sfspark.org: Neil Gehani Lightning Talk
Recording: sfspark.org: Neil Gehani Lightning Talk
hi thanks thanks Alexi thanks for having us here uh as he said I'm from mesosphere how many people have heard of mesosphere yeah that's pretty cool okay so just a little uh brief history mesosphere mesosphere the m in smack uh is what the company was formed on uh we actually build a distributed cloud-based operating system what does that mean that means that we we treat your data centers whether they be on the edge in the core or in the cloud and make it look like one one virtual data center if you will and that allows you to take your workloads and deploy them uh anywhere and and manage them and that's essentially what uh what dcos is so that's what I'm going to talk about I'm going to hit next and he's going to go there um so if you take a look at just what's happening with data and I'm I'm just using this as an example you could apply this to healthcare but I'm using the connected car example here you can see the amount of data that's being generated I think there was an article that was posted there was four terabytes of data per car per day right that's a lot of data to ingest uh and that data comes from all kinds of of pieces of information and they come into somewhere that gets processed where do they get processed where do they live how does it actually happen uh and that can happen at the edge you can have like we have a customer that actually has data centers on ships right they're mobile the data centers are actually moving most people when they think about a data center they think about it at being fixed in a building somewhere right and we don't think about them that way we think about data centers can be very very small they can be mobile uh they're in airplanes they are smaller versions they and they're going to be in cars and Trunks of cars that are processing tons and tons of data and making decisions in real time that's where machine learning comes in that's where AI comes in to take all of this data and to be able to then use the information to actually make decisions so this is the one example that we're talking about but there can be other examples as well hit the next button so what we try to think about it because we're a platform company fundamentally we we want to enable smack stack and we want to make it really really easy that's our mission our mission is to make this smack stack extremely easy to build it and deploy it at scale um so we allow people now in in our platform to be able to deploy any one of these big data services with basically one click or one Command right you don't need expertise to do this you also don't need expertise to upgrade maintain manage it monitor it all of those things that changes the game when you have it in a single silo or a single place when you have it in a cloud-based operating system then those challenges are are are quite U difficult to handle but we try to make that really really simple and that's our mission and that's what we're trying to do here uh deploying each data service is extremely timec consuming uh and we try to address that it takes a lot of people to do that uh also operating this thing is extremely error prone uh so we now have the SDK which allows uh people to build like lightbend is one of our partners uh so we actually build spark Cassandra Kafka uh all of these things are built on the SDK so we handle all the upgrades all the maintenance while the while those things are running so you could upgrade from one version of Kafka to another version of kka or you could downgrade right it's it's pretty straightforward and pretty simple and as you know Cassandra Kaa they they have their own way of doing those things but we enable that the other benefit that we have in the platform is to make sure that these big data services and stateless applications that have to use these Services the developers are building they can run on the same shared infrastructure and that's important because the resources that you need today if you think about the way people use VMS it's like very lightly used we try to make sure that those resources are right you know you get to 60 70% utilization rather than 5 to 10 which is what what's on average when we think about resources we think about all the resources in isolation so if you think about CPU memory dis Network volumes as well as gpus then eventually it'll be tpus as well if if Google gets you know Google's working on those things so because we deploy in any infrastructure it doesn't really matter whether this Amazon Azure or Google because you can move your workloads easily if you're running it on dcos then we can take advantage of those specific resources you can have a spark job scheduled to use the GPU like we we actually support nvidia's GPU uh out of the box it's available within dcos so you can take a job and say I want to run it on these nodes that actually have the GPU running and you can schedule the jobs on gpus and those resources are shared uh across your entire infrastructure okay that's our goal that's our mission that's what we try to do we want to make the deployment really really really simple so all of the smack stack that you can think of is basic basically one click install it's a tile install it it will start running okay that's how simple it is today all the upgrades the operations are metrics uh management you can write to statsd push all the metric statd we are the producer of those metrics so we take all the infrastructure metrics as well as enable people to push the metrics through D through dcos and then you can consume it anyway we we can use the the same Kafka to consume the the metrics through Kafka and then Kafka can send it out to Prometheus elastic sta elastic elastic somewhere else that it can go so your end consumption can be different but we produce all the metrics and we allow you to run those metrics infrastructure uh out of the box same with the logging infrastructure is out of the box and this is the advantage of having both stateless containerized apps as well as stateful applications running on the same uh shared infrastructure with same shared resources the other thing we're also working on in terms of uh resource utilization is to make sure that we can optimize for these big data workloads so if you have spark and wants to wants to schedule a job on a GPU or if you're scheduling a job on a set of resources you can also prioritize these things you can have reservation and quotas for particular type of streaming jobs okay or any type of big data services that you want to do so light B will talk more about about that as I mentioned Nvidia is is something we support already out of the box so you can now see which jobs are running on the GPU and you can decide that I want to run this particular job or this particular task or this particular workload and leverage the node that has gpus there you can actually schedule it on that specific node so that allows you to take your workloads if you need GPU Services then you can do that obviously if every node has the GPU services then we make it look like one giant pool of GPU services that GPU shares that are available to you the one difference between CPU and GPU is that today if you want to run a job you're going to take the whole C whole GPU where CPU can actually do it in shares so if I have a thousand nodes and that's I have I don't know how many of cores there are we think about them as as CPU shares and the workload gets shared among all of these CPUs but if you're running a GPU you'll say Okay I want to run this job on this particular set of notes that these are the only notes 10 out of a thousand have gpus in them and you want to pin it to that you will pin it to that but then you're going to use that whole GPU you can't like cut the share uh take the shares the GPU but uh once we can work with Nvidia and partners to make that better then we can also have that available as shares okay so this is this is what we're trying to do want to make sure that uh we providing data Rich applications here with the smack stack and we are the platform Ider so we allow you to run all of these things out of the box these are all available today so no no reason to worry about it you can hit the next one and that's it thank you a