And that was the 90s, after Lisp had been around for decades...

It's hard to say how many time was spent on GC in a Lisp program in the 70s. The installations were different, programs were different. How much time did a Multics Emacs written in Maclisp spend in GC? I have no idea. A macsyma running for some hours? AN early rule based system? Probably there were taken some benchmark numbers, but one would need check the literature to find them.

To see how different benchmark numbers were:

http://rpgpoet.com/Files/Timrep.pdf

There are also pathological cases where Memory is mostly full and time would be spent mostly in GC, or where a lot of virtual memory is used, and much time is spent in IO paging in/out memory.

Note also that the invention of generational GCs was triggered by the availability of larger virtual memory based systems and the programs using it - available to the market outside of the original research. Before that, most of them were quite a bit smaller or very rare. Systems like the Symbolics 3600 were extremely rare and expensive. When they first shipped, GCs were copying and compacting, but not generational. Systems like that did not exist on the market before.

The claims being debated here are, as I understand them, my half-remembered claim that it was common to spend ⅓ to ½ of your execution time in the garbage collector until generational GC, and Simon Brooke's claim that actually 5% was normal and 10% was in pathologically bad cases. I think Gabriel's book shows conclusively that allocation-heavy programs could easily spend ⅓ to ½ of their time in GC, or 80% to 90% in pathological cases. But Brooke could still be right about the typical case, since when allocation is expensive, people avoid it in performance-critical code when possible.

I agree that the problem mostly went away once generational (aka ephemeral) GC was shipped. In the essay, I dated that to Lieberman and Hewitt in 1980, but in a more recent discussion with Allen Wirfs-Brock, I learned that actual generational GC didn't ship until 1986. (And Lieberman and Hewitt's paper, though submitted in 1980, wasn't published until 1983.)

Given the question of whether non-generational GC would cost more like 5% or more like 25% of your performance, the performance of generational GCs is somewhat of a false trail. Hopefully nobody was claiming generational GCs would eat double-digit fractions of your CPU time. I wasn't, at least.

Ephemeral GC btw. means something slightly different. An ephemeral GC is mostly concerned with the detection and collection of short lifetime objects in main memory.

The main problem is that architectures, implementations, actual computers, programs were so different, that you could find all kinds of numbers. Even a generational GC can get extremely busy and may not prevent from having full GCs over large virtual memory from time to time. It's also a huge difference if the GC was incremental (the work was spread over the whole compuation) or if the system stopped for much of the GC work.

But come on, I have used emacs for 25 years, and on a daily basis it stalls for annoying amounts of time while I am just doing something simple like editing a .cpp file. Today. In the year 2017.

The performance claim is from my memory of reading LISPCraft, the 1984 edition written before generational GC, where Wilensky claims that something like a third to half of your program runtime is typically spent in GC, so consing is much more expensive than it appears. But I don't have a copy of LISPCraft here, and I read it in probably 1990 or so, so I might be misremembering.

How can we verify the truth about these conflicting claims about LISP systems from the early 1980s? Maybe Gabriel's book has a more definitive answer?

(A possible confounding factor is that, at the time, the frequently-executed parts of Lisp programs were written in a more imperative style to reduce the load on the GC.)

I'd like to claim that one of the four configurations of VAX Franz Lisp shone here, since it spent only about ⅐ of its total execution time in the GC. However, it spent roughly as much absolute time in the GC as the other three Franz Lisp configurations; it's just that it took 6× as long as normal to do the non-GC part of the benchmark; "normal", for context, was several times slower than SAIL. It's not clear that this counts as a performance improvement in any sense!

But, since many of his benchmarks use 0 GC time, I don't think I can justify the statement that GC just before the introduction of generational GC normally used 50% of a program's total run time; most programs did not consist primarily of evaluating tree-rewriting rules, which obviously are allocation-intensive. (Boyer isn't the worst, though. On Vaughan Pratt's Deriv symbolic differentiation benchmark, SAIL spends almost 90% of its total CPU time in the GC, though some other collectors handle it better.)

Reading through some of the other benchmarks, Franz Lisp commonly comes out rather GC-heavy, though less so than SAIL; on div2, for example, Franz spends 80% of its time in the GC. It may be significant that Wilensky was using Franz to write the book I got my original impression from.

I'm somewhat surprised that the FFT benchmark is written to use GC time, since normally I think of FFTs as not requiring any allocation at all, much less significant numbers of allocations of separate objects...

So, in conclusion, I do think it's justifiable to say that there were significant classes of programs for which all Lisp systems continued to spend ⅓ to ½ of their time in the GC, or 80% to 90% in bad cases, right up until the introduction of generational GC shortly after Gabriel's book. But it's probably not justifiable to apply that statement generally to all Lisp programs. Your 5%–10% number sounds quite reasonable as an overall average.

Keep in mind:

a) not every machine had FP hardware, so FP instructions might be emulated in software.

b) dealing with floating point numbers may mean allocate float objects. For example not every Lisp had a direct array of floats. The array then would be a general array pointing to float objects on the heap.

c) not every compiler/interpreter was clever enough to keep floats in registers (or similar) while doing float computations.

Even today, allocating floats can be a problem in Lisp applications.