🚀 go-pugleaf

In article <14957@bellcore.bellcore.com> karn@ka9q.bellcore.com (Phil Karn) writes:
>>It turns out that ISO IP is more efficient in its use of
>>memory; you know at the first fragment the length of the entire message,
>>so you can preallocate...
>
>I don't quite understand this. If you represent packets as linked lists of
>dynamically allocated buffers, then there is no need with either DoD IP or
>ISO 8473 to preallocate memory when reassembling fragments.
> [...]
>It is true that knowing how long an entire datagram is could save you some
>CPU time if you're so memory-starved that you won't be able to reassemble
>it; you could toss all of the fragments as soon as you receive them instead
>of running out of memory halfway through reassembly and having to toss the
>partially reassembled datagram.  I don't see this as a big advantage,
>though, since any system that runs out of memory often enough for this to be
>a signficant performance factor is going to have many other, much more
>serious problems.

As Chris Kent and I wrote (after Dave Mills raised the issue) in our
paper "Fragmentation Considered Harmful" (SIGCOMM '87), one problem
is that because IP doesn't tell you how much space is needed, you
can run into pseudo-deadlock when several large packets are being
reassembled.

If the rate of incoming fragmented packets is high enough, it doesn't
matter how much memory you have, because the situation is unstable:
once you begin to run out of memory, you're likely to see new fragments
arrive faster than the old ones time out.

The real problem, of course, is that IP-style internetwork fragmentation
is generally a bad idea ("Harmful", in fact).  Issues of "efficiency
in the use of memory" are second-order compared to efficient mechanisms
for avoiding fragmentation.  I certainly agree with Phil that variable-
length fields are a false economy.

-Jeff