SNIA Experts on Data
Listen to interviews with SNIA experts on data who cover a wide range of topics on both established and emerging technologies. SNIA is an industry organization that develops global standards and delivers vendor-neutral education on technologies related to data.
SNIA Experts on Data
Breaking Barriers in Data Movement with SNIA SDXI
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Are you ready to gain a fresh perspective on the future of data movement? This episode features an insightful conversation with Shyam Iyer, chair of the SNIA Smart Data Accelerator Interface (SDXI) Technical Work Group. Get ready to explore why a standardized data accelerator interface is crucial in our increasingly complex data landscape and how SDXI is paving the way with an extensible, forward-compatible, and interconnect technology-independent solution. From the intricate challenges of software mem copies to the potential of emerging interconnects like CXL, Shyam's expertise provides an enlightening look at the pivotal technological shifts shaping data movement's future.
As we progress through the episode, Shyam offers an exclusive insider's view on the formation and growth of the SDXI Technical Work Group. With a community that now includes nearly 30 member companies, this group has made significant progress in standardizing data movement and acceleration. You won't want to miss the updates on the progress of the 1.0 standard and the future plans of the group. This episode offers a unique opportunity for chip vendors, storage vendors, and anyone interested in the field to learn about the possibilities of getting involved in the group's ongoing efforts. Tune in and discover how your organization can leverage these exciting developments in the data movement space.
For more information on SDXI, check out the SNIA Educational Library.
SNIA is an industry organization that develops global standards and delivers vendor-neutral education on technologies related to data. In these interviews, SNIA experts on data cover a wide range of topics on both established and emerging technologies.
About SNIA:
Welcome to the SNEA on Data Podcast. Each episode highlights key technologies related to handling and optimizing data. You are listening to SNEA on Data Podcast setting the standard for memory-to-memory data movement. In this episode, Philip Alsop, editor of Digitalization World, interviews Sham Iyer, chair of the SNEA Smart Data Accelerator Interface, or SDXI, Technical Work Group.
Philip AlsopOkay. So SNEA, you have a smart data accelerator interface or SDXI Technical Working Group, and I believe it was set up to address the current issues around memory-to-memory data movement. So I guess just a little bit of background on you know flesh out what that actually means.
Shyam IyerYeah, if you look at software stack today, our current data movement standard is a software mem copy. You know this is because a software mem copy enjoys a very stable instruction set architecture with CPUs, but software mem copies take away from application performance. So, if you look very closely, multiple layers of software perform mem copies just for doing things like context isolation. And these layers exist for various use cases like virtualization, do having containers or just for creating separation between user and kernel software and most devices today, even if they're IO devices, they perform DMA for things like network or storage IO. They don't perform memory-to-memory DMA. And while memory-to-memory DMA acceleration has existed, it has been based on proprietary vendor implementations with no standardization between vendors.
Shyam IyerSo something else we want to consider here is like we need an accelerator interface that is one form factor independent or implementation agnostic. So, just like traditional compute devices like CPU, memory, data movement and acceleration is equally important to other kinds of compute accelerators like GPUs, fpgas, even computational drives and other terms that people like to throw around like GPUs, xpus and IPUs. And memory is no longer limited to traditional and proprietary DRAM-based interfaces. It is now possible to access memory via interconnects like CXL, which is another standard that is brewing in the industry. So you can also have memory with different kinds of properties, like persistence, or it could exist in device memory as MMIO, so in whole memory. Tiering with different classes of memory and interconnect type is needed. These different usage models and environments need a holistic, standards-based approach to future proof applications that will take advantage of this exhibition.
Philip AlsopOkay, so you've explained quite a complex landscape, but your simple and inverted commerce objective is to create a standard data mover interface. So I guess what does that involve? Bearing in mind what you already outlined is quite a lot of them, different things going on.
Shyam IyerYeah, absolutely so. Smart data acceleration interface, or STXI, is creating the industry's first standard for a memory to memory data movement and acceleration interface that is extensible, forward compatible and independent of IO interconnect technology. That would be your short tagline.
Philip AlsopOkay, yeah, and in terms of the actual practicalities of that, I guess, of what you described earlier, you're going to try and support a number of what you would deem sort of desired data, movement characteristics or attributes. So you're going to try and define what must have and, I guess, as you say, open to all, before anyone sort of adds any other layers. Is that correct? That's right.
Shyam IyerSo, among different things, stxi is envisioned and designed to support data movement between different address spaces. So this is address spaces, which includes user address spaces or address spaces in different virtual machines. And, to be fair, data movement should occur without mediation by privileged software. So if an application wants to do the data movement, it shouldn't have to transition into kernel software just to be able to move the data. Of course, once the connection has been blessed or it has been established, you can make that happen, but it kind of has that separation of user and kernel that is needed, but not when the actual data movement is happening. You need to allow abstraction or virtualization by privileged software.
Shyam IyerAnd something that is key that we are designing into is the capability to sort of co-ass or suspend or resume an architectural state of this data mover. And why is that important? If you ask me that, it's because we want to enable live workload migration, or some folks call it machine migration, between different kind of servers. So the ability to be able to stop, suspend and resume is important for this and compatibility is very important. So that's why we're trying to design here as attributes is to enable forward and backward compatibility so that now you can build interoperable software and hardware. Something else that I want to highlight here is, when we build a standard like this, we can incorporate additional offloads that leverage the same architectural interface, and overall the DMA model is meant to be concurrent, which means multiple DMA accesses can happen at the same time and one is start getting the other.
Philip AlsopOkay, and in terms of the inspiration or the seed for this, did you as an association identify the problem and then sort of start up the technical working group, get people joined, or did various vendors come to you and say, look, this is a bit of a Wild West scenario and we need some kind of formalization around a standard to help us all move forward? Where did the impetus come from?
Shyam IyerYeah, absolutely so. This trick was. The technical working group that we call was formed in July of 2020. And it started as a specification that was contributed by Dell, vmware and AMD and since it got formed inside SNEA, the technical working group has grown steadily to almost 28, 30 member companies and almost 80 plus members being part of this technical working group.
Shyam IyerMany members have contributed in various ways towards improving the specification since its initial contribution by the founding members and kind of their contributions have resulted in improvements in many areas, like defining an architectural behavior of the start-stop operations, architectural context states and architectural states of the function and SCSI function implementation. Improving things like how does cache placement work or cache injection needs to work, having a different memory of different groups of SCSI function implementations, error reporting I mean I can list a whole bunch of things. But something that very critical that we did was the language of the specification is very interconnect agnostic, so you could take any interconnect and still be able to implement SCSI with that, say, psei or CXL or maybe I don't know someone else has a proprietary interconnect. They can still implement a standard SCSI implementation of this.
Philip AlsopOkay, so the work has been going on on the specification for about, I guess, a couple of years, so what else is there needed to do, because I presume you're working towards creating and ratifying some kind of a standard. So what's the roadmap for that look like?
Shyam IyerAbsolutely so. Since we started in July 2020, last year, in September 21, we released our first public draft that went through about eight months of public review. We just released a candidate for 1.0 in September this year, which is expected to become the 1.0 standard before the end of the year. So, along with the draft specification, we also included header files that are auto-generated from the specification so that implementations can start using these standard structures and begin their implementation. In terms of ratifying an initial standard, we're currently in that 60 day membership and public review period, which means once that finishes before the end of the year, we would incorporate any review feedback that comes in and that will become the 1.0 standard and 1.0 standard to be precise, and after 1.0, we kind of plan to work on future versions of the specification to incorporate additional features.
Philip AlsopOkay, so there's a lot of right about that In terms of opportunities. I mean, if there are people, it sounds like you've got a pretty big sort of community already. But I mean, are there opportunities for other storage vendors, I don't know, end users, whatever to get involved, or is it pretty much you've got everyone you need or where you at?
Shyam IyerNo, it takes a village to build right Absolutely. As we bring activities around version 1.0 to a conclusion, we are also ramping activities in compliant software implementations. So common software stacks and structures across all OS implementations are being discussed within the Tweet and being proposed. So participating in these discussions or code contributions can ensure you're ready to take advantage of them in your applications. So already some chip vendors are participating actively in the Tweet.
Shyam IyerSo if you are a chip vendor looking to develop STXI compliant implementation, you may want to join the Tweet to ensure you are in compliance. Reading a spec and implementing is certainly possible, but you might want to come in and make sure your interpretation is good. You may also want to make sure that the next revision of the standard, which will include more features or that we are going to work on, meets the technical requirements of your company. And, for any reason, if you cannot join the Tweet, you may also provide feedback using SNES feedback portal. But I would say, if you're interested and actually this makes a lot of sense, joining the Tweet is the best possible way to start getting into it. It's a fun group. It's a lot of members in it, a lot of very senior people with a lot of expertise in it, so I would definitely ask people to join.
Philip AlsopOkay, and in terms of so on the developer side, I'm just thinking for the end users. So if I'm a storage user in an enterprise, whatever shape or size, what will this standard meet? I mean, will I notice anything? Or is it all under the covers, so life will just go on merrily as it always has? Or will there be sort of certain benefits or characteristics of my storage that I'll notice? Go, hey, that works better or faster, whatever it is, yeah, how do the end user find out and notice what's going on? I guess Well.
Shyam IyerThank you, this is a very important question. This work is a framework to unlock innovation from various acceleration vendors and make it available to sort of various software stacks. By leveraging a memory to memory data movement standard, wenders can innovate around the specification and offer many different types of accelerators directly benefiting the applications Because, remember, we are not making the applications go through a layers of stack just to be able to use the accelerator interface. This includes storage applications and non-storage applications, so it includes applications that run bare metal or they run virtualized. They could be used by column mode software or user level software. So many network and storage stacks already perform multiple buffer copies. Stxi assisted network and storage implementations can eliminate or accelerate many such redundant and unoptimized software memory buffer copies.
Shyam IyerThere are also some use cases we are envisioning for virtualized environments which are very specific to them, like, for example, the spec is fulfilling the need for when virtualized servers want pre-filled large memory buffers, filling them with zero is a necessary and important step for security. You don't want your data getting leaked when you're spinning up and spinning down VMs, and STXI provides an interface that accelerates even these kind of memory operations. That's what I meant, that different kinds of applications may use this Also. Many operations that involve inline data transformation of the data as it moves from, say, a source location to a destination location can get accelerated. So, for example, these data transformations could be compression, encryption or other kinds of offloads that involve data movement from one location in memory to another location in memory. Let me also kind of take you to a part exercise.
Shyam IyerIf you have a system that has different tiers of memory, like DRAM memory or memory which is persistent, a memory that is behind, like a device, or if the memory is far across a memory fabric, like CXL or a link, then STXI is being designed to target all such memory and memory address spaces. To accelerate data movement and, as various software stacks and storage stacks begin to embrace CXL, stxi will become an important standard that incrementers need to consider.
Philip AlsopOkay and just quickly on the time, you've told us obviously that you're hoping the standard will be oneo will be ratified by, I think, by the end of the year. But in terms of the vendors, bring out solutions I mean, I imagine some of them are already. Obviously, if they're working on the standard, they pretty much know what it's gonna be. Hopefully, so they're working. I mean, will it be soon, in 2023, that STXI compliance solutions, or will people have to wait, maybe to the second half of the next year? What's that sort of timeframe?
Shyam IyerYeah, no, it's a good question. I think a lot of the I guess the vendors are thinking about the implementations and some of them are further ahead and some are still working towards it, but it would be a safe bet to think that 2023 would be the year you might start seeing implementations.
Philip AlsopOkay, it was being lovely to chat to you and a great new technology solution, whatever. So be interested to see how it develops. We can maybe catch up again, but for now, Tram, thank you very much for your time. Thank you.
Shyam IyerThank you, phillip, it was great chatting with you and thank you for spreading the word. Thanks again.
Kristin HauserThank you for listening. For additional information on the material presented in this podcast, be sure and check out our educational library at sniaorg wwwsniaorg.