CPU multi threading doesn't work on Intel's hybrid core architecture

[DDoSolitary]

What happened?

Running iree-benchmark-module on i7-13700 shows that wall time equals cpu time, indicating only a single thread is used.

> iree-benchmark-module --module=model.vmfb --device=local-task --function=torch-jit-export --input=@in.npy
2026-07-26T22:48:03+08:00
Running /data/models/.venv/lib/python3.13/site-packages/iree/_runtime_libs/iree-benchmark-module
Run on (24 X 5100 MHz CPU s)
CPU Caches:
  L1 Data 48 KiB (x12)
  L1 Instruction 32 KiB (x12)
  L2 Unified 2048 KiB (x12)
  L3 Unified 30720 KiB (x1)
Load Average: 0.19, 0.13, 0.15
***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
***WARNING*** ASLR is enabled, the results may have unreproducible noise in them.
-----------------------------------------------------------------------------------------------------
Benchmark                                           Time             CPU   Iterations UserCounters...
-----------------------------------------------------------------------------------------------------
BM_torch-jit-export/process_time/real_time        422 ms          422 ms            2 items_per_second=2.36932/s

While on an older platform multi threading works as expected:

Working Directory: /data/models/build/models/example
Test timeout computed to be: 10000000
2026-07-26T08:05:47+00:00
Running /data/models/.venv/lib/python3.13/site-packages/iree/_runtime_libs/iree-benchmark-module
Run on (16 X 2593.91 MHz CPU s)
CPU Caches:
  L1 Data 48 KiB (x16)
  L1 Instruction 32 KiB (x16)
  L2 Unified 1280 KiB (x16)
  L3 Unified 43008 KiB (x1)
Load Average: 0.43, 0.27, 0.20
***WARNING*** ASLR is enabled, the results may have unreproducible noise in them.
-----------------------------------------------------------------------------------------------------
Benchmark                                           Time             CPU   Iterations UserCounters...
-----------------------------------------------------------------------------------------------------
BM_torch-jit-export/process_time/real_time        111 ms         1182 ms            6 items_per_second=8.99493/s

Root-cause

1. The sysfs-based topology detection code incorrectly uses cluster_id as a synonym of NUMA node id/isolates the clusters
   Exacerbated by
2. By default, only the "current" NUMA node is utilized (--task_topology_nodes defaults to "current" vs. "all")
3. cluster_ids are not continuous, leading to a latent failure in enumeration (task_topology_nodes "all" cannot be used as a workaround)

The sysfs-based topology detection

iree/runtime/src/iree/task/topology_sysfs.c

Lines 294 to 303 in 16396e4

	for (uint32_t cpu = 0; cpu < processor_count; ++cpu) {
	uint32_t cluster_id = 0;
	if (iree_sysfs_try_query_cluster_id(cpu, &cluster_id) &&
	iree_sysfs_is_valid_cluster(cluster_id)) {
	if (!CPU_ISSET(cluster_id, &cluster_set)) {
	CPU_SET(cluster_id, &cluster_set);
	++unique_clusters;
	}
	}
	}

For example, on a i7-13700 there's a total of 10 clusters with 8 P-cores = 8 clusters and 8 E-cores = 2 clusters:
every P-core forms its own cluster while groups of E-cores form other clusters

> grep . /sys/devices/system/cpu/cpu*/topology/cluster_id | sort -t: -k2 -n                                                                                                                                                                              
/sys/devices/system/cpu/cpu0/topology/cluster_id:0
/sys/devices/system/cpu/cpu1/topology/cluster_id:0
/sys/devices/system/cpu/cpu2/topology/cluster_id:8
/sys/devices/system/cpu/cpu3/topology/cluster_id:8
/sys/devices/system/cpu/cpu4/topology/cluster_id:16
/sys/devices/system/cpu/cpu5/topology/cluster_id:16
/sys/devices/system/cpu/cpu6/topology/cluster_id:24
/sys/devices/system/cpu/cpu7/topology/cluster_id:24
/sys/devices/system/cpu/cpu8/topology/cluster_id:32
/sys/devices/system/cpu/cpu9/topology/cluster_id:32
/sys/devices/system/cpu/cpu10/topology/cluster_id:40
/sys/devices/system/cpu/cpu11/topology/cluster_id:40
/sys/devices/system/cpu/cpu12/topology/cluster_id:48
/sys/devices/system/cpu/cpu13/topology/cluster_id:48
/sys/devices/system/cpu/cpu14/topology/cluster_id:56
/sys/devices/system/cpu/cpu15/topology/cluster_id:56
/sys/devices/system/cpu/cpu16/topology/cluster_id:64
/sys/devices/system/cpu/cpu17/topology/cluster_id:64
/sys/devices/system/cpu/cpu18/topology/cluster_id:64
/sys/devices/system/cpu/cpu19/topology/cluster_id:64
/sys/devices/system/cpu/cpu20/topology/cluster_id:72
/sys/devices/system/cpu/cpu21/topology/cluster_id:72
/sys/devices/system/cpu/cpu22/topology/cluster_id:72
/sys/devices/system/cpu/cpu23/topology/cluster_id:72

Removing the cluster_id code and fallback to physical_package_id, performance is immediately back to normal:

diff --git a/runtime/src/iree/task/topology_sysfs.c b/runtime/src/iree/task/topology_sysfs.c
index 083f014408..dfd00c1d76 100644
--- a/runtime/src/iree/task/topology_sysfs.c
+++ b/runtime/src/iree/task/topology_sysfs.c
@@ -141,11 +141,11 @@ static bool iree_sysfs_try_query_cluster_id(uint32_t processor,
   char path[256];

   // Try cluster_id first (kernel 5.16+).
-  iree_snprintf(path, sizeof(path), "%s/cpu/cpu%u/topology/cluster_id",
-                iree_sysfs_get_root_path(), processor);
-  if (iree_sysfs_try_read_uint32(path, out_cluster_id)) {
-    return true;
-  }
+  // iree_snprintf(path, sizeof(path), "%s/cpu/cpu%u/topology/cluster_id",
+  //               iree_sysfs_get_root_path(), processor);
+  // if (iree_sysfs_try_read_uint32(path, out_cluster_id)) {
+  //   return true;
+  // }

   // Fallback to physical_package_id (socket/package).
   iree_snprintf(path, sizeof(path), "%s/cpu/cpu%u/topology/physical_package_id",

2026-07-26T23:11:20+08:00
Running ./build/tools/iree-benchmark-module
Run on (24 X 5100 MHz CPU s)
CPU Caches:
  L1 Data 48 KiB (x12)
  L1 Instruction 32 KiB (x12)
  L2 Unified 2048 KiB (x12)
  L3 Unified 30720 KiB (x1)
Load Average: 4.91, 1.29, 0.52
***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
***WARNING*** ASLR is enabled, the results may have unreproducible noise in them.
-----------------------------------------------------------------------------------------------------
Benchmark                                           Time             CPU   Iterations UserCounters...
-----------------------------------------------------------------------------------------------------
BM_torch-jit-export/process_time/real_time       63.9 ms          709 ms            9 items_per_second=15.6597/s

Note: physical_package_id is not NUMA node id either, there is /sys/devices/system/node/node*/cpumap.

#22455 #8469

What component(s) does this issue relate to?

Runtime

Version information

compiler and runtime are manually built on git commit bc908fd

[github-actions]

Hello @DDoSolitary 👋

Thank you for submitting an issue to IREE! It looks like it's your first one.

The volume of issues is huge, and triage times may vary. If you need quicker assistance on the issue, consider posting it on IREE Discord and providing short context there.
Feel free to identify the most frequent contributors to the corresponding part of the codebase via Git history/CODEOWNERS and (gently!) tag them below.

---

Action required: acknowledge IREE project policies

IREE is a Linux Foundation project. All participants are expected to follow the LF Projects Code of Conduct.

Please note that all contributions to IREE, including GitHub issues, must follow our IREE AI Tool Use Policy. This means:

• Issue reporters must fully understand the primary issue and the impact of the problem.
  For bug reports, specifics of the use-case, workflow and/or crash scenario usually suffice.
  For improvement suggestions, this usually implies deeper understanding of suboptimal code, IR design choices, assumed performance tradeoffs, etc.
• Contributors must write the summary of the issue themselves, so that the exact impact and scenario is clear to other participants.
• Issue authors are expected to review and understand all LLM/AI-assisted segments of the report, if any. E.g.: CLI commands and IR reproducers, model/framework details, lowering pipeline details, assumptions about problematic behaviors of IREE components.
• Substantial use of LLM/generative AI tools must be disclaimed in the header or footer of the issue description.

Before the issue can be triaged, we kindly ask you to reply to this message and confirm that you understand and accept the proposed policy. As required, please adjust your initial issue description.

---

If you have any questions, feel free to leave a comment here, or ask away on IREE Discord.

Thank you,
The IREE Community

[devtbi]

thanks for summarizing the discord thread (https://discord.com/channels/689900678990135345/1529430463004020757) :) . I've edited some details and changed the language to be more neutral.