AcctIdx: when bg processing in multiple threads, avoid bin collisions (#20019)

runtime/src/accounts_index_storage.rs

@@ -96,6 +96,7 @@ impl<T: IndexValue> AccountsIndexStorage<T> {
         exit: Arc<AtomicBool>,
         in_mem: Vec<Arc<InMemAccountsIndex<T>>>,
     ) {
+        let bins = in_mem.len();
         loop {
             // this will transition to waits and thread throttling
             storage
@@ -106,11 +107,11 @@ impl<T: IndexValue> AccountsIndexStorage<T> {
             }
 
             storage.stats.active_threads.fetch_add(1, Ordering::Relaxed);
-            for bucket in &in_mem {
-                bucket.flush();
+            for _ in 0..bins {
+                let index = storage.next_bucket_to_flush();
+                in_mem[index].flush();
+                storage.stats.report_stats();
             }
-
-            storage.stats.report_stats();
             storage.stats.active_threads.fetch_sub(1, Ordering::Relaxed);
         }
     }