Advanced hardware enabling modernization and scalability of SQL Server - BRK3218
All. Right I've turned 45 on my clock so in. Order. To respect everybody's time here let's get the party started shall. We so. This. Is hardware advanced advanced. Hardware enabling. Modernization. In scalability, obstacle server my, name is Arjun is Fernandes and I'm a principal program manager with the Microsoft sequel server engineering, team. My. Job has two aspects of it if one aspect, of it is the business aspect, of my job which is handling, the relationship, with hosters. Which are the largest sequel, server customers, in the world those who actually serve sequel. Server provide sequel server as, two other end customers, as a service and. My. Technical, role is. A feature. P.m. for the sequel server engine managing. Their persistent, memory feature which were definitely going to talk about tonight today. Okay. So here's. The agenda, we're. Gonna get started with a little bit of reminder. Of how. Modernization. Of the lack of modernization, in your infrastructure, can. Actually, impact your business seriously. Impact your business so. We're gonna talk about that and then, we're going to talk about the, latest generations, of CPUs and how they can make a difference in the performance and. Scalability of your sequel server instances. We're. Going to talk about securing, clips which is a new technology, that were embracing confidential. Computing, which is a new thing. That were fully embracing, here at Microsoft it's. Actually not new at, all if, you guys have each and every one of you guys has a smartphone, has a secure Enclave on your phone we'll. Chat about how that. Actually translates into your, servers. Incoming what the concept of the concept of confidential, computing. Persistent. Memory my, favorite topic we're, definitely going to talk about that out of this act language. We're. Also going to talk about the latest generation in shared storage technologies. This. Is always, a fan, favorite and. Then we're also going to talk about the latest generation hyper-converged. Solutions, because these are so I think I'm basically touching on all the buzzwords here, it's a pretty. Much you can do buzz word with bingo here so to speak if. You notice I have an accent and I. Can. I am happy to repeat something I said if you guys don't understand what I'm saying it's big fast so. Please raise your hand if you notice I, did. Not quite see with all the lights but you know if you raise your hand I'll try to slow. Down and, try. To not, mumble as much anyway.
Please, Turn off your cellphone's much appreciated and. Let's. Get started so your. Business actually hurts when you don't modernize. You. Think you're actually saving your, business your business money by not upgrading. This. Is actually a very common misconception I see it all the time with customers, that are still. Running a Windows Server 2008, r2 and sequel Server 2008 r2 please, raise your hand if you're running a sequel Server 2008 r2 totally fine not gonna charge you there's. A lot of you there's, a lot of you so. We're. Definitely gonna have a little, chat about that in just a minute. But. I really want to emphasize a few points here they. Impact to your business and all modernize and basically translates, into a few things number, one is still performance, okay, your performance, is not gonna get better when. You don't modernize period, right, you can only optimize. So, much you, can only spend so much time from your engineers and your developers to get the code faster but, the reality is at the hardware that you're running on and the, software that you're running on so whatever a platform, you're running on if you especially if you're running on an old version of sequel server like. 2008, r2 that's. Definitely going to go in detriment of your performance, so this. Is something my good friend Glenn Barry from sequel skills says nobody. Ever complains about a fascicle server but it only takes one slow query for people to. Start. Racing the red flag and start you know great grade creating, the p0, alerts and all that stuff right so. It's always better, to, over-provision, blog then you know, try to stay within the boundaries soft performance right there. Here. You can coughing your developers, when you don't modernize. If. Your developers want to use a feature that's, been available in sequel server for many years, and you're, still running on 2008, r2, you're. You're, basically. Turning. Them off yeah they're really not going to be excited about what they do of what they're gonna go work somewhere else the, reality is that they are not excited about what they do they're, gonna start looking forward, for a job somewhere else and you make. Their job infinitely. Harder right there are things that you could solve easily with new features that we have released over the years in the engine, yet, the, the fact that you are not operating it's limiting them. This. Can force you into a very expensive CSA, if you guys haven't heard about CSAs. We. At Microsoft gave you the traditional, mainstream, support and then the extended support and then when all is said and done when. We call in the support still, a bunch of customers come to us and say oh you, know what I still want to continue you. Can you please support me can you please continue to make patches for this, we'll. Do that for you but we'll charge you and, we're going to charge you big-time and when, I mean big-time it's. Big time but, this it's going to cost you way more than the actual cost of the license over the time that you enjoyed, it okay, so we'll talk about that a little bit further when we talk about sequel. Server 2008 r2 and a support. This. Can definitely force, you into an instance problem right so if you have your single server to fascinating r2, instance that's, running happy you don't even want to touch it because you know if you mess, a little bit with it you're afraid that things are gonna you know go creates. Parts and a bit fire is gonna come, up in your datacenter well. What, are you gonna do right you, don't want to touch our instance I can't throw another database in there because it's fixed and you, know resources are fully allocated on that machine there's nothing else I can do with it you know what let. Me just put another instance of sequel server right here and that. Just means an, additional, cost your business so. What why. Do you think you have so many sequel servers right. Not. That we another, we don't enjoy it as Microsoft, right we're making a lot of money from you guys but the reality is I you. Could do a lot better than that we, head to an solid Asian is a big big big thing. This. Definitely. Affects and negatively affects your business agility. Here's. The story I used to work for a company, where the.
Fact That they were running on a single psycho server instance, limited. Then on, the. Actual, features, that they could support they. Were actually, intending, to build a new, feature that needed to query the database very. Heavily, and because, that database. Was monolithic, servers just one instance I was running the entire business which is going to do it that server cannot withstand upload, we. Tried many things couldn't. Get it to work because it was a monolithic design not. Only that but, we couldn't we couldn't we, couldn't even get it operated when you were. Oh there, because you know it was the main concern that the Davis Google could something could go wrong with. A new hardware well, something will go wrong during, that migration and, that. Limited the business once, they finally migrated, away and implemented, this feature that, business made an additional four million dollars on a yearly basis just, by upgrading the sequel server instance, and being. Able to run those queries that were that were couldn't run before so. You think about that think about what that means for your business if. Your your, business sees you as a, blocker, you're. In trouble. If. You're not modernizing, because. The business, is preventing, you from modernizing, well, then the business is also blocked because they cannot move forward with things that they wanted to do because. Of technology and, these are all problems that are totally so verbal. There's. A set. Of good sessions, tomorrow, here at ignite that my. Colleagues. From the sequel, server Tiger team are delivering. And they, are fantastic, they, are about upgrading, your, secret servers the right way that's there's actually. Quite a few I'll mention them at the end but you guys definitely want to attend those it's. Not just about the infrastructure, it's not about the process of getting, Seco celebrated, right. So, it's a whole journey it, doesn't take very long it just takes a little effort and know that how you're doing the right thing when you do want to do the right thing. This. Is actually very key. It. Can not, modernising remain remaining, on role version or sequel server can. Leave you out of compliance. So. If you're if you're a company that wants, to grow and needs to be under need for example needs to be under PCI compliance, if, you're running a version of sequel server before which you're getting no patches, for, non security vulnerabilities. That are being patched or newer versions you're out of compliance, and. That's. A big deal that can, cost you but I can encourage your business into stiff. Penalties. So. It's a big deal. This. Is something, that I was discussing with somebody yesterday and, we will fully agree with agreed. On this the. Tangible, plus intangible cost to. Your business of number, in icing is much. Higher than the cost of removing the blockers the. Thing sad, removing. The blockers it's, also hard to it's hard to quantify. But. If you attend that session tomorrow that my colleagues, are gonna, word they're going to discuss or grade, strategies. For sickle sort of modernization, strategies for single server then, you're going to feel a much better position to do this okay, so I highly recommend that you guys attend that session. So. Here. We go this. Is the, why 2019. Problem. Sequel. Server 2008, end sequel server 2008, r2 are both going end of support, you. Only have literally, did the math just. After midnight you. Have 200 native six days to upgrade that's about, nine point four months. Have. You guys started yet raise, your hand if your 2008, r2 and you have started. Okay. Great I like, it I like it that's great so. You. Definitely want to attend that session tomorrow because that's also going to give you a ton of more ideas on how to accomplish your upgrades successfully, okay. So I'm, very glad to see that. So. The thing about Indo support for sequel server 2000 or 2008.
In 2008, r2 is, that if you run on Prem. You're. Not going to get any more badges patches. For, security, issues or. Anything, starting. On July 9th, on, of 2019. So next year, if. You're running on the cloud if you're running on it sure VMs, or Azure managed instance you. Obviously, get. Support. And get, secured data you. Get security patches free for three years so. If you're running on Prem and you don't have a plan to, migrate to the cloud yet you, should get you should get going, on that because, that's going to give you an additional three years to work on your upgrade path okay. So our VMs sequel, server 2008, r2 will still get three. Years of patches, in the cloud the. Alternative. To that is, getting. A very expensive custom, support agreement, that, will basically, cost, you 75. Percent of your license cost per. Year so. You're gonna pay. What. Two point, the cost of your license so, it's gonna cost you a lot of money your. Best bet is to either go cloud or, get, going on that upgrade. They're. Going guys. Let's. Talk a little bit about the latest generation of CPUs, one. Thing that I stress. To all of my customers especially the big coasters right all the, big companies that offer sequel server to other customers is. We. Have a perk or licensing, model okay. We have had this ever, since the sequel server 2012 and, that. Means that every, CPU, cycle, costs you money, so. If your queries, are taking too many CPU cycles that's actually costing you money I was. Talking to a friend of mine who works on a company that makes a monitoring, tools this morning they were making a tool that actually that is able actually actually going to be able to tell you exactly how, much I call a query cost you to run in Azure. You're. Able to say to tell these things right because you know exactly how long your VM has been running what, kind of instance you provision for your oracle database and how long you're pretty good to run so hence you could calculate exactly, how much of query is running so that's great right you're not going now go to your to. Your manager. And say hey look, this. Code, sucks it's costing us a lot of money that's actually that's actually a great idea so. If. You're running on an old CPU. Your. Query search just went to a slower but the very nature that those Abused are old they. May have higher. Clock, frequencies but, guess what the clock frequency doesn't, matter when, the architecture result a newer, newer, CPUs, even a lot lower frequency. Run, way faster than the older CPUs, so. There's. Some that there's something I want to I want to show you this comes from our friends at Intel this. Is a demo that they did I believe they're showing it at their booth as well so you may want to go visit them and talk to them about this but. This is basically a comparison, of, how. Sequel. Server runs, on, a. Hash. Well CPU, versus. How it runs on a coelom Platinum CPU skylake, CPU so. Let's get that going for a little bit it's just gonna take a couple minutes but check, it out you. Have your traditional analytic. Queries right so you're super in priority shipping, modes I'm an auto part this kind of revenue. Blah-blah-blah-blah-blah. So, check out the right that's, actually running on a xenon platinum CPU and then 80 81. 82 PU so pretty. Expensive CPU, with like 56. Oh, logical. Course but. Check out how much faster, ran than a box running a sequel server 2008, r2 that's. Just running on sick with Windows return 19 a pre-release, build and, sequel server turning 17. So. It's, not it's. Not just about the CPU though. CPU. Is one thing CPU is going to give you yes, it's, going to give you a great it's going to give you a nice boost, but. The reality, is that leveraging. The features in sequel server, that. Are fundamental. To performance, and the grading. Eventual, innovation, that has gone into the product as, of late like columnstore. There's. Other things I should be using today and you're not using them because you're running an old version of the code server ok, so this is another yet another reason to modernize.
It's. Something that you guys don't need to think about they're, preventing, your business from being agile. Imagine. How much faster, everything will be in your enterprise if. You actually kept up with the times, right. So. Yeah. Like I said this takes oh this takes about a couple a couple of minutes and I'm not gonna waste you guys this time just watching paint dry but the reality is that, it's. An average about a 12x improvement, of performance from, both hardware and software. Right. In some cases they actually saw 25 X increase. In performance for, data warehousing workloads. They're. Not impressive, 25, times faster, just. By operating and using. One feature rebuilding, an index. It's. Kind of amazing so. This is something that everyone should think about if, you're if I'm running on 5-year old hardware on a 10, year old operating system in. A nine, year old eight year old version of single server I'm hurting. My business badly. Taken. The CPU is not just about picking the cheapest. Or. Picking. The fastest. My, friend Dylan Berry again from sequel skills he he has an extensive, set. Of person patience on how to pick the right hardware physical server and I highly encourage you guys if your infrastructure, people if. You're if you're in charge of making the CPUs and the the kind of servers specs, for, your infrastructure. I highly. Recommend that you look into them but, here's an example, I. Actually picked this from one of his presentation so special, thanks to him. You. Can see here. It. Doesn't go very well but the top line you can see the T PCE score. For. That box there's, only 680. 3.79. Only 26 or 94, they, are the exact same generation. They're both V Force if. They're both ePHI, v4, but. Because of the difference in clock speeds in the core count and. The. Way that those superuser our. Packaged, orsk or so. That's different skews you. Get completely, different performance characteristics, so. This is something that you guys don't think about you know you should just, get a latest thing on no it's not just about that and. It's, so much more than that you actually need to think about these things as you make purchasing, decisions to, be intelligent, about the amount of money that you're spending on your single server instances, if. You if, you go out the top you're actually getting worse performance and you're paying top of the price. So. Why. Do that, be. Smart when you pick CPUs it's really the best way, to go about this. So. Let's talk about the neck the next section of the presentation which is Sakura glaze I mentioned. Computational, computing, as something that's fundamental, to Microsoft, moving forward every. Single one of you guys has a secure, Enclave on your smartphone, so I pretty much every single smartphone, today has a secure. Enclave and what, that actually means it does is that there's a protected, area memory, a hardware, backed protected. Area of memory that no, process, can access unless, it actually has. Privileges. To do so okay. So the. Way that we're translating, this, into a, single server is, the ability for us to now, protect, secrets. Within. The engine itself, even. Work from from, within a local administrator from, from looking at it, so, this. Will be this will be the output of windbg. Run. Against an area memory dump of the in area of memory that's actually protected by an English. You. Cannot make any sense of that data. Right. It's actually, protected, by, either, the operating system in the case of DVS. Which is one of the ways to implement secure, anklet that's, how what we are rolling out on, sequel. Server 2019, CTP 2.0, and, then there's hardware backed secure, Enclave, that's, ignored new technology, from Intel call into FS, GX, which. We are we have plans to support shortly. After the release of CTP, 2.0, for. More information, on that you guys want to go to my. Colleague jacobson, session on. Confidential. Computing on single, server the, link is at the end of the presentation, so you'll be able to tell, the code etc well. What does this mean for sequel server. If. You think about how always. Encrypted works always.
Encrypted Keeps keeps, the encryption, keys. At, the client level so, that way the database administrator, has no way to make sense of any of the data and, all the encryption, happens. At the client level that. Means that on the sequel. Server side we're, actually restricted, on the amount of operations, that we can perform in, that data itself because, we don't know on the engine itself we don't know what. We can do with that data right. Because. You, know it's not like I can perform just a like search. On that index that's now encrypted, I can't. Write, as the engine I even, the engine itself has no visibility into what, data is being kept in there so. With, a secure Enclave we, actually can perform some operations, within the unplayed itself without. The administrator, the machine or the database administrator or even having access to this area memory but, the engine can actually see the clear text and then perform operations within, that data itself, so. That's actually interesting right we're keeping. The data. In. Ciphertext, for everyone, else around the engine, pretty. Neat concept, so. This is something that you guys definitely want to look into if. You're looking to enhance your security for your single servers remember. This talk is not just about performance, right this is about modernity, in scalability, hey, this is definitely a part of modernization, you're. Protecting your enterprise it should be your one of your utmost, important. Tasks. Now. This is a very favorite topic, persistent. Memory support. You. Guys don't know what persistent memory is back. In the 60s, we, had something called magnetic, core memory and, that's what it looked like, it. Was actually persistent, memory because. It was magnetic, right so unless you actually erased, the, magnet the. Memory, will remain in there whatever you wrote to that would, actually remain in there, for. A good while and. This, is actually the kind of memory that most computers used back then until. Intel, came up with the 1103. Chip which. Basically. Disrupted. The market and you. Know put these things to rest it. Was an integrated circuit. This, thing's lasted for a really long time. So. Persistent. Memory kind of started. With the beginnings of computing, and then, it was kind of like not that interesting, for a while because the cost was a little bit too high to make it work on on, ICS, etc, but. This is how my magnetic, sorry. Persistent, memory looks like today, in the year 2018, this. Is a an invading, - n hey. This is, basically. The same thing that, you use from your computers you know when you open up your server you have your CPU and you have your memory it's, literally the same size of a din okay. It's just not, be Ram this. Actually has the RAM but. In addition to the DRAM it, has a piece of Flash. Okay. In a super capacitor, so, in the event and that you lose power, the. Nvm, is smart and realizes, that it's lost its main power source and what, it does is that it uses the power and the super cap so. Actually flash that information, that's. Sitting on the DRAM, and flush. It down to the flash, okay. So once it's persisted, down the flash it's, safe so, now you can carry your system from, one prac to the next and then, you can plug it in and then, when it boots again, the. System realizes, that it's kind of an envy dim and, even in census that it's coming online so what it does is that it reads that information, that I was sitting on the flash and places. It back in the RAM so. It's basically persistent, storage and. It's not persistent, on the DRAM itself but, it's persisted on the flash within it so. That's very interesting and. Because now you can access you. Basically can treat this as a persistent, storage device and you. Know that as long as the super cap can actually, backup the data to flash you're. Not going to run into any corruption in the event of a power loss or, even if your server crashes and you need to restore it so that's, a very interesting proposition isn't, it you. Can run things, at speed of memory, you. Can actually treat your storage, as, fast, as you Street your rare so. The whole bond knowing main architecture, of computing, where you have your CPU. Your primary. Memory and your secondary, memory can, estates starts, to blur doesn't, it, and. So if you have to think about the implications that this can have on, software. That, is mainly i/o driven, like, seco server then. You really start thinking about the power that this can have and the, disruption in fact that this can cause on the industry, the.
Thing About these particular, language in - and devices. Which, are actually, everywhere. These. Days have been around for a really long time you can find a 16 gig one for about a thousand bucks, so. They're not that expensive and all things considered, the. Problem with them is the capacity. Okay. So sixteen, gigabytes is where you can find some vendors have a roadmap for denser. Invadin. - and solutions, but the reality, is that you can only get 16, gigs today so, if you do the math or, really, how many dims can you possibly fit on a server. Times. 16 gigabytes it's not a lot you. Can still run some you know pretty interesting, workloads in there if you say if you have a 300. Gig database and it's, super intensive, can. You run it in there yeah sure, no problem, but if your database this multi terabyte, that's. A different kind of proposition right you're. Doing, a whole different ballpark at that point. So. What changes, the game are these, guys. These. Are the new envy games from Intel they're called octane DC. These, and B dens can. Hold up to 512. Gigabytes, each. So. Think about that for a second. 512. Gigabytes even, if I place four of those on a box I'm talking two terabytes of memory right, there. Isn't. That nice. So. Now you, can fit a lot more databases, into those things the, difference, between these, and the ones that I just showed you is that, these actually use a different medium that. Is persisted. Itself, that. Medium, is called 3d at 3d crosspoint you've, made a third of it so, the. Evolution, of, SSDs. Was basically, NAND. Flash, so. Your traditional, flash and now, the new medium which is a 3d, crosspoint from intel ok now granted, ICS these are still ubiquitous, are everywhere they're, actually. Getting much cheaper, but, the new kid on the block the one that's fast and gives, you even better than the agency then SSDs it's, the 3d crosspoint minutes so the obtain drives are my dad Intel, cells but, now they have taken that. That. Medium. And actually. Made persistent, memory solutions, all of them so. That's really interesting, because now you can do capacity. At speed. So, really low latency, now. Granted, the latency on these devices is never going to be as fast as the latency on DRAM so, an envied end - and will always be faster than and invade. Them from Intel using, all obtain, dcmp dips, naturally. Because the medium takes a lot longer to write and brief to write to and read from okay. So DRAM is always going to be faster than that that's why DRM is so much more expensive by. The way that brings up another point, which. Is a, lot. Of you guys are still buying 128. Gigs of ram why. Are you doing that, why. Are you doing that you're hurting your business okay. Ram. Is actually, getting a lot cheaper than it used to be so. Buying a server with 256. Even half a terabyte of RAM should be the norm reading, you. Shouldn't be looking at a small amount of RAM it was smaller rounder Ram it's not gonna help you right. You want to run more things and there's. New things that we're working on the single server edges are actually going to take a little bit more memory, than than, they used to so, more RAM is going to benefit you to work out even more. So. Think about that as you make your purchasing decisions on on that as well. Anyway. Until. Hasn't, formally we GA, this product yet so it can say more, of a we if we and we we have seen on the on the lab on the testing that we have performed like cetera but let me just say these are very nice okay. So these devices are great. We're, going to see a couple of demos oh and. The. Sequel server side of things. Starting. With sequel Server 2016 we. Implemented. In me a feature that, took advantage, native. Advantage, of the, fact that you, could have an invading device on your, box and that was called tale of the log caching, what. That did is that basically, we. Basically shortened, the amount of operations, that we were have to go through the amount of code that you would have to execute in, order for you to to. Harden a log buffer you. Know when your secret server actually performs rights against your database it creates log records those, locked records are actually accumulated, in this, these. Structures. Well. Something we call the lock cache the, creation we correct these lock buffers and then we flush these log buffers to disk, okay. So. Now that there's no disk. We. Couldn't. You know remove, a lot of the code that we had you, know thinking that we were going to if Scott was gonna take us quite, a while potentially. A min and potentially. Several. Milliseconds, to complete single, i/o we. Could do those a lot faster so now we removed, a lot of overhead that we had to, deal with the fact that the, underneath.
Portion. Of the. Disk where your transaction, log will be. Will. Be on disk so we supported, in 2016, that's a lot L of the non caching you could create up to, 32, gigabytes I believe of a. Space. On an MDM for, you to place your active, portion a transaction log and it was only about the your active, portion of the transaction log and nothing else so the rest of your transaction log will, be this packed as always, so. We had a mechanism for, moving. You know those are that, will, active, pushing a transaction log into and move it around from. MV diem today's cakes etc, but. This was the beginning right we saw this as an opportunity to do something you. Know that, played nice with new advancements, in technology, as. We. See this, storage another. Storage for the overall OAM and hardware. Industry, shift. And embrace. And bidam's a lot more, then. We decide. To. Do something more interesting with with, persistent, memory so. Available, with sequel server CTP, 2.0, which, we released here at ignite, we. Now have what's called enlightenment, of database, files and this, essentially, means. The. Method. Of access. Of every. Sequel server file of cycles, or on Linux is, performed. Through a memory. Mapping, so. Rather than us going through the entire storage, tag and performance. Or expensive, syscalls to place those iOS in the kernel and the kernel which. Is memory map the. File and then. Once. We have it mapped in user space then. We just mem copy things, so. Rather than performing, those expensive, i/o operations, we, can now, that we're running on a payment device we. Can basically say, you. Know what I don't have to go through this entire set, of code here, because this stack is actually a lot of overhead I can, just map it and I can just copy right. From DRAM to the persistent, memory device. That's. Very interested, in using it, so. The. SIRT the instructions, are you have to follow in order to enable this on sequel several minutes are at that link that I just published at the bottom you can go visit at any time by. The way sequel PMON at microsoft.com, is also an email, address that hits all the PM's are involved in this so. You're welcome to send us feedback on this feature. This. Is very, nice right it opens the possibility of us running a much. Faster. Speeds. On sequel server on Linux now. There's. Also the possibility of you running just in block mode okay. So we have had this for a long time, let. Me just talk about block mode really quick before I move on I explain. To you guys how in. Lag networks if. You think about and Vadim's for a second they, could also be treated as discs and. Because they are persistent, devices, and so, the first thing that the operating systems did back then it's. Just to support MV dancers block devices just like we would treat a disk or an SSD it's.
Just Much faster right, you will go through the entire storage stack to, access those. So. You, could leverage. Your. Mvgame investments. As block, on any. Version of C Co server as. Long as a version of C Co server is supported. On the operating, system and the, operating system, supports optimal device. That's. Cool did you guys know that no. So that's really cool I give you guys some table that actually explains that yeah. With. All the different versions of signal server etc well let me explain let me go back and explain that light then line may feature a little bit this. Is what this is roughly, a high-level architecture of, how. You would actually access, an invading device on, Linux. You. Got your user space you, got your kernel you, got your different ways of accessing you're performing, tax against that device so direct access against the device. There. Are several. Ways of implementing, this at. The time that we started working on this, file. System tax was, the more appropriate, or the more developed solution. Device. Tax is something that's has come up recently, we. Are exploring that space we. May or may not pursue, it but. Right now the way the things work is basically, like, this. That's. The area that enlightenment covers daily. Basically, removes the, layer of EFS and the, page cache altogether but. We have to go through the file system in order for us to understand. That, IO needs to be hardened ok. Like, I said this possibility. In the future of a garage for device tax. But. The memory mapping happens, for the kernel level and then hold the application all the memcache actually, happen at the user space, so, there's a lot less context, switching from, the sequel server engine, into. Or. Other, CPU rings. Like. The kernel because. We don't have to perform those expensive, sitacles anymore so, the performance can. Be dramatically. Faster. So. These are this this is the expanded support for persistent memory devices you. Take. Into consideration that, impeding - and has been supported since Windows Server 2016. Remember. The versions of secure servers are supported, on the Windows Server 2016. So. Starting, with she co-starred with Windows, Server 2019. And, going. All the way back to sequel server 2014, we. Will support, high capacity, and Vadim's from Intel or, any other manufacturer, the Compson starts money creating. Putting. Those in the market as. Long as the operating system supports, it you can run sequel server 2014, and above on. Using. That storage as block mode. So. This is really interesting. If. You have a severe, bottleneck. On storage, that. No. Flash. Array can takes knowing, via me device configs, then. Persistent, memory ask block, in. Block mode can, potentially be a solution for you. This can have implications on, any. Sequel, server workload, that's our intensive, okay, including, some TV ok, so are these are very, nice solution to store user Explorer if, you have any questions, about what. The, features are supported, on. Which, versions secure server feel free to to. Come, to us this, table should explain it though sequel. Server on Linux obviously wasn't available in 2016, or 2014. So. We're not back porting in lineman to 2017. Yet there's no plans to do that before, you guys ask. So. That's part of the value prop for sequels over turn 19. Windows. Server 2016 though, you're, welcome, to use all the way back to 2014, and blocked out no problem, okay.
So. Let. Me demo persistent, memory in block mode and, so, thing. About my computer is that I've actually connected to three different VPNs, right now not, kidding you so. I'm hoping that no not, a single one of them disconnected to bear with me for a second. There we go. So. I. Know, this is kind of small I'll try to make it bigger promise. You. Saw her this. Is a sequel server 2019. CTP, 2.0, instance or, maybe, a build that we released that we came, up with a little bit earlier than we actually see to be to ctp 2.0 bill don't, worry about that. Do. I have two databases in here atpc eh like, database. Sitting. On nvme and another one sitting on a premium device, okay. So I'm. Going to show you guys the performance, of query, one from, the TP CH specification. It's. Just. Really. So. They're. With me. There. We go okay, so you click somewhere else and things just work so. I'm going to just set, statistics. Are you on and drop clean buffers. So. That way a free up. For a ripple, access. To every single page that I need to surprise my query yes. So. I ran this query. Which. Is the again q1, from the cpc-h specification. And. What. I can see is that this took roughly. 11. Seconds, okay. That. Is sitting on the TPC, HM DME database okay so this database is just sitting on a PCIe. Nvme SSD. It's, a pretty fancy solution, nowadays actually pretty good it's, pretty good performance for this query but. Look what happens when we actually run this against a premium to be CHP mm database. So. Again I'm gonna say drop, team buffers to make, sure that I'm not cheating and that every single i/o is a physical i/o I'm. Just gonna run that query again. There. We go 5 seconds plus 5 seconds flat so. Actually, closer to 6 seconds but you get my drift so. The one took about 11 seconds the other one took 6. 6. Seconds, that. Is a nice improvement in performance isn't it very. Nice so. Let me tell you guys this. Has the potential to change a lot of things for you okay. So a lot of you're very. Heavily. Restricted. Workloads. But you cannot make any faster this can, potentially make things a lot better for you, ok. So. That. Was the demo for block mode. And. So what we have next. See. I, believe we have another demo yes we do so, we have the demo for persistent, memory using, alignment all. Unfortunately. The. Boxes, that are refusing for this were not cooperating, so, I cannot. Demo live however. I did take snapshots. Screenshots. Of everything. Before. I make as, I was doing my initial testing so, I am going to show you those screenshots and I apologize that I can not do a live demo for this but. This. Is the first this. Is the first screenshot. This. Is actually executed. Against. 8 nbme. Device. Using, Allah without, using enlightenment okay, so, this is just running a TPCC, like, workload, okay, against an nvme device, again, without enlightenment. If, you look right. Here. Right. There the, transactions, per second, we're about thirty four thousand okay. So again that's the performance on PCIe, nvme, without. In leiden which. Is what most of you guys would be running today okay. So if we were running sequel server 2017, today this is likely the scenario, that you would be in, and. By, the way there's zero tuning involved here okay I was purportedly. Lazy around, building these demos so, I wouldn't use any fancy trace flags or anything like that this, is out-of-the-box, stuff, okay. So the next screenshot. That's. Sitting, on P man so let's look at this number again. 70,000. Transactions. Per per second, all, standing on P man so. You can tell clearly, how the much, lower latency, of a premium device can. Have a very, nice in impact. On your workload it can dramatically. Make it better okay. So enlightenment, is one of the things that you're definitely want to look into. That's. One that's one sequel server or Linux by the way that particular, optimization doesn't.
Make A lot of sense on Windows ok. We are definitely exploring, it or whether we are going to we're going to do it or not on Windows but the, gains that we may get may, not be worth it okay. So, we're remember. This is CTP 2.0, okay we have plenty of time until GA to. Figure out how things are going to look like in the end but, the reality is that this isn't this is this is in product right now so, if you're lucky. To, get your hands on a premium device especially. One of the one. Of the fancy ones from Intel then, you can test, this with your data bases today again. And if you do have mvgame, so I want to leverage champions and wanna test your worker and I need any help from that please contact me again seek OPM at microsoft.com. So. We have a few, more things in store here. They, degeneration, our shared storage, you. Know it's. A lot of you guys that are kind of like allergic to sands I can. Actually, understand why because, I worked for a sander and grow until not too long ago I can. Tell you, sands. Today are dramatically. Better than they used to be and, I mean dramatically. Okay. The. Performance, and that's nice set of features that they can offer these days are really. Really great so, if you're looking for just shared storage solution, because you want to you, know continue to follow our model or maybe you have your investments, already on your, fibre channel infrastructure, etc or, your. You have your widely deployed ice, crossing infrastructure, or maybe you're running VMware right and you're it's, just following the shared, storage solutions, or, that model for for, vmware or even for hyper-v right. You. Can have some great solutions here they're. Typically named based okay so typically named flash your traditional, SSDs, some. Of the newer ones though they can actually the, leverage, Intel, 3d crosspoint so. The latency between the. Device all of, those, storage. Arrays can definitely vary, they. Typically, have very good ladies very good latency typically under one millisecond, and on.
The One millisecond, is the only when you want to be for. Sequel server workloads, okay, this is 2018, if, you guys come to me and you tell me that you're it's taking you ten milliseconds, to perform a transaction log right I'm. Gonna refer you to the first part of the presentation where. We talked about modernization, and how that's impacting your business right. Because how many means how many transaction, log writes will, then take you to get to a second, and then how much to a minute etc etc you, can do the math and how this is already hurting just doing the math in your head right. These. Devices, these storage arrays are nowadays so you can have excellent, bandwidth, excellent, throughput. I have, seen anything, from a. Gigabyte. Per second on the lower end of the spectrum some. SKUs all the, way to ten eleven even twelve gigabytes, per second at report on some devices and. So if you think about that for a second if you're performing very, expensive, the data warehousing scan, and trying a ghostie, PCH. That I just showed you, those. Can, definitely benefit from having. That massive bandwidth available, too if. You have a lot, set of analytic workloads that can definitely drive better, performance, from or extract, better performance, from from that storage array you're, gonna benefit from this ok, now. Granted, the, performance, you can get from those. Nvm, solutions that I mentioned before is equally. Or even, faster right, because. You're there's basically no network, no sand in between you, and the storage device itself. This. Is a like a different, paradigm but. It's equally valid today ok. And a lot of you guys are, kind. Of shifting away from it it's fine I fully. Embrace hyper-converged, I've only embrace pmm as well but, a lot of you guys knew you also need to look into if you want to remain on the shared storage model you, want to look into the new devices, typically. These half-block services, already built-in right. So they can give you snapshots, and they can give you replication, they, can give you things that make. Things a lot easier and that. Makes a lot more sense to do on the storage side and. They do on the database side it depends, on this it, depends on what you're dealing with right. So. The offerings come from a variety of vendors and this is just a list, of the ones that are are the on, the, Gardner. Upper. Right. Hand corner. Blah, blah this. That's just a list of those vendors right there's a lot more I. Can. Tell you I've worked with many of these they're all great they all they will do fantastic. This. Is the key, because. Of the services, that they offer most of them has have, snapshots, most, of them have replication, they. Are actually, excellent, for, performing, tests and validation, of workloads and if, you remember what this session is all about which, is about modernization. Of single server one. Thing that you can do is leverage these features, for, testing, and validating, your sequel server workloads, so. The way that you would go about this is take a snapshot of your production, system, either. Attached'. That, sac take that snapshot and attach it as a copied volume or mirrored volume or whatever it's called to the, new system that. You're building. Test. Your workload there, maybe. Within the same array or take, that take one of those copies of the volumes and then replicate them to another array and run. That work out there, that. Sounds, a little bit faster, than taking backups, and restoring things doesn't it typically. It is. Especially. If you have one of the very fancy, arrays or do metadata, only copies, I will, you just press a button and you automatically, get a copy of the data that. Is. Significantly. Better than what you could do with your desk today so you did your you're, directly, attached storage they, I mean they run fast they have great Nancy but they don't have the, additional. Services, that are very beneficial when. It comes when it comes to testing and validation so this is why I like them so much because. If you're running one of these arrays you. Can just clone, your database and start running tests on it yes. If you're running within the same storage array then you could potentially bring it down by just running the wrong query against. The test, database you're performing so, you need to be careful quality, of service etc, etc you, could. Use resource governor to limit the amount of i/o so you're issuing, against that array etc, etc but. These are potentially. Very good solutions for you to look for you guys to look into ok. I was. Going to do a demo of this but unfortunately due to technical glitches we couldn't, get it to properly. Function, but, anyway this. The, idea of behind sure storage so you can even achieve better. Performance then, then. You can on my own on some of the MV and DME solutions so, I'm gonna demo that for you guys this, actually comes from from.
A Friend some peer storage here we. See. If I can actually get to that box this. Is a Cisco CC. 480 box so. Remember, that TP, CH 5 database, that I told you about where I was working furnished before. How. About we, drop, clean buffers in this box as well and, then. We run the q1 again. So. This is storage array with all those services and I mentioned before look at that. Sorry. Oh sorry. What happened here, haven't. Switched. Sorry. Sorry. About that. Thank. You all. Right so uh opening. Buffers. And. Now we run the query. So. Here you go three, seconds that's. Not bad huh so. Of. This obviously depends right if you have a bunch of nvme drives a PCIe, nvme drives, on your box you're definitely gonna achieve, tons. And tons of terms of bandwidth I'd much much later bandwidth. If you have one of the super fancy storage arrays you can achieve this, bandwidth, if, you have tons of key mm in, your box you can also achieve that kind of bandwidth okay, this is so relative I'm not saying one is better than the other what, I'm saying is you, can definitely impact, your workloads. On a very positive way by, leveraging by. Leveraging today's. Storage. Technologies. Oh. Guys. Hold on I'm trying to get this under control there, we go so, latest. Generation of hyper-converged, there's. A lot of hype a pervert solutions, out there okay. A lot of them quite. Honestly. We. Just came up with the best record for hyper-converged performance, Intel. And Microsoft just ate okay. This. Is something that we just announced literally, on Monday we, broke the record on performance, on hyper-converged. Okay. This. Is the setup okay. It's, 12, nodes from, Intel. 384. Gigabytes, of DRAM. It's. Got 2, - socket. There, are 2 socket boxes with. Twenty eight cores. The. Future, Xeon scalable processors. 1.5. Terabytes, of persistent, memory use, test cache for. Each box. Okay. And then 32, terabytes, of latest, generation of vm e, as. Capacity, and then. Fully, interconnected a twenty. Five gigabit four 1250. Actually 50 gigabits, per second across, all of the nodes running. An insider preview of secrets of Windows Server to nineteen and, then that's. The bill number right there so using hyper-v and storage, space is direct. You. Guys want to see how. A mini I opted we got on this thing. That's. The number. Those.
Are Four K random eye ups but you get to. Get the drift. This. Is insane. Performance. Okay, this is. Amazing. If, you guys do the math based, on 4k ops this. Is basically 52, gigabytes, per second, of throughput, at. Microsecond. Latency. So. The performance that you get from these hyper convert solutions, where, you can now virtualize, your hyper-v, infrastructure, for example. And. Run all your sequel server so right on top of hyper-v, that's. The kind of performance you can get from your storage. Amazing. So. This is a great. Great, break new set of a new. Set of technologies that, just came out. Anyway. The. Session is actually not going to be 75 minutes so you can clearly tell these. Are the related sessions that I will want everyone to attend from here. Upgrading. With predictable performance yes, it can be done the, session is actually very important I'll tell you why. My. Colleague. Pedro Lopez on the sequel server Tiger team is going to talk to you guys about database compatibility, which. Is a very important, feature for modernization. The. Reason why database. Come in DB compat all, it is very important, is that we're actually changing or, we have already changed the, verbiage, now and then in the way that was perceived by folks out there we, are no longer using it as, a patch. Whenever you see or not more no longer telling you use. DB compat for a little bit until you get until you can migrate we're, now telling you we're fully going to support DB comp ad as, long as it's supported by the underlying store, database, engine we. Will support, you running on that DB compact level for a really long time doesn't matter until when until you're ready to switch to the next day become pad so, rather than you certifying, for a version of sequel server you, certify, for a particularly be compatible. That's. Very interesting because now you can take your database that's, running on a given the, DB comp at level and just place. It on another newer. Version, of cycle server that, can still give you better performance and, leverage all the features as, long as our DB compat mode is. Available, in that version of sequel that is. One way through which you can actually achieve a much. Quicker migration. Off of secure Server 2008, r2 okay. So keep that in mind you definitely, want to go to that session, okay. The, other session that's really good the one that I mentioned the beginning modernizing. Seco server the right way this, one is going to give you all the tools that you need to leverage and, tell you about the process that you need to follow that, playbook that you need to build in order to get you from vibrational secure server to the next, okay. Then. There's practical. Guidance to make yours equal to your tier 1 Seco server role. That. One is all about making sequel server faster, so, it's not just about infrastructure, which is really what we talked about here it's. Also about, following. All the best practices and doing all the right things to get you to the best place that you can be intersperse. Then. There's also the session that they did yesterday on Windows Server 2019, the deep dive where, they show all the new features, that sequel, server actually, leverages, because if you run fast on Windows sequel. Server will run faster on top of that ok. You. Definitely want to check that out and then, the one that relates to the security and clips if you're interested, in confidential. Computing that's, the one that you want to go to BRK three one five seven computers local PD which it goes over anklets, okay. With. That said please I appreciate, it I would appreciate it if you guys would evaluate this session if, you have any questions, please feel free to come up and in to, the microphones and we'll do a little QA session, thank. You so much.
2018-10-27 14:36
