High-Performance Memory for the World Computer

Transcript

[Music] Okay, thank you so much. Uh, I'm Muriel and yes, I have onampus parking, right? So, um, uh, yes. So, basically, um, I'm on leave from MIT here. Uh, I lead the network coding. I've been leading the network coding and reliable communications group in the research lab for electronics. I'm a full professor in the electrical engineering computer science department. uh and hold the NC chair for software science and engineering. It's a um school of engineering wide chair. Um I'm going to talk about optimum and first always like to um recognize our backers led by 1K axe. Uh and then let me go straight into

what do I mean by memory. So what we mean by memory is basically if you really are going to think and you know this is uh this is u basically an ETH event uh that web 3 is the world's computer decentralized and so on and then you think okay you know when I've put together a computer what things did I put on the computer right so other than if you're my son and then it has these lights so that it like looks like it goes fast in which case you know you're on your phone. Um you basically uh have the processing part and you have the memory part. We'll talk a bit more about that later. And what we do is the memory part. And memory means anything

where it's the data itself, not the compute part, but the data itself. Whether it be basically the ephemeral memory, the data propagation. Uh that's optimum P2P. That's our first product. And I'll be showing you some of the results. uh from our hoodie test net there. Whether it be the readonly memory such as the ledger uh which is basically the historical state. It's like a readon. You do you add to it but you don't you don't modify it. um or whether it be the actual state which corresponds effectively to the random access memory the read write and doing that in a

decentralized fashion is doable um but is not easy unless you have the right tech and we'll go a little bit into what that right tech is. So this is what we mean. We're doing the memory. And again, when you think about putting together a computer, it's not the same folks who are writing your operating system, who are making your CPUs or your GPUs or whatever you're using, and who are making the bus and the memory. It's not the same people because it requires specialized expert knowledge and specialization in what you're doing. Okay. So let's go back uh to this idea of the computer and let's put a little bit more uh thought and uh

structure around it. So on the left hand side I have the web three stack. Now you know different people have somewhat different layers in there but I think pretty much everybody would agree that there are at least these layers right and we can talk afterwards if you want to stick other layers. I'm I'm happy to do that. And what I'm doing here is separating the green layers which are basically really around the computation from those purple layers which are really about the data. Okay. And the networking is about the data and it's about memory. It's treasure memory.

Okay. Um and if we think okay how does this actually relate to a computer? Well, we can go back to what's called the vonoyoman or the the Princeton model. And you know, it came out basically right after World War II when people had done things like the enigma machine and so on. They were really starting to think about what is a computer. And Vonoman said, well, basically this is the pieces that a computer should have. It has an input. It better have some way for me to tell it what to do. It has an output. It has to tell me what what it did. Uh and in the middle there's the compute part.

Let's think of that as the central processing unit which has the control say the EVM. It has the actual hardware the arithmetic or logic unit sometimes called ALU and then it has what we would recognize as the bus to transport stuff and the memory unit and that's it. That's a computer. And this has stood the test of time, right? I will argue that any, you know, whether it be your phone, whether it be anything that you have is actually abiding by that general structure. And again, there's variations on this, but by and large has stood the test of time. Okay, let's look at the top. Let's look at the green part. uh you have the OS maybe an EVM maybe an

SVM maybe whatever you have and then you have applications running on those OSS that part if you look at the web three part and you know let's say the classical single computer part actually looks fairly homorphic okay yes there's all kinds of subtleties and so on but it looks pretty recognizable it doesn't look weird let's look at the bottom part the purple part that is not looking recognizable, right? Uh you have very ad hoc solutions for readwrites like memp pools, right? Does your computer have a does anybody here have a computer that has a mempool in it?

I hope not. Otherwise, let's talk. Um you know, you have these giant full nodes for archival information. Uh instead of actually reading the data, those nodes are so big that you have to check that the data is available, which is DA. your computer actually has something like that behind the scenes but you know you usually don't don't display it. Um and then the part of the bus particularly if we're talking say uh in the Ethereum context and this has been of course in the news a lot with lip P2P and so on um is actually gossip which is definitely not what's happening inside your computer. Okay. So with this framework and this sort of you know

intellectual discipline of you know we should look at what a computer does it doesn't mean you will never do something different but it means if you're doing something different you should have a very good explicit thought out explainable reason why you're doing something different okay uh we basically look at the transition memory the DRM and the DROM which is basically the reader only memory and what I want to look at mostly today is the optimum P2P which is basically the ephemeral memory the bus because that's our first product why does it make sense that it's first product it makes sense it's first

product because by the way you know you need to access the memory so if accessing the memory if your IO to the memory is slow then who cares how good your memory is that becomes your bottleneck So it makes complete sense. Okay. So you can say okay this is very nice. This is conceptual but you know how are you actually going to do it? So recall that uh I've been leading the network coding and reliable communications group here at MIT and you know I am by trade by uh um by training an information theorist.

So, we're going to use information theory and in particular, we're going to use coding and in particular, we're going to use a type of coding that I've co-invented here in my lab at MIT called RLNC. Now, this does not just, you know, roll off the tongue, but it's actually very simple. The idea behind memory is that you're representing data, right? You have a data representation which can be recoverable.

How do you make data representation be robust and efficient? You code which means that you represent the data not necessarily as the original data but you represent it as equations and those equations can be invertible have to be invertible so that you can recover the data. And the way that we're doing this set of equation constructions is that we're using a linear approach which is random. And the randomness critically has to work over a particular type of uh finite fields that are sufficiently large that

have a large enough cardality that this all works out. And you know we have tons of papers. This work has won tons and tons of awards all over the world. So, it's obtuse, but actually it's really intuitive. You're just going to make these equations. And these equations are infinitely composable. So, because it's random, it's inherently decentralized. And because it's random, it's infinitely composable. I like to joke that when I started working on this, my colleagues would do wellness checks on me because they were like, why on earth are you working on this decentralized stuff, Muriel? Like nobody gives or you know

your your own your own word there, right? Um and and basically I was like but because it actually is amazing that you can do things in a decentralized way here and this is why we're called Optimum in an optimum way even though it's decentralized. And you know my mind was so blown I couldn't stop working on it. Um and this is what it does. So this is uh for optimum P2P. So the uh our protocols called mum P2P as which you see up there. This is our live hoodie test net.

This is just a snapshot. Up top you see basically how long we're taking uh in this particular time for the latency you see it's around 150 milliseconds and on the bottom what you're seeing on the y ais on the ordinate is the number of times the how many x were better than the best node with current lip p2p gossip sub the current approach taken by the we're six times better. Sometimes we're more.

I don't think we've ever seen us go under four times. Okay, that's what it does for you. So, a little bit of math goes a very very very long way. So um we've gone to validators with this because validators like things to go fast me increasing their APY. This is what they're trying to do. Um you know the calls are always visibly short because we have like some really long presentation and about you know three slides in they're like I get it just give it to me. Thank you Muriel. I

don't want to hear your presentation. Um, and basically this is how we've been signing up um, in our private test net. The the hoodie, you saw the the graphana um, dashboard there because it goes fast. Um, so basically this is where we are right now in terms of a node operator partners and they've been fantastic. They give us I have to say that's something I've really really appreciate in this community is the level of openness, intellectual um curiosity and collaborative collegial approach. Um and uh our val our validator partners have definitely

embodied those qualities. Um so who benefits? Well, obviously the validators as we just said L1's and L2s we have also signed a few of those and in the end of course end user DAPs etc because you know things are going faster gas fees are lower you know without using you know using math not using extra hardware so with that um I'd like to make a plug for people who are interested in coming on day here at MIT to our um hackathon where you will actually hack uh what we call a flex node which is a

node that's actually propagating those equations and you can start seeing some of these benefits yourself. So, uh, if you have no plans for Sunday, or even if you have plans for Sunday, uh, I think we have a couple spots left, uh, catch somebody with an optimum hat or, uh, an optimum shirt. There you go. There and there. Uh, and, uh, the jacket and, you know, see about about registering. Um, and, you know, you really have to do it to to basically see what it does and see what it can uh see what it can do for you. Um, and with that I shall stop and leave it for questions.

>> Yes, thank you. That's great. A quick one. Are there any other bottlenecks currently that would stand in the way of the of achieving the 6x? Like what else is out there to block that at the moment? >> Yeah. Uh so what's out there to block that at the moment is basically the current way in which uh gossip systems operate and I can give you a very high level uh you know view of why this is. So what gossip does it's decentralized transmission of messages. Uh so suppose that I'm basically have some messages and I send them here to the first row and then they start sending them etc etc. Okay. Um,

if a node, say gentleman in the second row is a node and he doesn't have enough messages coming in, he's not going to be able to send it to the third row, right? Or it's going to take a while. On the other hand, if a node has too many connections, then what's going to happen is you're going to get a lot of duplicate and that's going to clog up your network and create congestion. So uh in lip P2P there's a really uh subtle difference uh between those two effects and you're basically trying to you know enhance one without going you know they enhance the connectivity without going too much into the congestion and the problems that

that raises. Okay. What we do is we send equations. Every equation that you get pretty much is going to be useful so you don't get repeats. That's it. You don't get repeats. So you're like, "Yeah, absolutely. Just give me your you give me your equations. The chances that I get an equation that's not useful is almost nil." Right? This gets back a little bit into the math. So it's really around the design, the protocol, and the math. Um, oddly enough, this optimum approach and actually my former student Bernhon Hopler who was a student here was he was my PhD student then he went uh to uh CMU to to to be a professor in computer science proved what's called uh basically optimum

he he showed that this approach RLNC uh uh achieves um basically the same behavior it's optimum pipelining as though you had an omnisient centralized controller that's seeing everything and is actually telling everybody when to go. Okay, so this is the thing. It's not that this that going back to what I was saying, well, I couldn't stop working on this. It's not that you go, okay, well, I decentralized, so I'm going to have to put up with all of this, you know, inefficiency. The decentralized approach without state information does as well as a centralized approach with full state information, which is

mind-blowing, right? It is mind-blowing. I don't know. That was a really long answer to your question, but I hope it was at least an answer. Yeah, absolutely. Yes. Oh, somebody has a mic. Whoever has a mic. Sorry. I think >> I do. >> Oh, sorry. Yes. Um so you you had said that uh earlier in the presentation that there was uh supposedly like a minimum size uh for these equations to be able to to encode information from >> minimum size of the field representation. >> Yes. So so in terms of uh you know some sort of bite or binary representation what what kind of like >> oh yeah it it's can be very small that's a good question. So basically uh you know you don't have to go to say

primes if that's the way your your mind is going um you know which is a natural way to go if you think of finite fields uh an extension field even just of a binary will work right so rough rule of thumb is the cardality of the field take the inverse of that that's the probability that one of your equations ends up being to our discussion before a repeat which is not the end of the world it's a little bit right so if you have say, you know, eight bits, right? The probability that you get a repeat equation is, you know, one over 2 to the 8, which is very small, right? Compared to the probability before where you had tons and tons of it. Does that answer your question?

>> Thanks. Yes. >> Uh, great presentation. It's good to actually see someone working on uh this memory bottleneck that Ethereum always had. uh after the upgrade I think in 2023 the Shanghai Capella upgrade you saw many um people having issues like liquid staking with all the transaction being stuck on the queue. I was one of them had my transaction stuck for a few days. Yes. So um it's good to actually see you have something on a test net. >> When is the manet?

>> A very good question. Um, so you know, I'm I'm hoping at the end of the year Q1, I think Q1 is is fairly is is fairly um conservative. Yeah. Thank you. >> Let's give a round of applause. >> Thank you.