Steel Bank Common Lisp

"Peter Hertmuller" <phertmuller@...> writes:

> I'm having problems running sbcl. I'll try to be as concise as possible:
> 
> I start SBCL (via SLIME in a terminal), start my program, and after
> sometime, SBCL crashes, with a message like this:
> 
> Argh! gc_find_free_space failed (first_page), nbytes=8.

This means that you've run out of memory (a maximum of 512MB can be
allocated on Linux/x86 for and sbcl compiled with the default
settings).

>     Gen Boxed Unboxed LB   LUB  !move  Alloc  Waste   Trig    WP  GCs Mem-age
>    0: 128844     1    27     0    13 527686944 172768  2050840    0   1  0.0000

And this says that practically all of the memory was used by newly
allocated small objects. Which really shouldn't be happening: while
these objects are being allocated, garbage collections should be
triggered, which move some of them from the 0th generation to older
ones.

Is your application maybe disabling the GC or alternately disabling
interrupts?

-- 
Juho Snellman

Benjamin Lambert <benlambert@...> writes:
> Is there any documentation available that describes what the report  
> from sb-sprof means? The main thing I'm confused about is what are  
> the differences among "self", "cumul", and "total".

No, it's not documented. One reason is that I keep wondering whether
"cumul" and "total" should be switched around, which we really
shouldn't do once they've been formally documented.

"self" is time spent executing the function itself. "cumul" is the
time spent in the function or functions called from it. "total" is the
sum of the "self"s of this function, and all functions above it in the
list.

Or maybe "cumul" and "total" were the other way around. I
forget... ;-) It should be obvious from the report which is which.

> But also I wonder exactly what the profiler samples.  Does it just
> sample the function that is currently executing or does it sample
> the entire call stack?

Neither. By default it samples eight levels of the call stack on x86
and x86-64, and two levels on other platforms.

-- 
Juho Snellman