hot path optimizations in uma_zalloc() & uma_zfree()

[ant]

Hi folks.

I just tryed to make buckets management in perCPU cache like in
Solaris (see paper of Jeff Bonwick - Magazines and Vmem)
and got perfomance gain around 10% in my test program.
Then i made another minor code optimization and got another 10%.
The program just creates and destroys sockets in loop.

I suppose the reason of first gain lies in increasing of cpu cache hits.
In current fbsd code allocations and freeings deal with
separate buckets. Buckets are changed when one of them
became full or empty first. In Solaris this work is pure LIFO:
i.e. alloc() and free() work with one bucket - the current bucket
(it is called magazine there), that's why cache hit rate is bigger.

Another optimization is very trivial, for example:
-   bucket->ub_cnt--;
-   item = bucket->ub_bucket[bucket->ub_cnt];
+   item = bucket->ub_bucket[--bucket->ub_cnt];
(see the patch)


The test program:

#include <unistd.h>
#include <sys/socket.h>

main(int argc, char *argv[])
{
 int *fd, n, i,j, iters=100000;

 n = atoi(argv[1]);
 fd = (int*) malloc(sizeof(*fd) * n);

 iters /= n;
 for (i=0; i<iters; i++) {
  for (j=0; j<n; j++)
   fd[j] = socket(AF_UNIX, SOCK_STREAM, 0);
  for (j=0; j<n; j++)
   close(fd[j]);
 }
}



The results with current uma_core.c

> time ./sockloop 1                                    # first arg is the
number of sockets that created in one iteration
0.093u 2.650s 0:02.75 99.6%     5+180k 0+0io 0pf+0w
> time ./sockloop 1
0.108u 2.298s 0:02.41 99.1%     5+181k 0+0io 0pf+0w
> time ./sockloop 1
0.127u 2.278s 0:02.41 99.1%     5+177k 0+0io 0pf+0w
> time ./sockloop 10                                # number of iterations
is changed according to arg (see code)
0.054u 2.239s 0:02.30 99.1%     5+181k 0+0io 0pf+0w
> time ./sockloop 10
0.069u 2.199s 0:02.27 99.1%     6+184k 0+0io 0pf+0w
> time ./sockloop 10
0.086u 2.185s 0:02.28 99.1%     5+178k 0+0io 0pf+0w
> time ./sockloop 100
0.101u 2.393s 0:02.51 99.2%     5+179k 0+0io 0pf+0w
> time ./sockloop 100
0.085u 2.505s 0:02.60 99.2%     5+180k 0+0io 0pf+0w
> time ./sockloop 100
0.054u 2.441s 0:02.50 99.6%     5+178k 0+0io 0pf+0w
> time ./sockloop 1000
0.093u 2.739s 0:02.84 99.2%     5+181k 0+0io 0pf+0w
> time ./sockloop 1000
0.085u 2.797s 0:02.89 99.3%     5+180k 0+0io 0pf+0w
> time ./sockloop 1000
0.117u 2.689s 0:02.82 98.9%     5+179k 0+0io 0pf+0w


The results of first optimization (only buckets management)

> time ./sockloop 1
0.125u 1.938s 0:02.06 99.5%     5+180k 0+0io 0pf+0w
> time ./sockloop 1
0.070u 1.993s 0:02.06 100.0%    5+180k 0+0io 0pf+0w
> time ./sockloop 1
0.110u 1.953s 0:02.06 100.0%    5+177k 0+0io 0pf+0w
> time ./sockloop 10
0.093u 1.776s 0:01.87 99.4%     5+180k 0+0io 0pf+0w
> time ./sockloop 10
0.116u 1.754s 0:01.87 99.4%     5+181k 0+0io 0pf+0w
> time ./sockloop 10
0.093u 1.777s 0:01.87 99.4%     5+181k 0+0io 0pf+0w
> time ./sockloop 100
0.100u 2.182s 0:02.29 99.5%     5+180k 0+0io 0pf+0w
> time ./sockloop 100
0.093u 2.174s 0:02.27 99.5%     5+180k 0+0io 0pf+0w
> time ./sockloop 100
0.078u 2.158s 0:02.24 99.1%     5+180k 0+0io 0pf+0w
> time ./sockloop 1000
0.101u 2.403s 0:02.51 99.6%     5+180k 0+0io 0pf+0w
> time ./sockloop 1000
0.124u 2.381s 0:02.52 99.2%     5+180k 0+0io 0pf+0w
> time ./sockloop 1000
0.125u 2.373s 0:02.51 99.2%     5+178k 0+0io 0pf+0w



The results of both optimizations

> time ./sockloop 1
0.062u 1.785s 0:01.85 99.4%     5+180k 0+0io 0pf+0w
> time ./sockloop 1
0.124u 1.722s 0:01.85 99.4%     5+180k 0+0io 0pf+0w
> time ./sockloop 1
0.087u 1.759s 0:01.85 98.9%     5+177k 0+0io 0pf+0w
> time ./sockloop 10
0.069u 1.684s 0:01.75 99.4%     5+181k 0+0io 0pf+0w
> time ./sockloop 10
0.070u 1.673s 0:01.74 100.0%    5+180k 0+0io 0pf+0w
> time ./sockloop 10
0.070u 1.672s 0:01.74 100.0%    5+177k 0+0io 0pf+0w
> time ./sockloop 100
0.077u 2.102s 0:02.18 99.5%     5+180k 0+0io 0pf+0w
> time ./sockloop 100
0.116u 2.062s 0:02.18 99.5%     5+180k 0+0io 0pf+0w
> time ./sockloop 100
0.055u 2.126s 0:02.19 99.0%     5+178k 0+0io 0pf+0w
> time ./sockloop 1000
0.077u 2.298s 0:02.39 98.7%     5+181k 0+0io 0pf+0w
> time ./sockloop 1000
0.070u 2.340s 0:02.42 99.5%     5+178k 0+0io 0pf+0w
> time ./sockloop 1000
0.054u 2.320s 0:02.39 99.1%     5+179k 0+0io 0pf+0w


the patch is for uma_core.c from RELENG_5, but i checked
uma_core.c in CURRENT - it's the same regarding to thiese
improvements. I don't have any commit rights, so the patch
is just for reviewing. Here it is:

--- sys/vm/uma_core.c.orig Wed Jun 29 21:46:52 2005
+++ sys/vm/uma_core.c Wed Jun 29 23:09:32 2005
@@ -1830,8 +1830,7 @@

  if (bucket) {
   if (bucket->ub_cnt > 0) {
-   bucket->ub_cnt--;
-   item = bucket->ub_bucket[bucket->ub_cnt];
+   item = bucket->ub_bucket[--bucket->ub_cnt];
 #ifdef INVARIANTS
    bucket->ub_bucket[bucket->ub_cnt] = NULL;
 #endif
@@ -2252,7 +2251,7 @@
  cache = &zone->uz_cpu[cpu];

 zfree_start:
- bucket = cache->uc_freebucket;
+ bucket = cache->uc_allocbucket;

  if (bucket) {
   /*
@@ -2263,8 +2262,7 @@
   if (bucket->ub_cnt < bucket->ub_entries) {
    KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
        ("uma_zfree: Freeing to non free bucket index."));
-   bucket->ub_bucket[bucket->ub_cnt] = item;
-   bucket->ub_cnt++;
+   bucket->ub_bucket[bucket->ub_cnt++] = item;
 #ifdef INVARIANTS
    ZONE_LOCK(zone);
    if (keg->uk_flags & UMA_ZONE_MALLOC)
@@ -2275,7 +2273,7 @@
 #endif
    CPU_UNLOCK(cpu);
    return;
-  } else if (cache->uc_allocbucket) {
+  } else if (cache->uc_freebucket) {
 #ifdef UMA_DEBUG_ALLOC
    printf("uma_zfree: Swapping buckets.\n");
 #endif
@@ -2283,8 +2281,7 @@
     * We have run out of space in our freebucket.
     * See if we can switch with our alloc bucket.
     */
-   if (cache->uc_allocbucket->ub_cnt <
-       cache->uc_freebucket->ub_cnt) {
+   if (cache->uc_freebucket->ub_cnt == 0) {
     bucket = cache->uc_freebucket;
     cache->uc_freebucket = cache->uc_allocbucket;
     cache->uc_allocbucket = bucket;


if one will decide to commit first optimization (about buckets),
then there must some adjustments be made also
regarding correct statistics gathering.

Regards,
  Andriy Tkachuk.