About the talk
These days, taking the revolutionary approach of using clean-slate learning-based designs to completely replace the classic congestion control schemes for the Internet is gaining popularity. However, we argue that current clean-slate learning-based techniques bring practical issues and concerns such as overhead, convergence issues, and low performance over unseen network conditions to the table. To address these issues, we take a pragmatic and evolutionary approach combining classic congestion control strategies and advanced modern deep reinforcement learning (DRL) techniques and introduce a novel hybrid congestion control for the Internet named Orca1. Through extensive experiments done over global testbeds on the Internet and various locally emulated network conditions, we demonstrate that Orca is adaptive and achieves consistent high performance in different network conditions, while it can significantly alleviate the issues and problems of its clean-slate learning-based counterparts.
01:30 The step scenario
02:55 Current learning-based designs' motto
04:59 How Orca works
06:57 Orca's performance
Soheil Abbasloo received the Ph.D. degree in electrical engineering from the New York University in 2020, where his research was centered around congestion control, ultra-low latency networks, machine learning-assisted network design, scheduling and resource management, datacenter networks, and cellular networks. Before attending the New York University, he worked as a Software Engineer and a Network Protocol Designer with the Network Industry for more than four years.View the profile
Lowered one. This is so I'm going to talk about are there. When you're going to talk of war and mine, then be supported and classic uristix designs in particular, design decisions. I look at these two key questions. Christmas Cy using anime supporters in the context of congestion control might be a good idea changed from Lynchburg technology to see if I could beat. You have all of these schemes out, there is a bit to the Lord, because if you had talked to figure to go Ajit or
apartments, we still lack a. That's exactly where they need a support. Just can't come to the buy because these approaches potentials to Green Top with mysticism. Now, let's see wine using. Anyways, that congestion control might not be a good idea. Tonight and will perform two simple experiments in the first one, I brought up in this works. As long as the one shown here. Is that wise minor every 10 seconds remaining five hundred times more than the funeral through files. And that is clearly a nap time for any congestion control
cover. Show me, see how car atlanta-based congestion consoles perform in this unseen Network. As you can see here, would you have serious performance problem? The bottom line here is our scenarios series problems. Experimented on by using a 50-foot per second link, and send in traffic or dad for about 2 minutes. I'm sure you hear an overhead of to speak yet again, which is across if you receive this game and also today's platform controls, you remind me to purchase a
problem this time. All together. You see that we have a confrontation between modern approaches and classic approaches. Approaches replacing passenger seat designs. Instead of you like to be pregnant to give my dad trying to replace anyone. The key for us is to have the best of both worlds. On that note, we have a 2 level controller on a PS4, use a block to perform periodic actions prove it. You have a fine drink called show using plastic Christie's, which happened to be acknowledged,
been trying to have fun at the same time, I'm having adopted. Here is the big picture of orca. We do have classic here as soon as I stop doing its job Mission from this board game called Soap. Opera Digest is useful for the other Weiser observations and then sushi. The key thing for disregard is to motivate the engine to hoard minimizing, Bend all these implants will go to the agent and it's basically tries to come up with a better plan. I was looking down on you and
see how did you're the worst but in a more qualitative legal. Hemorrhaging of a scenario in which we suddenly dropping capacity, would like to see how I see you have a steady going to rain. And the other one is we have someone jumping Adeline is his son because we have someone show, you things were pushed toward your fines to probe for mobile. The Asian learn to control. Game Classic Ford TriStar. Which means we have Lori Morris and this maze agent wish to contest him in Oakland Oregon.
In Yonder scenario, we do the opposite. We said only increase the capacity of line to see again system, which means spaghetti Hardie board, because of that, the congestion. Should go higher. Classic bar tries to promote violence class dressing basically, how's the system to review board of Alabama? Now, put see how Orca performs in two experiments that I mentioned in during this time. Tears of the Uber app, you can clearly see here in the door cuz over at is much lower than the other than a major design and support of the job
and that were there happens to be. No over had a partner who speaks your workout has bacon Hair Dynamics compared to the other, like they describe two reasons, why? It's radar, we used, is it be humid as the classic jersey block for about 6 hours, over a small set of sing a better picture of a sperm from a steering, the training. We also designed Clayton, State concert choir which is basically or got weed out the classic heroes fund. Here are the results during the training
three times better. And that is important because of dots to level architecture that I ordered described. Has, she become a cyborg in the inside. Where is the Mayo has been from? Addison Burns scenarios, It Ain't Nothin. Here, I'm sharing with you is also one of those experiments call the Intercontinental scenario. Of course, you can check out to compare the pepper miles of those games with each other. This graph shows, the average normalized results, x-axis and y-axis
is circulatory. System described would be top right side of the graph to see how it does its job compared to the others. And here it is. Chicago fluid in his presentation. Let me point out a single over here that's promising results of our work has not been sending me. What did you put aside more than 3 games if I have to reset. It seems like they they wanted us to wrap a suit to a great work. I'm going to read out some ocean. To the first question is coming from Kong Mei. Dong is curious if the performance of orchid with change significantly, with a different job or function, or
if it concerns different questioned, everyone function is going to park in the system. You fly, if you change it, of course, of behavior of the system. Completely changes the equation from cost of John Carpenter's curious. How does Orca perform in Wireless scenarios with Bangla translation at much? Smaller time scale. Yeah. Actually you we have something for my son to sell our scenario which you know I've been to a higher than the fluctuation so yeah you can check out the paper
for that. Do they know? The last question is from Maria Maria's curious. If you could comment on how fair or tiles two other PCP versions us and of course a fairness self unite in the paper. Every show that you know, we have both properties. question about the set of the overhead, gravity-defying overhead like If he misses Asian, we counted that as that at work. Question, I had was. Jamie is also songs of combining light classical sticks with the machine learning in the sense that
it's just learning the prom and WhatsApp. Yeah, like how do you sort of Remy in comparison to all the trouble that you were sick last night? He has to, basically go. And, you know, how to find a brute-force matter to figure out Keno in spending the entire space. What would be in at the best action as possible in a sort of going through a decision tree? You know, and I'll try to figure out in the entire space. What will happen and we also in terms of what we observed was Remy in the real, sorry, so I can have internet InterContinental that nature, that's all. I actually was a very bad
job. So in terms of the weather, There are so many condition control methods can the situation of data centers and work as well as a quick question about. It's not a scheme. It's basically, you can customize version of that for yourself in an environment, what we did in this paper was we focused on the internet version of the internet. So you can imagine that you if you're interested in having as a memory customized version of the Orca, you can have it, you know, drainboard your own this part and then you can use, you know, which works better in
your network. You might be no end up losing DCTC instead of a cubic. So the choice of cubic princess for us was because of some, you know, probably just give it gas. Ernest, Brandon s do sororities in Indiana. Latency and stability under.
Buy this talk
Buy this video
Our other topics
With ConferenceCast.tv, you get access to our library of the world's best conference talks.