Re: Half a Job?
Sure, you can get better performance with custom software and hardware. Homa's bet is that by targetting a specific niche area with known performance issues, and network latencies, and by putting the code into a library that's already well understood, you can gain alot of benefits.
It's not meant as a general replacement for TCP/IP, especially not at the desktop. And it's not meant to replace TCP in all hyperscaler instances either.
But for google/FB/AWS/HPC groups, if you can update a library and get better performance and not have to re-write your applications, that's a huge win. But this only works for a very specific use-case.
TCP/IPs strength is that it works over 5 feet as well as it works over 15,000 miles, without any change needed to the program. But as someone who spent alot of time evaluating different tools (bbcp, lftp, rsync_parallel, etc) and various WAN accelerators (Silverpeak, and some others I forget about now) it's not an easy problem. TCP suffers horribly from the Bandwidth-Delay-Product, which means that as bandwidth goes up, the default size of buffers needed for in-flight packets rises quite a bit as well. So you either increase your buffers, or your goto parallel streams, or you have applicances which handle this at the edge for you and basically lie to the application that their packets made it there and to just keep sending data.
So that's one instance (long distances, high RTT, high bandwidth) where TCP fails.
But Homa is attacking the issues at the other ends of the scale, sending data 100m between lots of different hosts. And efficiently using the link speed we already have, isntead of just buying ever faster (10mbs/100mbs/1g/10g/25g/40g/100g/400g) ethernet links. Who else remembers FDDI? Woot! Token Ring? It got high utilization rates out of existing bandwidth.
Ethernet wins because it just throws switching silicon at the problem. Which at short distances doesn't really address the problem Homa seems to be trying to solve.
I for one would love to see something like this get deployed, I'm sure all the big guys are looking into something like this because if they *don't* have to replace TOR and Core switches all the time to get better performance, then that's a win. They can amortize the huge cost over a much huger installed base.
And yes, Homa will never be a generic replacement for TCP/IP, it's not meant to be.
No one runs Infiniband over WAN links, it's not what is solves, but people don't whine about how Infiband can't be used to send cat pictures from the UK to NYC.