CTOA Briefing - Storpool
Transcript
you have the ball and it's on you great um try to make it non-salesy and quick so i'll just give you some background so um we have some buzzwords here software defined storage but as uh i think we spoke uh last week we were doing this since 2011 very late 2011 so buzzwords like software defined didn't exist back then vsan was not even announced from vmware etc so i have a fun story about that but i will keep it for a bit later uh so we're
focusing on walk level storage uh so that's basically anything that's mission critical um databases vdi virtualized environments that are large etc so they need performance highly available shared storage we refer to it as ssd sds 2.0 with the idea of this is storage software that's not only api driven and real storage uh you know sds but it also is able to replace million dollar boxes sandboxes it's faster than local ssd but comes with all the data management features so it has quality of service built-in backups
uh snapshots clones whatnot that's you know features that you would expect from high-end uh storage array uh but it's as fast as all cos is this so that i think is very unique on the market to achieve that the third uh kind of bullet we build everything from the ground up we see a lot of companies that are just taking existing open source solutions or some kind of getting some components that they oem and they build the solution when we have taken the harder approach to
build everything from the ground up and recreate things that are entire domains for example our networking management we are kind of before intel came with the spdk and uh that set of technologies for kernel bypass acceleration of network packets on the network for ssds we've reimplemented or implemented our own technology around that and we have our own quorum management system and on disk formats and whatnot so you basically have everything underneath the you know we present we come about as a local uh drive but we're
not local to to the host and then everything underneath is our own ip so we typically replace uh cenro flash race mostly we compete with the guys like uh solidfire now netapp and netapp all flash and hp's three parts and you know kind uh like the vnx softwash and uh nimbo storage now hp and all these guys so so let's let's pause there for a sec because that's uh interesting approach and statement so from a are you saying that you're not only the faster the local
ssds i i can buy that and that's fine the the second piece was that you say it's it appears as a local drive to the server correct presented to as a uh as part of the local file system now we're a box so we're coming underneath that we appear as a road device and they can uh you know format it with a file system be dmfs or kind of uh uh any other file system so it's like rdma but your your is directly you're directly presented
to as a low-level device the system sees it form acid etc but is presented through this kind of proprietary network stack that you guys have built for basically already we have a a client a driver that's sitting on the host in hyper commercial deployments especially or we can do what's called now composable infrastructure or segregated deployment where we have a rack of servers on one side for compute and a rack of servers on the other side for doing the storage uh and we can present from
a single storage platform storage for multiple it stacks meaning let's say you're doing vmware on one side hyper v on the other side we can iso the storage um service to them but we also have native kubernetes csi plugins so you can from the same store called storage cluster you can service kubernetes with all the apis that come through the kubernetes csi driver or openstack or bare metal so we have i think very versatile one of the most versatile storage systems and why like on a
very high level you can think of stropo as uh the vsan for non-vmware stacks predominantly because in vmware um people would i will typically choose vsan and i think about vsan as an add-on to the hypervisor it's not a storage product per se because you cannot take hyper-v workloads and and utilize vsan storage right it's right kind of in their stack and it's a an add-on to the hypervisor uh if you want to have a proper storage solution that general storage like a storage layer where
that we can run hyper converge or not we can run with vmware with hyper-v with bare metal with you know we can expose storage to kubernetes to open stack you know we're general purpose storage solutions whereas like if you think about storage spaces directly with microsoft or vsan with vmware they're just add on to the hypervisor and are not a storage product but in itself for the wider market uh so we're kind of that general uh storage layer that can service multiple stacks from the same
system uh and it's deeply integrated with all of stuff so one of the things i was uh just complaining about the other day was how exactly how integrated vsan is with vsphere when you don't have a management layer above that it can complicate like people hear the idea of running workloads on the same nodes as storage and it seems like it makes sense until you actually have to do it so you know uh and you have to do it at scale so once you know if
you're in a small environment it's not that bad but once you start to scale out to a hundred a thousand nodes and storage and cpu are using the same underlay the same physical bare metal servers it gets to be a just to be a pain without having something to orchestrate it all so are you guys promoting kind of that use of of multiple uh of uh mixing computing storage notes or do you just say you know what build a storage array with x86 and then go
and then consume us as you would any other storage array all right i'll take a you know step back so as i said strop was typically used in what we refer to as new or modern i.t whereas everything is api driven everything is automated self-service everything's scalable and it's typically managed by a cloud management system of higher order typically we're for non-vmware or we're mostly in the kvm stack like if you think about the i.t consumer 15 years ago that would be a fortune 500 enterprise
that would spend all the all the doors on it infrastructure now a lot of the doors that are spent on it infrastructure are spent by major cloud companies be public private companies that are typically using kdm and linux if you think about how much vmware or netapp is aws using well the answer is zero right it's a strange question to ask for these new cloud companies and stroke is kind of like a tool for these new companies from this point of view uh we're typically running
with something like openstack or call stack or some of the other cloud management red hat virtualization layers that are typical for the new age or the modern i.t clouds and not the fortune 500 guys and in this case we're deeply integrated we're very lightweight as software meaning you know you can run stor stropo hyper converge on the same physical nodes or segregate it doesn't really matter you know because it's all managed orchestrated by a cloud management system already something that would be the cloud director for
example in the vmware stack so if you're thinking about if uh so i'm having i'm having trouble mentally separating because i keep thinking about the vsphere and vsan integration when you're when we think about hypervisor use but when you separate yourself out from the hypervisor and you're you know just a linux machine and on top of that linux machine you have kvm running inside of that linux machine you have some local storage disk that you're using to provide as the physical underlay for store pool you
have a client on that you have then you also have the client on there that a lot that presents the storage to kvm so at the highest level at the uh i'm sorry at the lowest level you have linux the linux kernel storage pool the storage pool software that creates your virtual storage array uh virtual pools of storage then on those same machines you can have a client that consumes the storage which uh uh makes it available which makes storage uh devices available to kvm within
kvm orchestrating kvm is a openstack uh or a uh kubernetes which provides storage vsf or some type of presentation software uh or this uh csi driver for uh kubernetes so you have basically these three levels of extra abstraction you have the storage array the virtual storage array you have the driver orchestration layer or uh os that's consuming that storage and then you're presenting that storage via some type of cloud service whether it's ceph or uh do you guys do object storage at all or support not
yet not the hdi uh this type of storage isn't people typically don't use this for object storage it's not a great use case for object storage in general but uh when i'm consuming if i need volumes in a container so for containers it's like a wonderful abstraction layer because i i can consume the csi driver i can uh get storage presented to me via csi and then i can create applications with uh persistent layers of storage that's backed by a traditional approach to storage software and
storage services correct yeah okay i have so i think here are a couple of diagrams one is the more high-level generic diagram you have servers and local drives in these servers and store port build installed you know a basic linux installation on each server and stroke installed on top and it will aggregate and create a virtual storage array of the capacity and then we can either consume it on the same servers if you run uh other workloads could be bare metal or virtualized in this case
with kvm with our own driver or you can do iscsi and present storage service to other with vmware hyper-v bare metal et cetera and we have also a csi gigantic maybe this is another representation of the same thing but a great gray rectangle here is a server a physical box and have two components of our software the store pool storage server that manages local drive stocks to other storage servers and present and create this virtual storage appliance and then you have our own driver that can
be used to consume storage service and you can decouple them you can for example in one storage cluster you can have boxes that are running hyper converged boxes on the far right side that are only compute node and boxes that are storage only and only present their devices to to the pool okay i got it yeah so i get back to the more general stuff so we're a software company we basically deliver software and we have you know some partners that can help and provide solutions
bundles with hardware uh we're focusing on primary flash we can also cover hdd backup etc but we're kind of typically you know tier one than you know workloads not a silicon valley company we see a lot of companies especially 2014 to 16 there were a lot of companies that we recor we referred to as marketing defined storage they had nice slider but zero customers and you know very if any product very weak if any product and we're kind of the i guess the complete opposite we're
smaller and under the radar but we have a very you know solid customer base with people that are voting for surplus as they call it the mortgage on their house and the college of their case like they're really using straw poll benefiting from it and um we're organically wrong vendor with a very very um solid product so um kind of um there's a question around optimizations for particular designs or hardware implementations you mentioned sdbk uh then you guys kind of predated sp dbk so you created
your own optimization at the network stack uh there's also dpk there's all the things that intel has done from a optimization perspective so do you guys run any differently on intel versus amd uh we can do both like we have a very wide compatibility list on the component level so you can have any vendor via dell or hp or supermicro doesn't matter we're certifying components and we are compatible with most popular ssd which is typically until samsung micron are the most popular these days cpus from
intel and from amd we're running both in production we even have an arm port so we can actually run our on arm arm and yeah basically also simulating for the more um uh you know problematic let's see or more uh picky components in in the stack which is typically the mix the network interface cards we've seen because some some of them to cause trouble although they're both you know you're most of the knicks in the world are based on broadca intel chipset so they're largely the
same devices but we've seen some particularities with some okay um here we are latest generation as we say but also very mature we've been running production for eight plus years and we have now 20 major versions and global spread of customers and one of the very few uh storage companies that are actually making profits or kind of i i recall early when we started shortcut it was maybe i don't know 2013 and a lot of people are saying hey you're a small startup how can i
trust you with my data because you might disappear and a couple years later uh funny funnily enough the largest storage company in the world emc got acquired and discontinued a third of their uh storage uh products product lines so that was a kind of a storage is not what it used to be so the funny thing is that you know they were looking for uh elusive um uh security in a increasingly secure world so we're still here and now emc is part of dell and you
know they don't have a lot of the product lines that they used to have so that's an interesting observation here um yeah i mentioned some of the things uh were very good solution for companies that want to build a modern private or public cloud and this these clouds are typically not built on vmware or hyper-v or solutions that are managed by gui these are solutions that are api first and everything is done through the api is automated so um you know if if you think about
where i t is going at least our you know how we see the market is um things like ascend is going away like if you think about kubernetes or openstack movement you're thinking about how can i have infrastructure as code how can i have software run on standard server and emulate storage array emulate network accumulate whatever i i need to do and it you know it doesn't it's not a very good fit to have a kubernetes stack on one side and a netapp box on the
other or they want to get rid of all specialized hardware so i think this is the world we were kind of very good at and for that reason on point seven we typically appeal to guys that are managing um a sizable infrastructure as a service platforms be it for hosting their own applications or being a pass or eos platform or msp and cloud hosting or some sas companies etc so different use cases uh but in all case we're focusing on the block so that's typically databases
virtual machine disks uh vdi uh environments or larger applications that want their metal and need a lot of performance and short storage on the back end so let's talk kind of hero numbers what's the how big does this get uh there you go we're getting into the here's like a small system of example of what our smallest system is faster than you know most of the systems that are can be delivered by other vendors on the market so we typically nvme systems is the typical thing
we deliver these days uh triple replicated sure storage system does less than 100 microseconds latency and easily does above million ios per server and it's a scale out system meaning 10 servers would do can pick up a peak at about 10 million iops so tremendous performance and that's with all the functionality so you have solutions on the market for example like accelero that are very fast but are coming as hardware and also they don't have a lot of data management uh functionalities um so they trade
performance at the the cost of you know data management features and we're having all these inside and still tremendous you know performance uh very um good soa a fully managed solution so we can also help our customers so for some of the their storage problems they you know become our problems and already storage expert that you know helps them to have a very fast and robust solution that's working and they can focus on core business uh the system can scale online so it can start at
a 10 terabyte a very small system and it can online scale to multi-petabyte system to your question you know you can have a system that that does let's say two three petabytes in a single cluster most customers typically do something similar to the aws availability zones so they'll say i'll do one petabyte system and then i'll have a store pull against a single pane of glass that can manage multiple clusters so it can appear as one system and shift workloads and virtual machines from one to
the other but you have the availability and blast radius concepts so you can comply with security or design considerations it doesn't make sense yep um [Music] let's see i have this [Music] here is a representation of where we play we've taken here the uh basically four types of storage block on the bottom containers file and object and blocking containers you do natively file an object you can do with a gateway architecture and here is like where we play we also kind of slice the different tiers
of storage and we've taken some prices on the around right hand side from aws the public cloud just as a just to show that there are multiple use cases at very very different price points that you know have uh 10 20 times difference uh sometimes more so store pool is playing mostly in the tier what we call zero tier one uh systems so that's nvme or general purpose sub ssds and we can reach down to also uh primary hd and backup hdd systems and this is
where we play basically in this type of uh segment in terms of use cases and uh storage tiers that we cover and then if you want to do for example a small object storage typically we've seen a lot of guys interested in object storage but they do fairly small object object buckets so that can easily be kind of a gateway architecture we're running a menial for example gateway in a virtual machine on store posts or having high availability etc and running a small let's say 20
30 terabyte bucket of um object storage uh if you need that or file with an nfs gateway in a similar fashion so we can in a way have a story for object file and block but you know the book and containers is very native all right uh some customers here we have you know nasdaq in dubai we have a lot of workloads from mission critical pretty much all of our customers are most of our customers are public clouds or publicly traded companies in india and we
service workloads for very demanding customers like european space agency nasa siemens the deutsche stock exchange we have a lot of you know workloads like that that are running on top of store pool and i think that's that one on the five slides that i said i'll show you otherwise i have plenty of other stuff but i think on very high level this is kind of like a crash course to the company uh boeing actually i ended up ended up being a half an hour a little
bit less than a half an hour exactly so we uh we have plenty of times for slides and conversations
