Skip, those 3 ideas have been around forever and a day in HPC. I first saw flow control by the receiver in the HIPPI interface from LANL back in the (early) 90s. I'll read the thesis but the talk was disappointing for me.
I am still unconvinced that the proposed solutions address a problem we've seen in 9p as implemented on Plan 9. The big problem we measured in high delay-bandwidth networks is the serialization of RPCs on the fid in muxio(). To some extent, the way 9p is used requires that: it makes no sense to kick off readahead on devices, for example; and if you do a write-read-write on the mouse location you want the write to complete before the read, and you want the read to complete before the write (consider moving the mouse pointer: change it, read it, change it again: there's no way to overlap those, they have to complete). So I'm not seeing suggestions here that align with the measurements we made of 9p back in the day. But you know 9p very well and I'm willing to be educated. On Mon, Sep 28, 2026 at 12:43 PM Skip Tavakkolian < [email protected]> wrote: > I suspect that there is more nuance, especially given that it is > coming from Ousterhout. The parts that caught my attention were: > > 1. flow control is done by the receiver -- because, as they assert, > network congestion in data centers mostly happens on the downlink to > the destination host. > 2. message-based > 3. short messages are given priority > > The third one makes me wonder if 9P's Rwrite and Rread should be > treated differently by both receiver and sender. Maybe they could have > flags that throttle the sender (I think this is what in Homa does with > GRANT replies). > The first point is already done in 9P, but it is at a higher layer > where TCP has already done congestion control. > > Synopsis of the protocol: > https://github.com/PlatformLab/HomaModule/blob/main/protocol.md > > On Mon, Sep 28, 2026 at 8:04 AM ron minnich <[email protected]> wrote: > > > > I think the proposed solution would benefit from a description of the > perceived problem. > > > > 9P *already* multiplexes lots of I/O across multiple QIDs. > > > > We've been here before, so many times. Floren spent some time back in > 2008 measuring a couple hundred million 9p packets transactions, in the lab > and across the US. You can read the result in his thesis. Multiplexing > across multiple QIDs was not the issue. > > > > I'm unable to understand how Homa and QUIC could solve any of the > problems I've measured over the last 30 years. > > > > How would QUIC help 9p performance? I.e., what bit of code can you point > to in the kernel that QUIC would improve? Where are your numbers? > > > > > > > > On Sun, Sep 27, 2026 at 9:32 PM Thaddeus Woskowiak < > [email protected]> wrote: > >> > >> Another protocol approach to improving 9P performance might be QUIC. > >> You could multiplex IO on multiple open qid's without hitting > >> congestion issues. It is also designed for the internet and runs over > >> UDP to avoid protocol ossification making it a bit more practical. It > >> also performs a fast, simple handshake and integrates TLS. > >> > >> https://en.wikipedia.org/wiki/QUIC > >> > >> On Tue, Sep 22, 2026 at 1:48 AM Skip Tavakkolian > >> <[email protected]> wrote: > >> > > >> > I will preface this by acknowledging that I know very little about the > >> > details of the protocol and I could be totally off the mark. > >> > > >> > I only recently found out about Ousterhout's latest project from this > >> > recent talk. Homa protocol is based on the paper referenced below. My > >> > first impression is that the 9P (RPC) on top of IL seem very similar > >> > to one of Homa's features -- the receiver-driven congestion control. > >> > I'm also wondering if a variation of IL could implement these > >> > features, or if 9P could benefit from use of Homa. Has anyone done > >> > any deep thinking? > >> > > >> > Ousterhout's talk > >> > https://www.youtube.com/watch?v=eZ8WWZzoaR0 > >> > > >> > Homa: A Receiver-Driven Low-Latency Transport Protocol Using Network > Priorities > >> > https://people.csail.mit.edu/alizadeh/papers/homa-sigcomm18.pdf > > > > 9fans / 9fans / see discussions + participants + delivery options > Permalink ------------------------------------------ 9fans: 9fans Permalink: https://9fans.topicbox.com/groups/9fans/Td0379c45387a908c-M932365fc5a1a98b4ceaddee9 Delivery options: https://9fans.topicbox.com/groups/9fans/subscription
