GC leaks debugging

[Andrew Haley]

On 12/04/11 19:42, Erik Groeneveld wrote:
> Hans, Andrew,
>
> Having concluded for myself that fragmentation causes the heap to grow
> indefinitely, it tried to find a workaround.  Because changing
> different (environment-, build-, runtime-) variables didn't help, I
> started looking at the code itself.
>
> I found that all memory allocation calls from GCJ eventually come down
> to GC_allochblk(), so I started gathering some statistics about it.
> It turned out that it wasn't called that often at all, so I just added
> a forced collect to see if my assumptions were right, risking much
> slower runtime of course.  I tried:
>
> @@ -50,6 +52,13 @@
>      /* Do our share of marking work */
>          if(GC_incremental && !GC_dont_gc)
>             GC_collect_a_little_inner((int)n_blocks);
> +
> +    if (n_blocks >= 8) { // 32 kB and bigger often occur in fragmented heaps
> +           GC_gcollect_inner();
> +           printf(">>> forced collect <<<\n");
> +    }
> +
>      h = GC_allochblk(lw, k, flags);
>  #   ifdef USE_MUNMAP
>         if (0 == h) {
>
> I ran my test, and ignored it slowness (only noticing that it was not
> so much slower).  But it works:
>
> Before: 29,000,000 docs, 820 MB heap, OOM.
> After: 67,000,000 docs, 490 MB heap.  Disk full ;-(
>
> So frequent collection can certainly avoid fragmentation in this case.
>
> Now the most curious of all: it is even faster as before:
>
> Before: 1306 docs/second
> After: 1582 docs/second
>
> Apparently, it is better to collect a small heap more often than a
> large heap less often.

This is very interesting.

> Now this hack helped me to assert my assumptions, but it also works
> well enough that I am going to try it to relieve some of the stress
> that has been plaguing some productions systems for quite some time
> now.
>
> Meanwhile, I'd like to pursue a better solution - less of a hack.  Any
> interest in helping out?

Before you go any further, it's worth remembering that you're using an
old version of the GC.  I've been told that "Of course, [the new gc]
will require modification of boehm.cc in GCJ" but not why.

It looks like you're making progress, but I urge you to move to the
new gc or your time may be wasted on the old one.

Andrew.