you're in system so I guess our time is up absolutely thanks for for being here around always a pleasure you and I will see each other next time I'm sure
