perftune.py: work in progress scylladb#2

scripts/perftune.py

@@ -258,7 +258,7 @@ def auto_detect_irq_mask(cpu_mask):
     PUs_key = 'PUs'
 
     # List of NUMA IDs that own CPUs from the given CPU mask
-    numa_ids_list = run_hwloc_calc(['-I', 'numa', cpu_mask]).split(",")
+    numa_ids_list = [int(n) for n in run_hwloc_calc(['-I', 'numa', cpu_mask]).split(",")]
 
     # Let's calculate number of HTs and cores on each NUMA node belonging to the given CPU set
     cores_PUs_per_numa = {} # { <numa_id> : {'cores':  <number of cores>, 'PUs': <number of PUs>}}
@@ -293,20 +293,17 @@ def auto_detect_irq_mask(cpu_mask):
         # Let's allocate a full core for each 32 HTs.
         # Let's distribute IRQ cores among present NUMA nodes
         num_irq_cores = math.ceil(num_PUs / 32)
-        num_cores_per_numa = max(1, num_irq_cores // len(numa_ids_list))
         hwloc_args = []
+        numa_cores_count = {n: 0 for n in numa_ids_list}
         added_cores = 0
-        for numa in numa_ids_list:
-            for i in range(num_cores_per_numa):
-                hwloc_args.append(f"node:{numa}.core:{i}")
+        while added_cores < num_irq_cores:
+            for numa in numa_ids_list:
+                hwloc_args.append(f"node:{numa}.core:{numa_cores_count[numa]}")
                 added_cores += 1
+                numa_cores_count[numa] += 1
 
                 if added_cores >= num_irq_cores:
                     break
-            else:
-                continue
-            # if the inner loop exited on break - get out
-            break
 
         return run_hwloc_calc(['--restrict', cpu_mask] + hwloc_args)