[ih] History from 1960s to 2025 (ARPANET to TCP)

[John Day]

No, for some time now the problem has been too much memory.  That is why there is buffer bloat.

There is a paper by Raj Jain from the early 90s on Myths of Congestion Control in which he points out that several of the myths are: faster processors, more bandwidth, more memory or more of all 3 will solve congestion. All of those can make the problem worse.

The solution has been known since the late 80s: early notification to back off before it happens and thus also avoid lots of retransmissions and buffer bloat.  But the Internet didn’t do that. They adopted a solution to maximize retransmissions and is predatory.

As has been pointed out for 50 years, datagrams were adopted by CYCLADES, which was a network to do research on networks, as the minimal packet and a building block. It would be more fruitful to treat them like that rather than as a religious icon.

Actually, the Denning paper is on buffers for terminals in timesharing systems. But the difference is so huge, it is obviously a ’no-brainer’, except it is seldom followed. Of course it is no-brainer except when it isn’t: buffer bloat.

Take care,
John


> On Jan 5, 2026, at 23:20, William Westfield <westfw at mac.com> wrote:
> 
> 
>> … pooled buffers require 3 or more orders of magnitude less buffer space than static allocation of buffers. …  The practice at the time was that VC nailed everything down and tried to make it deterministic, i.e., static allocation. This demanded either more memory (expensive then) or fewer virtual-circuits of less capacity.
> 
> Ah, yes, the good old days when Windows were supposed to be some indication of memory resources on the endpoints, instead of a mechanism for network-side congestion avoidance…  (although, I remember getting dramatic LAN throughput increases by implementing an “optimistic window strategy” in Tops20.)
> 
> Hmm.  Should the question be re-examined now that “memory is cheap”, or has the increase in required buffering kept up with the advances in memory?
> 
> BillW
> 
>