malloc: Consistently apply trim_threshold to all heaps [BZ #17195]

Trimming heaps is a balance between saving memory and the system overhead
required to update page tables and discard allocated pages. The malloc
option M_TRIM_THRESHOLD is a tunable that users are meant to use to decide
where this balance point is but it is only applied to the main arena.

For scalability reasons, glibc malloc has per-thread heaps but these are
shrunk with madvise() if there is one page free at the top of the heap.
In some circumstances this can lead to high system overhead if a thread
has a control flow like

    while (data_to_process) {
        buf = malloc(large_size);
        do_stuff();
        free(buf);
    }

For a large size, the free() will call madvise (pagetable teardown, page
free and TLB flush) every time followed immediately by a malloc (fault,
kernel page alloc, zeroing and charge accounting). The kernel overhead
can dominate such a workload.

This patch allows the user to tune when madvise gets called by applying
the trim threshold to the per-thread heaps and using similar logic to the
main arena when deciding whether to shrink. Alternatively if the dynamic
brk/mmap threshold gets adjusted then the new values will be obeyed by
the per-thread heaps.

Bug 17195 was a test case motivated by a problem encountered in scientific
applications written in python that performance badly due to high page fault
overhead. The basic operation of such a program was posted by Julian Taylor
https://sourceware.org/ml/libc-alpha/2015-02/msg00373.html

With this patch applied, the overhead is eliminated. All numbers in this
report are in seconds and were recorded by running Julian's program 30
times.

pyarray
                                 glibc               madvise
                                  2.21                    v2
System  min             1.81 (  0.00%)        0.00 (100.00%)
System  mean            1.93 (  0.00%)        0.02 ( 99.20%)
System  stddev          0.06 (  0.00%)        0.01 ( 88.99%)
System  max             2.06 (  0.00%)        0.03 ( 98.54%)
Elapsed min             3.26 (  0.00%)        2.37 ( 27.30%)
Elapsed mean            3.39 (  0.00%)        2.41 ( 28.84%)
Elapsed stddev          0.14 (  0.00%)        0.02 ( 82.73%)
Elapsed max             4.05 (  0.00%)        2.47 ( 39.01%)

               glibc     madvise
                2.21          v2
User          141.86      142.28
System         57.94        0.60
Elapsed       102.02       72.66

Note that almost a minutes worth of system time is eliminted and the
program completes 28% faster on average.

To illustrate the problem without python this is a basic test-case for
the worst case scenario where every free is a madvise followed by a an alloc

/* gcc bench-free.c -lpthread -o bench-free */
static int num = 1024;

void __attribute__((noinline,noclone)) dostuff (void *p)
{
}

void *worker (void *data)
{
  int i;

  for (i = num; i--;)
    {
      void *m = malloc (48*4096);
      dostuff (m);
      free (m);
    }

  return NULL;
}

int main()
{
  int i;
  pthread_t t;
  void *ret;
  if (pthread_create (&t, NULL, worker, NULL))
    exit (2);
  if (pthread_join (t, &ret))
    exit (3);
  return 0;
}

Before the patch, this resulted in 1024 calls to madvise. With the patch applied,
madvise is called twice because the default trim threshold is high enough to avoid
this.

This a more complex case where there is a mix of frees. It's simply a different worker
function for the test case above

void *worker (void *data)
{
  int i;
  int j = 0;
  void *free_index[num];

  for (i = num; i--;)
    {
      void *m = malloc ((i % 58) *4096);
      dostuff (m);
      if (i % 2 == 0) {
        free (m);
      } else {
        free_index[j++] = m;
      }
    }
  for (; j >= 0; j--)
    {
      free(free_index[j]);
    }

  return NULL;
}

glibc 2.21 calls malloc 90305 times but with the patch applied, it's
called 13438. Increasing the trim threshold will decrease the number of
times it's called with the option of eliminating the overhead.

ebizzy is meant to generate a workload resembling common web application
server workloads. It is threaded with a large working set that at its core
has an allocation, do_stuff, free loop that also hits this case. The primary
metric of the benchmark is records processed per second. This is running on
my desktop which is a single socket machine with an I7-4770 and 8 cores.
Each thread count was run for 30 seconds. It was only run once as the
performance difference is so high that the variation is insignificant.

                glibc 2.21              patch
threads 1            10230              44114
threads 2            19153              84925
threads 4            34295             134569
threads 8            51007             183387

Note that the saving happens to be a concidence as the size allocated
by ebizzy was less than the default threshold. If a different number of
chunks were specified then it may also be necessary to tune the threshold
to compensate

This is roughly quadrupling the performance of this benchmark. The difference in
system CPU usage illustrates why.

ebizzy running 1 thread with glibc 2.21
10230 records/s 306904
real 30.00 s
user  7.47 s
sys  22.49 s

22.49 seconds was spent in the kernel for a workload runinng 30 seconds. With the
patch applied

ebizzy running 1 thread with patch applied
44126 records/s 1323792
real 30.00 s
user 29.97 s
sys   0.00 s

system CPU usage was zero with the patch applied. strace shows that glibc
running this workload calls madvise approximately 9000 times a second. With
the patch applied madvise was called twice during the workload (or 0.06
times per second).

2015-02-10  Mel Gorman  <mgorman@suse.de>

  [BZ #17195]
  * malloc/arena.c (free): Apply trim threshold to per-thread heaps
    as well as the main arena.

1 files changed, 10 insertions, 3 deletions

diff --git a/malloc/arena.c b/malloc/arena.c
index 8af51f0..d85f371 100644
--- a/malloc/arena.c
+++ b/malloc/arena.c
@@ -658,7 +658,7 @@ heap_trim (heap_info *heap, size_t pad)
   unsigned long pagesz = GLRO (dl_pagesize);
   mchunkptr top_chunk = top (ar_ptr), p, bck, fwd;
   heap_info *prev_heap;
-  long new_size, top_size, extra, prev_size, misalign;
+  long new_size, top_size, top_area, extra, prev_size, misalign;
 
   /* Can this heap go away completely? */
   while (top_chunk == chunk_at_offset (heap, sizeof (*heap)))
@@ -694,9 +694,16 @@ heap_trim (heap_info *heap, size_t pad)
       set_head (top_chunk, new_size | PREV_INUSE);
       /*check_chunk(ar_ptr, top_chunk);*/
     }
+
+  /* Uses similar logic for per-thread arenas as the main arena with systrim
+     by preserving the top pad and at least a page.  */
   top_size = chunksize (top_chunk);
-  extra = (top_size - pad - MINSIZE - 1) & ~(pagesz - 1);
-  if (extra < (long) pagesz)
+  top_area = top_size - MINSIZE - 1;
+  if (top_area <= pad)
+    return 0;
+
+  extra = ALIGN_DOWN(top_area - pad, pagesz);
+  if ((unsigned long) extra < mp_.trim_threshold)
     return 0;
 
   /* Try to shrink. */