Homa: The End of TCP for AI Clusters — John Ousterhout, Stanford
Read the talk
Homa: Keeping Small Messages Fast in AI Clusters
John Ousterhout explains why short coordination exchanges can stall GPUs, how sender-driven congestion control compounds the delay, and how Homa combines message lengths, receiver grants, and switch priorities to reduce it.
From a talk by John Ousterhout
At a glance
Ideas worth remembering
When GPU nodes must all finish a coordination exchange before computing again, one slow small message can stall the group. Tail latency becomes a throughput concern as compute phases shrink toward milliseconds.
Incast concentrates traffic at a receiver’s last-hop queue. Sender-side feedback arrives after queues begin forming and must coordinate several senders through delayed, limited information.
Homa exposes message lengths, keeps messages independent, and uses receiver grants to favor messages with less work remaining while limiting simultaneous incoming traffic.
Switch priority queues let short messages bypass queued bulk traffic, complementing the receiver’s control over future transmissions.
The presented benchmark reports about thirteenfold better short-message P99 latency and almost twofold better performance for the longest messages versus TCP. Those results motivate application experiments; they do not establish equivalent AI throughput gains or a measured advantage over RoCE.
When throughput depends on latency
AI networking has traditionally served enormous transfers: gigabytes of data moving between machines for work such as exchanging weight gradients. For those transfers, sustained throughput matters most. Connection setup and a slow start contribute relatively little to a transfer that runs for a long time. John Ousterhout begins by acknowledging that TCP and RDMA perform pretty well in this setting. His argument for a different transport starts with a change in the work the network must serve.
Ousterhout sees inference and agentic workloads becoming more granular, with smaller chunks of computation separated by smaller exchanges of data. He distinguishes these from training workloads, which still involve massive transfers. He also narrows his use of RDMA in the talk to RoCE, RDMA over converged Ethernet. The comparison concerns data-center transport for mixed traffic, rather than every possible use of remote memory access.
Small exchanges often carry metadata or coordinate work. A node might check whether an entry exists in a distributed KV cache, or nodes might synchronize at a barrier after a period of computation. The useful measurement becomes the round-trip time: send a small request, perform a little computation elsewhere, and receive a small result. Moving many gigabits per second does not by itself make that exchange finish quickly.
The more important measurement is often tail latency, such as the ninety-ninth percentile, rather than the average. Consider a workload distributed across several GPU nodes. Each node computes, then participates in an exchange before the next compute phase can begin. During synchronization, the GPUs sit idle. One slow exchange can hold up the entire group because all exchanges must finish before computation resumes.
The cost depends on the relative duration of computation and communication. In Ousterhout’s example, a few milliseconds of synchronization after five seconds of computation barely matter. If the compute phase itself takes milliseconds, a synchronization phase that also takes milliseconds consumes a significant share of the cycle. That is how latency for a small message can limit overall throughput and waste GPU resources. His audience poll draws more raised hands than he expected, though it supplies no numerical estimate of how widespread the problem is.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
How incast traps a short message
Ousterhout identifies incast as a typical cause of poor tail latency: several nodes simultaneously send data to one destination. With equal-speed links, three senders can collectively transmit three times as fast as the destination can receive. Packets accumulate at the last hop, in the top-of-rack switch’s egress queue for that destination. The bottleneck is the shared downlink to the receiver.
A short message arriving at that same queue waits behind packets from the large transfers. Its own size may be tiny, but its completion time includes the wait for traffic already ahead of it. If the switch exhausts its buffer space, it drops packets, adding timeouts and retransmissions to the delay. Congestion control must therefore limit how much the senders inject into this shared bottleneck.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The sender learns about congestion late
In the sender-driven protocols Ousterhout discusses, each sender must discover congestion and reduce its transmission rate. The difficulty is placement: the sender makes the decision, while the queue may be at the far end of the data-center network. An older approach detects missing acknowledgments after packets have been dropped and treats the loss as a reason to slow down. By then, the queue has already overflowed.
Switch feedback can signal trouble earlier. Ousterhout describes ECN marking: once a queue reaches a threshold, the switch marks passing packets before the buffer overflows. The receiver observes those marks and passes the signal back to the sender, for example in an acknowledgment. The sender can then reduce its rate without first waiting for packet loss.
Earlier feedback still leaves a hard coordination problem. In his explanation, the sender receives one bit of information indicating congestion somewhere. That signal does not tell it precisely how much to reduce its rate or when to increase it again. Several senders may receive the same warning and adjust simultaneously, even though their combined behavior determines whether the destination’s queue grows or drains.
Control lag makes those adjustments unstable. Information takes time to return, and a sender may need several round trips to approach a rate that matches the available bandwidth. Meanwhile, other transfers begin or finish. The target changes as the control loop tries to reach it, producing oscillation between sending too much and too little. Ousterhout acknowledges improvements from more than twenty years of research, but argues that sender-side control retains a fundamental disadvantage: in this feedback scheme, queues must build before the sender learns that congestion exists, so delay has already begun.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Message lengths give the transport something to schedule
A separate problem comes from hiding application messages inside a byte stream. Ousterhout illustrates it with a TCP socket: the application sends distinct messages, but TCP sees serialized bytes without their message boundaries. The transport cannot use those boundaries to know how much of a message remains or to favor a short message. If an application sends two large messages followed by a small one in the same stream, the small one waits behind them. This head-of-line blocking can occur before the short message even gets a chance to compete in the switch.
Homa approaches data-center transport through a clean-slate design intended to handle large and small messages together. It began as a student’s PhD dissertation, after which Ousterhout took on the work of moving it toward production. At the time of the talk, he had written a Linux kernel module, made it available on GitHub, and was working through the process of upstreaming it into the kernel. That describes an available implementation and an integration effort in progress; it does not establish that upstream inclusion had happened.
Homa’s fundamental unit is a remote procedure call: a request message from client to server, followed by a response message from server to client. Message lengths remain visible throughout the transport. When the receiver gets the first packet of a message, it learns how much more data the sender wants to transmit. That gives it a concrete amount of pending work to manage, rather than only a warning that a queue has started to fill.
Homa uses Shortest Remaining Processing Time, or SRPT, to favor messages with less work left to complete. Independent messages can bypass longer ones because they are not serialized into a single stream. Message lengths and independence serve different purposes: the length helps the transport choose which message to favor, while independence makes it possible to act on that choice without first finishing earlier large messages.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The receiver grants permission to send
Homa moves congestion control to the receiver, close to the last downlink where Ousterhout says congestion primarily occurs. The receiver can see the lengths of incoming messages once their first packets arrive and coordinate their progress together. This gives it more useful information about the competing transfers than a sender receives from a delayed congestion mark. The rationale concerns the receiver’s incoming traffic and bottleneck; it does not require the receiver to know every future transfer elsewhere in the network.
A sender first divides its message into packets and transmits only the first few without permission. Homa calls these unscheduled packets. The remaining scheduled packets wait for grant packets from the receiver. The receiver paces those grants over time, authorizing the next chunks of data. Initial packets reveal the message and its length; grants then control how the rest enters the network.
Suppose ten messages are arriving at one receiver. Granting all ten permission to send together would recreate congestion in the top-of-rack queues. The receiver can delay some grants to limit simultaneous traffic, and it can favor shorter messages when choosing which grants to issue. The same mechanism therefore controls admission to the bottleneck and implements the scheduling preference. It limits the bulk of incoming traffic without claiming to eliminate the initial unscheduled packets.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Switch priorities let short messages pass queued traffic
Receiver grants govern future transmissions, but Homa also needs a way for short messages to pass long-message packets already queued in a switch. It uses the multiple queues available at each egress port of modern data-center switches—typically eight, in Ousterhout’s description. Packet fields select a queue, and the switch preferentially transmits packets from the highest-priority queue.
Homa dynamically assigns priorities to favor shorter messages. Returning to the incast example, long-message packets accumulate in the lowest-priority queue, while a short message enters a higher-priority queue and bypasses that backlog. The three design choices work together: message lengths identify the work to favor, receiver grants regulate incoming traffic, and switch priorities preserve the preference when packets reach the congested hop.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A thirteenfold tail-latency improvement—and what to measure next
Ousterhout presents one benchmark used to tune and evaluate Homa. Several machines exchange messages ranging from about fifty bytes to a megabyte, with request and response messages of equal length. The comparison measures round-trip time by message length, with separate P50 and P99 curves for TCP and Homa. P50 is the median; P99 captures the slower tail. This is a mixed-message transport benchmark, rather than a reported end-to-end AI application result.
For short messages, he reports TCP tail latency above one millisecond and Homa tail latency below one hundred microseconds, an improvement of about thirteen times. That directly addresses the synchronization problem from the opening: a short exchange is much less likely to spend milliseconds behind bulk traffic. It does not establish a thirteenfold application speedup, because the effect on throughput also depends on how much time the application spends computing and waiting.
Favoring short messages might seem likely to penalize long transfers, yet Ousterhout reports that Homa is almost a factor of two better than TCP even for the longest messages in this benchmark. He attributes that result to run-to-completion approaches, which he contrasts with TCP’s fair scheduling. He does not explain the mechanism further in the talk. The result shows that short-message priority did not require worse long-message performance in this experiment; it leaves open how that tradeoff behaves across other workloads. The displayed comparison also concerns TCP, so these figures do not quantify an advantage over RoCE.
Ousterhout ends with a conditional adoption question: is high latency for short messages affecting your application’s throughput? He expects the role of short messages in AI to grow, while explicitly leaving that forecast to be tested over the next year or two. For workloads that already show the bottleneck, he recommends experimenting with Homa and suggests that tail latency could improve by an order of magnitude or more. The useful starting point is the application’s synchronization behavior, rather than assuming that every AI cluster needs a different transport.
He also makes a concrete offer to help adopters get started, answer questions, and fix bugs. Ousterhout describes stepping back from Stanford so he can spend one hundred percent of his time working on Homa. Readers interested in that invitation can find him through his supplied Stanford page. His closing emphasis is practical: measure the short-message bottleneck, try the available implementation where it matters, and work through the problems that arise.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Read the complete timestamped transcript
- 0:12
Please welcome to the stage the Professor Emeritus at Stanford University, John Ousterhout.
- 0:37
Good morning. It's really great to be here to talk about the network side of AI applications, and in particular, to make the case that latency matters, and it's probably gonna be mattering more in the future. But I just wanna say this is a talk that's unusual for me. I've never before given a talk where there are fog generators in the auditorium. Just a really San Francisco experience, I guess.
- 1:01
So it's, it's well known that AI workloads depend on really great networking performance in order to achieve their own performance. And of course, that's because the workloads are so large that they have to be distributed across machines, and then you have to communicate between the machines. But what I wanna talk about today is that it seems that those workloads are changing. And so I hope to do three things over the next fifteen or twenty minutes. First, to convince you that in fact the workloads are changing, and that whereas the workloads used to be completely dominated by large transfers, where
- 1:31
throughput is the key metric that matters, that we're seeing more and more smaller transfers where the latency is crucial. The second thing I hope to do is to convince you that legacy protocols like TCP and RDMA are poorly suited to this environment. They weren't designed for this environment, and unfortunately, they suffer from very high tail latency when you mix small messages with large ones. And I'll talk a little bit about why that's the case. Then third, I'd like to introduce Homa, which is a new
- 2:01
protocol we've developed at Stanford that actually was designed in a clean slate redesign to handle data center workloads like these. And in fact, it does quite well on those workloads and can reduce tail latency by an order of magnitude or more. So I'll tell you a little bit about Homa. So let's dive in. First, workloads. Historically, AI workloads have consisted of enormous transfers between machines, and that's all that really mattered. Gigabytes of data for things like weight gradients and, and so on.
- 2:32
In these workloads, what you really care about is throughput, how many gigabits per second you can pump through the pipes. And these are relatively easy workloads for networks because if it takes a while to set up the connection and start the transfer, it doesn't matter. The transfers go on for so long that all that really matters is the throughput. And so in these environments, TCP and RDMA perform pretty well. Uh, by the way, when I say RDMA, what I really mean is RoCE, RDMA over converged Ethernet,
- 3:03
which is the underlying transport that's used by RDMA for most purposes today. So anyhow, the old workloads, big transfers, throughput matters, uh, the legacy protocols work pretty well. However, it appears that the workloads are changing. They're becoming more granular with smaller chunks of computation and smaller exchanges of data, and this seems to be particularly true in the world of inference and also in agentic workloads. Not so much for... Training workloads are still massive
- 3:33
transfers. And so what's happening is that more and more there are small message exchanges, typically for things like metadata and coordination, such as checking to see if a particular entry is present in a KV cache that's distributed or doing barrier synchronization at the end of, uh, periods of compute. And for these workloads, what really matters is latency. That is, what's the round trip time to send some small piece of data across the network, do a little bit of computation and get a small result back again?
- 4:03
And in fact, it isn't ju-just latency or average latency that matters. What really matters is tail latency. That is, you'd like to know that if we send a whole lot of small messages, all of them will complete quickly. So for example, we typically measure things like ninety-ninth percentile latency, and if we have high tail latency, that can limit the overall throughput of the system. So here's an example. Suppose a common thing is to take a workload and split it up across several nodes, which do intensive computation using their GPUs
- 4:33
for some period of time. And then once they've all finished their computation, you do some small exchange between the nodes, exchange data, metadata, and then it'll go on to the next round of computation. And while that exchange is happening, that synchronization is happening, the GPUs are sitting idle. So if even one of those exchanges takes a long time, it turns out the whole process stalls. You need all of those exchanges to complete before you can go on to the next phase of
- 5:03
computation. Now, if the computation phase is, say, five seconds, and it takes a few milliseconds for the exchange, you know, not a, not a problem, and that's historically what it's been. But now with the agentic workloads where you're trying to pump out tokens relatively rapidly at a regular rate, the periods of computation are getting down into sort of the millisecond timescale. And if it also takes milliseconds to do that synchronization, then you're wasting a significant fraction of your GPU resources waiting
- 5:33
for the, the synchronization to occur. So I'm curious, I'd like to just do a, a quick audience poll here. Is there anybody here where you have reason to believe that the latency of small messages is impacting the overall throughput of your applications? If so, can you just raise your hand? See if there's anybody out there today. Actually, more hands than I expected, so f-- quite a few people out there are raising their hands. I think this problem is likely to get worse as the trends continue. So what's going on? Why is tail
- 6:03
latency bad? Well, typically the cause is congestion resulting from incast. So incast is when several nodes all decide simultaneously to transfer data to some destination node. And if they all send large messages, well, the links are the same everywhere in the network, so three nodes can transfer three times as fast as one node can possibly receive. And so what happens is that packets accumulate at the last hop
- 6:33
going to that destination in the top-of-rack switch at its egress port for the destination node. Then, if some other node decides it wants to send a short message to that same destination, the short message gets stuck behind the long ones in the queue there. And actually, that causes delay. In, in the worst case, so many packets arrive that the switch runs out of buffer space that it has to drop packets, and then there are, you know, timeouts and retransmissions that make everything even worse.
- 7:03
So somehow, we need some way to reduce the congestion in those queues. Somehow, we have to get the sending nodes to stop sending so fast, so the queues don't just build up without limit. So the way this is done historically, virtually all network protocols before Homa, including TCP and RDMA, congestion control is the responsibility of the sender. So senders somehow have to figure out that congestion is happening, and they have to slow down their
- 7:33
rate of transmission. Now you, you might wonder, why are senders doing it? 'Cause the congestion is way over at the other end of the data center network. How does the, the sender find out? Well, in the, in the old, really old days, the way they would find out is the queues would overflow and packets would get dropped. The sender would detect the packets got lost 'cause it wouldn't get acknowledgments back, and it would assume that means there's congestion and then slow down its rate of transfer. That's really expensive, so today there are better techniques that mostly involve the switches providing information.
- 8:04
So a top-of-rack switch, when it sees that the queue length for an ingress port has reached some threshold, starting to fill, long before the queue overflows, it starts marking all of the packets that pass through with what's called early congestion notification, ECN marking. And so when those packets pass through to the receiver, the receiver sees the marking in the packets, and then when it communicates back to the sender next, for example, to send an acknowledgment, then it includes that marking that goes back to the sender,
- 8:34
and now the sender sees-- The sender realizes, "Oh, there's congestion someplace. I've got to slow down my rate of transmission." So that's the basic idea. Unfortunately, getting this right is really hard. Really hard. It's very hard for the congestion to figure out exactly how to set its rates 'cause it gets one bit of information. There's congestion someplace, and there are multiple senders all sending to the same destination. They're all trying to make adjustments simultaneously. How much do you cut back, and how do I know when I can ramp up again? And
- 9:04
even worse, it's really hard to do this in a way that's stable because there's control lag. That is, it takes time before the sender finds out that there's congestion. And in fact, using this process, it typically takes several round trips for the sender to gradually adjust its rate to get just the right rate to match the available bandwidth. But by the time you do that in a network, things have changed. New transmissions have started, old ones have finished, and so these systems tend to never stabilize. They're constantly oscillating between
- 9:34
sending too much and, and sending too little. Now, this problem's been around for a long time. It's been known in the research community for more than twenty years now. There have been tons of papers published on it. There have been some improvements made, that's undeniable, but we're still a long ways from anything that works well, and the problem is with the fundamental nature of doing the, the congestion control on the sender side. It just doesn't work very well. So you end up with a lot of queue buildup. And in fact, you can see the only way to find out that there's congestion is if there's queues. And so by that
- 10:04
point, we're already experiencing delays. So that's a problem. There's one other problem with TCP and RDMA also is that their, their basic data model is a byte stream. Just a stream of bytes with no differentiation in it. So if you send a series of messages, say, through a TCP socket, they get serialized into that stream, and on this slide I've, you know, I've shown the messages appear like they have different colors in the stream. Well, there are no colors in real life. TCP has no idea where the message
- 10:34
boundaries are. And that also makes life hard. For example, you don't know how much more data is coming. If you knew how big the message was, you'd know how much more is coming. And you can't prioritize short messages, which we'd really like to do, get the short messages through faster. And you can end up with what's called head-of-line blocking, where somebody sends a series of messages to the same destination, and they send two really large ones and then a small one after that that gets stuck behind them in that stream, and so it gets delayed, and again, you have tail latency issues.
- 11:05
So all in all, TCP and RDMA are just not well suited to this environment. So what do we do? Well, what I'd like to do next is tell you about a new protocol called Homa that we've developed at Stanford, which was based on a completely clean slate redesign for network transport. If you could start from scratch and rethink how you do transport for data centers, how would you do it? And it turns out in Homa, virtually every major design decision is different from TCP and RDMA. TCP, for all the
- 11:35
amazing things it's done, is just not a good match to today's data centers, nor RDMA. So what Homa does particularly well is to manage the combination of large and small messages and to make sure that the messages have re-- short messages have really low latency. So this started off as a PhD dissertation for one of my students, Behnam Montazeri, and then the results were so great that I decided to make it my personal project to see if we could get it out of the lab and into production. As you may know, I'm, I'm not like
- 12:05
most professors in that I love to code, and so I turned this into my, my own programming project. I have created a kernel module for Linux. I'm currently working through the process of getting that upstreamed into the kernel. It's available on GitHub for download. So let me tell you just a little bit about how Homa works. I want to mention three things. First, it's message-based, not stream-based. In fact, the fundamental unit at Homa is a remote procedure call, which consists of two things, a request message sent from a
- 12:35
client to a server, and then a response message returned back from the server to the client. So the key thing here is that Homa knows about message lengths. They're buried in the transport all the way down to the bottom, and this has a bunch of advantages. First, it allows us to predict the future. As soon as a receiver gets the first packet of a message, it knows exactly how much more data the sender wants to send, and that's so-- provides so much more information for doing congestion control.
- 13:05
Second, Homa prioritizes shorter messages. It uses SRPT, Shortest Remaining Processing Time, first to try and prioritize shorter messages. And third, because messages are all independent, they're not s-serialized into a stream, every message is independent, shorter messages can bypass long ones, so they don't get queued behind long messages.
- 13:29
S- the second thing about Homa that's different is that it controls congestion from the receiver. Now, when you think about it, this makes sense because the congestion happens primarily at that last downlink to the receiver. And so the receiver has way more information. In fact, with Homa, as soon as it gets the first packet of a message, it knows exactly how much more is coming. So it has essentially complete information about congestion, and it can therefore respond to congestion much more quickly and much more precisely.
- 14:00
The way things work with Homa is that when a sender has a message to send, it breaks it up into packets, but it only transmits the first few packets, those are called unscheduled packets, to the receiver. Packets after that are called scheduled packets, and they only get transmitted when the receiver asks for them. So the receiver will send grant packets back. It'll pace them out and send those back to the sender over time, telling the sender, "It's now time for you to send me the next chunk of data." And the receiver
- 14:30
can delay those grants. So for example, if the receiver has ten messages that are incoming, there's no point in sending grants to all ten of them because then you'll just get congestion in the, in the top of rack queues. So it can use the grants to reduce congestion, and then it can also use the grants to give preference to its most favorite messages, which would be the shorter ones. So it's a way of, of implementing SRPT by favoring short messages.
- 14:57
The third aspect of Homa is that it takes advantage of the priority queues in modern switches. So modern data center switches have more than one queue at each egress port, typically eight, and they can be used in a priority mechanism where packets get transmitted, uh, preferentially from the highest priority queue. So I've shown only two queues on the slide here, but typically, there's more than that. You can specify in packets, using the various fields of the packet, you can specify which queue it should go into. And so Homa dynamically makes those
- 15:27
choices in a way to give priority to shorter messages. So if we go back to the incast example from a few slides ago, all of those long messages will pile up in the lowest priority queue. But if there's a short message coming, it will use a higher priority queue. And so it will immediately bypass all of the queued packets from the short-- from the, uh, the longer messages and get through to the destination more quickly. So how much of a difference does this make? Uh, here's a-- on this slide I've got
- 15:57
one sample benchmark that I use as part of my tuning and evaluation of Homa. It consists of a workload of a bunch of machines on a network that are exchanging messages back and forth of different sizes, ranging from very small to very large. And on this graph, you can see on the x-axis is the message length, so from about fifty bytes up to a megabyte. The y-axis shows you the round trip time for messages of that length. So this reque-- this uses request and response messages that are the same length. You can see TCP in green, Homa in
- 16:26
blue, and the y-axis is, is round trip time, so lower is better. And for each protocol, I've got two curves. One curve is the P50 curve. That's the median latency for messages of this length. And then P99 is the ninety-ninth percentile, i.e. tail latency for messages of this length. So I want to point out two things. First, the P99 for short messages is dramatically better for Homa. So with TCP, it's more than a millisecond tail latency. Homa is less than a hundred
- 16:56
microseconds, about thirteen times faster. Second, interestingly, you might think that because Homa favors shorter messages, that long messages suffer and get worse performance. It turns out that's actually not the case. Even on the longest messages, Homa is almost a factor of two better than TCP. I don't have time to explain that today, but it has to do with the fact that Homa uses run-to-completion approaches, which are much more effective than fair, than the fair scheduling used by TCP.
- 17:28
So just to wrap up, the role of short messages in AI appears to be increasing. I think, I think it's likely that it's going to continue to increase. We'll see over the next year or two if that happens. And I just want to pose a question to you. You know, as you're running your applications and measuring performance and seeing what the bottlenecks are, ask yourself, is high latency for short messages affecting your throughput? If the answer is yes, then just know there is a solution available. You should give Homa a try. You can probably
- 17:57
reduce your tail latency by an order of magnitude or more. And by the way, this is-- Homa is basically my life mission right now. I'm sort of semi-retired from Stanford, and the reason I did that is so I can spend one hundred percent of my time hacking on Homa. So I'd be delighted to work with you and help you if you decide you want to experiment with Homa. If you need help getting started, answer questions, bug fixes, whatever, you know, I'd be happy to work with you to try and make you successful with it. So, uh, if that is interesting, feel free to contact me. My email's on the slide, or you can Google me too and find me over the internet.
- 18:28
So thanks very much for listening, and hope to hear from some of you.