Implemented ThreadChunk limit.

diff --git a/lib/sorter/array.c b/lib/sorter/array.c
index 26b4a1b53048e1f11de89a653c36bcbdb513b292..d724f84128ed5a1bd14bd32e6bc68d0e40e52054 100644 (file)
--- a/lib/sorter/array.c
+++ b/lib/sorter/array.c
@@ -214,7 +214,8 @@ rs_finish(struct worker_thread *thr UNUSED, struct work *ww)
 {
   struct rs_work *w = (struct rs_work *) ww;
 
-  DBG("Thread %d: Finishing %d items, shift=%d", thr->id, w->num_elts, w->shift);
+  if (thr)
+    DBG("Thread %d: Finishing %d items, shift=%d", thr->id, w->num_elts, w->shift);
   if (w->shift < ASORT_MIN_SHIFT || w->num_elts < sorter_radix_threshold)
     {
       w->ctx->quicksort(w->in, w->num_elts);
@@ -223,7 +224,8 @@ rs_finish(struct worker_thread *thr UNUSED, struct work *ww)
     }
   else
     asort_radix(w->ctx, w->in, w->out, w->num_elts, w->shift, w->swap_output);
-  DBG("Thread %d: Finishing done", thr->id);
+  if (thr)
+    DBG("Thread %d: Finishing done", thr->id);
 }
 
 static void
@@ -298,6 +300,8 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
   for (uns i=0; i<buckets; i++)
     {
       uns n = cnt[i] - pos;
+      if (!n)
+       continue;
       if (n * ctx->elt_size < sorter_thread_threshold)
        {
          struct rs_work *w = ep_alloc(ctx->eltpool);
@@ -309,8 +313,17 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
          w->num_elts = n;
          w->shift = shift;
          w->swap_output = !swapped_output;
-         clist_add_tail(&ctx->rs_bits, &w->w.n);
-         DBG("Scheduling block %d+%d", pos, n);
+         if (n * ctx->elt_size < sorter_thread_chunk)
+           {
+             DBG("Sorting block %d+%d inline", pos, n);
+             rs_finish(NULL, &w->w);
+             ep_free(ctx->eltpool, w);
+           }
+         else
+           {
+             clist_add_tail(&ctx->rs_bits, &w->w.n);
+             DBG("Scheduling block %d+%d", pos, n);
+           }
        }
       else
        rs_radix(ctx, buffer, array, n, shift, !swapped_output);

diff --git a/lib/sorter/common.h b/lib/sorter/common.h
index f67f9bf9a1522c26fd9ff84db872d83ac61ade27..cd8053e11ce7b8d06737b24c8b3a342e54e096e1 100644 (file)
--- a/lib/sorter/common.h
+++ b/lib/sorter/common.h
@@ -16,7 +16,8 @@
 extern uns sorter_trace, sorter_stream_bufsize;
 extern uns sorter_debug, sorter_min_radix_bits, sorter_max_radix_bits;
 extern uns sorter_min_multiway_bits, sorter_max_multiway_bits;
-extern uns sorter_threads, sorter_thread_threshold, sorter_radix_threshold;
+extern uns sorter_threads, sorter_thread_threshold, sorter_thread_chunk;
+extern uns sorter_radix_threshold;
 extern u64 sorter_bufsize;
 extern struct fb_params sorter_fb_params;
 

diff --git a/lib/sorter/config.c b/lib/sorter/config.c
index 75d8fa49030876a8b4f3add48906c1a57b11426b..cc4bb3e66afb1ccd6f37974929179ad47dabb638 100644 (file)
--- a/lib/sorter/config.c
+++ b/lib/sorter/config.c
@@ -21,7 +21,8 @@ uns sorter_max_radix_bits;
 uns sorter_min_multiway_bits;
 uns sorter_max_multiway_bits;
 uns sorter_threads;
-uns sorter_thread_threshold;
+uns sorter_thread_threshold = 1048576;
+uns sorter_thread_chunk = 4096;
 uns sorter_radix_threshold = 4096;
 struct fb_params sorter_fb_params;
 
@@ -38,6 +39,7 @@ static struct cf_section sorter_config = {
     CF_UNS("MaxMultiwayBits", &sorter_max_multiway_bits),
     CF_UNS("Threads", &sorter_threads),
     CF_UNS("ThreadThreshold", &sorter_thread_threshold),
+    CF_UNS("ThreadChunk", &sorter_thread_chunk),
     CF_UNS("RadixThreshold", &sorter_radix_threshold),
     CF_END
   }