Array sorter thresholds are now configured as 64-bit numbers of bytes,

but they are recalculated to the number of elements internally with
proper clamping to ~0U.

diff --git a/lib/sorter/array.c b/lib/sorter/array.c
index e1805245ff4b1a609903fecca20120f7a8014416..a612e8f7841b66b9cd1c02fafc0089d72b15050d 100644 (file)
--- a/lib/sorter/array.c
+++ b/lib/sorter/array.c
@@ -51,7 +51,7 @@ asort_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts,
   for (uns i=0; i<buckets; i++)
     {
       uns n = cnt[i] - pos;
-      if (n * ctx->elt_size < sorter_radix_threshold || shift < ASORT_MIN_SHIFT)
+      if (n < ctx->radix_threshold || shift < ASORT_MIN_SHIFT)
        {
          ctx->quicksort(buffer, n);
          if (!swapped_output)
@@ -128,15 +128,16 @@ static void
 qs_handle_work(struct worker_thread *thr UNUSED, struct work *ww)
 {
   struct qs_work *w = (struct qs_work *) ww;
+  struct asort_context *ctx = w->ctx;
 
   DBG("Thread %d: got %d elts", thr->id, w->num_elts);
-  if (w->num_elts * w->ctx->elt_size < sorter_thread_threshold)
+  if (w->num_elts < ctx->thread_threshold)
     {
-      w->ctx->quicksort(w->array, w->num_elts);
+      ctx->quicksort(w->array, w->num_elts);
       w->left = w->right = LR_UNDEF;
     }
   else
-    w->ctx->quicksplit(w->array, w->num_elts, &w->left, &w->right);
+    ctx->quicksplit(w->array, w->num_elts, &w->left, &w->right);
   DBG("Thread %d: returning l=%d r=%d", thr->id, w->left, w->right);
 }
 
@@ -231,7 +232,7 @@ rs_finish(struct worker_thread *thr UNUSED, struct work *ww)
 
   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 * w->ctx->elt_size < sorter_radix_threshold)
+  if (w->shift < ASORT_MIN_SHIFT || w->num_elts < w->ctx->radix_threshold)
     {
       w->ctx->quicksort(w->array, w->num_elts);
       if (w->swap_output)
@@ -332,7 +333,7 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
       uns n = cnt[i] - pos;
       if (!n)
        continue;
-      if (n * ctx->elt_size < sorter_thread_threshold || shift < ASORT_MIN_SHIFT)
+      if (n < ctx->thread_threshold || shift < ASORT_MIN_SHIFT)
        {
          struct rs_work *w = ep_alloc(ctx->eltpool);
          w->w.priority = 0;
@@ -343,7 +344,7 @@ 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;
-         if (n * ctx->elt_size < sorter_thread_chunk)
+         if (n < ctx->thread_chunk)
            {
              DBG("Sorting block %d+%d inline", pos, n);
              rs_finish(NULL, &w->w);
@@ -409,7 +410,7 @@ predict_swap(struct asort_context *ctx)
   uns elts = ctx->num_elts;
   uns swap = 0;
 
-  while (elts * ctx->elt_size >= sorter_radix_threshold && bits >= ASORT_MIN_SHIFT)
+  while (elts >= ctx->radix_threshold && bits >= ASORT_MIN_SHIFT)
     {
       DBG("Predicting pass: %d elts, %d bits", elts, bits);
       swap = !swap;
@@ -422,12 +423,18 @@ predict_swap(struct asort_context *ctx)
 void
 asort_run(struct asort_context *ctx)
 {
+  ctx->thread_threshold = MIN(sorter_thread_threshold / ctx->elt_size, ~0U);
+  ctx->thread_chunk = MIN(sorter_thread_chunk / ctx->elt_size, ~0U);
+  ctx->radix_threshold = MIN(sorter_radix_threshold / ctx->elt_size, ~0U);
+
   ASORT_TRACE("Array-sorting %d items per %d bytes, hash_bits=%d", ctx->num_elts, ctx->elt_size, ctx->hash_bits);
+  ASORT_XTRACE(2, "Limits: thread_threshold=%u, thread_chunk=%u, radix_threshold=%u",
+       ctx->thread_threshold, ctx->thread_chunk, ctx->radix_threshold);
   uns allow_threads UNUSED = (sorter_threads > 1 &&
-                             ctx->num_elts * ctx->elt_size >= sorter_thread_threshold &&
+                             ctx->num_elts >= ctx->thread_threshold &&
                              !(sorter_debug & SORT_DEBUG_ASORT_NO_THREADS));
 
-  if (ctx->num_elts * ctx->elt_size < sorter_radix_threshold ||
+  if (ctx->num_elts < ctx->radix_threshold ||
       ctx->hash_bits <= ASORT_MIN_SHIFT ||
       !ctx->radix_split ||
       (sorter_debug & SORT_DEBUG_ASORT_NO_RADIX))

diff --git a/lib/sorter/common.h b/lib/sorter/common.h
index bd24376613a32553979b9402f9b6378c1e46058b..ddd7ba6fadb407ba46803961dca92e1d61e07683 100644 (file)
--- a/lib/sorter/common.h
+++ b/lib/sorter/common.h
@@ -16,9 +16,9 @@
 extern uns sorter_trace, sorter_trace_array, sorter_stream_bufsize;
 extern uns sorter_debug, sorter_min_radix_bits, sorter_max_radix_bits, sorter_add_radix_bits;
 extern uns sorter_min_multiway_bits, sorter_max_multiway_bits;
-extern uns sorter_threads, sorter_thread_threshold, sorter_thread_chunk;
-extern uns sorter_radix_threshold;
+extern uns sorter_threads;
 extern u64 sorter_bufsize, sorter_small_input;
+extern u64 sorter_thread_threshold, sorter_thread_chunk, sorter_radix_threshold;
 extern struct fb_params sorter_fb_params, sorter_small_fb_params;
 
 #define SORT_TRACE(x...) do { if (sorter_trace) msg(L_DEBUG, x); } while(0)
@@ -122,6 +122,7 @@ void sbuck_swap_out(struct sort_bucket *b);
 /* Contexts and helper functions for the array sorter */
 
 struct asort_context {
+  // Interface between generic code in array.c and functions generated by array.h
   void *array;                         // Array to sort
   void *buffer;                                // Auxiliary buffer (required when radix-sorting)
   uns num_elts;                                // Number of elements in the array
@@ -137,6 +138,11 @@ struct asort_context {
   struct rs_work **rs_works;
   struct work_queue *rs_work_queue;
   struct eltpool *eltpool;
+
+  // Configured limits translated from bytes to elements
+  uns thread_threshold;
+  uns thread_chunk;
+  uns radix_threshold;
 };
 
 void asort_run(struct asort_context *ctx);

diff --git a/lib/sorter/config.c b/lib/sorter/config.c
index 5e3da8dd9006155818f6d07e0ecdcd7360d6d838..9ff646b2c45f1bd87c97af71804e6b8685ce4668 100644 (file)
--- a/lib/sorter/config.c
+++ b/lib/sorter/config.c
@@ -22,9 +22,9 @@ uns sorter_add_radix_bits;
 uns sorter_min_multiway_bits;
 uns sorter_max_multiway_bits;
 uns sorter_threads;
-uns sorter_thread_threshold = 1048576;
-uns sorter_thread_chunk = 4096;
-uns sorter_radix_threshold = 4096;
+u64 sorter_thread_threshold = 1048576;
+u64 sorter_thread_chunk = 4096;
+u64 sorter_radix_threshold = 4096;
 struct fb_params sorter_fb_params;
 struct fb_params sorter_small_fb_params;
 u64 sorter_small_input;
@@ -44,9 +44,9 @@ static struct cf_section sorter_config = {
     CF_UNS("MinMultiwayBits", &sorter_min_multiway_bits),
     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_U64("ThreadThreshold", &sorter_thread_threshold),
+    CF_U64("ThreadChunk", &sorter_thread_chunk),
+    CF_U64("RadixThreshold", &sorter_radix_threshold),
     CF_END
   }
 };