Homa: The end of TCP for AI clusters [video]
youtube.com
youtube.com
The core idea: When a message arrives at the sender’s transport module, Homa divides the message into two parts: an initial unscheduled portion (the first RTTbytes bytes), followed by a scheduled portion. The sender transmits the unscheduled bytes immediately, using one or more DATA packets. The scheduled bytes are not transmitted until requested explicitly by the receiver using GRANT packets.
So it sends blind for short requests, then needs a go-ahead from the receiver. That's reasonable when the main application is a remote procedure call. It's reminiscent of QNX's networking protocol, which is also single packet message request/response but can also handle arbitrarily long messages.
What makes this work today is that per-packet processing overhead in hardware switches is low vs. per-byte overhead. In early software driven switches, per-packet overhead tended to dominate, and sending small packets was very inefficient. In modern hardware switches, where FPGAs are doing the processing, the per-packet overhead is low enough that small packets are not inefficient.
It's amusing that web stuff is so bloated today that any transaction under 1MB is considered "small". So this is not a suitable protocol for open web use.
[1] https://people.csail.mit.edu/alizadeh/papers/homa-sigcomm18....
Also, thanks for the clear summary and context.
1. No way to detect whole RPC loss. Since there is no outer connection state, if every packet in the send-side of a RPC is lost then there is no way for a server to detect that it should issue a resend. The RPC is just lost to the ether. This affects small messages, like messages that fit in a single packet, more since there are fewer packets in the send-side.
2. Related to the above, there is no builtin encryption support. So, if you want encryption then you need to layer it either above or below.
3. Benchmarked performance is awful. The 60 kB average message case in [2] Table 4 takes 5(!) hyperthreads to average 20 Gbit/s. That is just 4 Gbit/s per hyperthread. Even a totally naive one-packet per system call network protocol design and implementation should get to ~8 Gbit/s per hyperthread. 30 Gbit/s per hyperthread is easy with just a little focus on performance.
4. Despite all the performance design problems in QUIC (though still faster than Homa) it already solves basically every problem Homa is trying to solve in a much cleaner way. Stream IDs correspond to RPC IDs. Stream Max corresponds to Grants. Multiple streams under single Client allows prioritization.
Except you do not randomly lose entire messages. You can compact small messages into packets. You get more precise RTT time allowing more accurate pacing/congestion calculations. You get builtin encryption. It survives ossified middleboxs. It has multiple ack frames/packets reducing ack overhead.
The only real difference is that Homa uses explicit receiver Resend instead of implicit sender Resend. Except that actually consumes significantly more receiver resources in non-trivial loss scenarios especially due to Resend packets only supporting a single Resend span. It also incurs higher latency and has higher requirements on the entire lossy network path due to all the extra transiting data that you do not want to lose.
And those are just serious problems off the top of my head after reloading the RFC into my head. I can come up with some more if needed.
[1] https://github.com/johnousterhout/homa-rfc/blob/main/draft-o...
[2] https://www.usenix.org/system/files/atc21-ousterhout.pdf
However, software QUIC implementations are generally much slower than software TCP implementations. This is not due to any fundamental protocol design limitations, it is just that most/every QUIC implementation is poorly implemented for maximizing performance.
You can, of course, do multiple times more throughput than QUIC or TCP with a protocol better designed for performance, but that is likely orthogonal to your question.
2. TCP doesn't have encryption either. They should probably show DTLS working though.
4. You can't just say QUIC is faster without testing it.
That is a sender-driven resend and as such relies on the sender identifying the “send lost” condition. However, the entire protocol is designed around not doing that and thus you can only feasibly rely on “should have received a reply by now” as your timeout.
But Homa is intended to be a RPC protocol. So you send to server, wait for server to process command, then wait for server to send reply. Your timeout depends on the variable and heterogeneous command processing time.
Even if you were able to give a separate correctly tuned timeout for every possible RPC that is still awful. Any RPC with long processing time should not trigger the timeout until the expected reply time, but if that is far larger than the RTT then you are waiting a tremendous amount of time.
For instance, a select query that is only a few bytes long (and thus fits in one packet) could take seconds on a large database even though the database server is physically nearby and only microseconds away. In that case you would have to wait seconds before timing out instead of just microseconds like a ack-based design could achieve. The worst-case transport latency becomes application-specific instead of related to the application-independent transport parameters like RTT. That is troubling.
2. TCP can be forgiven for that given it predates asymmetric cryptography and even DES. Not considering it relevant in a new protocol this side of the millennium is much less reasonable.
4. MsQuic at 7.5 Gbit/s: https://microsoft.github.io/msquic/
And then in basically every other respect it is worse including compatibility.
Because the NSA and similar organizations will 100% infiltrate the physical infrastructure. It’s easier and harder if not impossible to detect.
This is self-evident, isn't it?
Other than the buzzword factor, why do you think name-dropping AI makes a difference?
That is relevant for musks in-space computation. There, latency between compute nodes, due to being hundreds of kilometres apart, cannot be low.
This pretty much limits the size of AI models which can be trained on a space datacenter to a single satellite to avoid incurring huge latency overhead both in training and inference.
With the smartest models getting bigger and bigger, that really doesn't look good for space based compute.
On the other hand, if the traffic follows a very predictable pattern, a custom implementation can be much more efficient. Specially nowadays machine learning can find much better solutions through reinforcement learning. And AI cluster data flow is much more predictable than what goes over the internet as a whole.
> Link Level Flow Control: InfiniBand uses a credit-based algorithm to guarantee lossless HCA-to-HCA communication. RoCE runs on top of Ethernet. Implementations may require lossless Ethernet network for reaching to performance characteristics similar to InfiniBand. Lossless Ethernet is typically configured via Ethernet flow control or priority flow control (PFC). Configuring a Data center bridging (DCB) Ethernet network can be more complex than configuring an InfiniBand network.[19]
* https://en.wikipedia.org/wiki/RDMA_over_Converged_Ethernet
> A sending station (computer or network switch) may be transmitting data faster than the other end of the link can accept it. Using flow control, the receiving station can signal the sender requesting suspension of transmissions until the receiver catches up. Flow control on Ethernet can be implemented at the data link layer.
When a link becomes saturated working out how to manage that is a hard problem.
I have only used RoCE once at scale, it was really finicky. We would get big waves to pause frames that stalled everything.
Consider a very simple topology:
A C
\ /
S1===S2
/ \
B D
S2 is receiving packets destined for C faster than it can forward them, but sending an Ethernet pause frame from S2 to S1 is not a very productive way to alleviate the situation, because it also disrupts any traffic that would be bound for D. It just moves the bottleneck elsewhere and causes collateral damage.
I agree that Ethernet flow control is insufficient, but given that NVidia in an act of brilliant foresight acquired Mellanox a decade ago, I’m fairly certain that this is how all these AI clusters are actually deployed, not using TCP, and maybe not even using Ethernet but infiniband instead.
if not, you get RoCE over ethernet.
Blog post: https://openai.com/index/mrc-supercomputer-networking/
Paper: https://cdn.openai.com/pdf/resilient-ai-supercomputer-networ...
Spec: https://www.opencompute.org/documents/ocp-mrc-1-0-pdf
MRC is an extension of RoCE that does out-of-order packet delivery, packet spraying across all available paths using SRv6 routing, combination of ECN and packet trimming for congestion control (so sender-based CC like TCP and unlike Homa).