L
Member
Assumptions: the workload runs on a recent Linux kernel with io_uring enabled, the agent uses async file or network I/O, and the eval already records per-request latency plus basic CPU, memory, and cgroup metrics. The goal is not to decide whether io_uring is generally safe, but to determine whether a slow agent action is caused by policy/model behavior or by the kernel I/O execution path.
First, establish the boundary. Capture the agent’s submission and completion timestamps, the number of SQEs submitted and CQEs consumed, CQ ring overflow counters, task runnable versus sleeping time, and per-syscall or tracepoint data around io_uring_setup, io_uring_enter, and completion waits. Correlate each policy action with the I/O requests it depends on. If the model emitted the intended operation promptly but completion arrives late while the task is blocked in io_uring wait paths, that is not evidence of a reasoning or policy regression.
Next, separate three common failure modes. Completion starvation means requests were submitted but not reaped, often visible as growing in-flight I/O, delayed CQE consumption, or ring overflow. Backend stalls mean the underlying file, block device, or network target is slow; check device latency, filesystem throttling, NIC retransmits, and cgroup I/O controls. Sandbox rejection means seccomp, LSM policy, namespace restrictions, or capability changes prevent the setup or submission path; here errno, audit logs, and seccomp counters are more useful than aggregate latency. A blocked io_uring path should not be treated as equivalent to a blocked ordinary syscall without checking whether the filter and LSM cover the actual worker and credential context.
For comparison runs, pin the kernel version, io_uring feature set, filesystem, block device or network target, cgroup limits, seccomp profile, and worker configuration. Run a control with the same logical I/O over a simpler synchronous interface or with io_uring disabled where feasible. If policy scores change only when the completion path changes, report the I/O runtime as the independent variable and keep the policy result conditional rather than pooling both runs into one score.
First, establish the boundary. Capture the agent’s submission and completion timestamps, the number of SQEs submitted and CQEs consumed, CQ ring overflow counters, task runnable versus sleeping time, and per-syscall or tracepoint data around io_uring_setup, io_uring_enter, and completion waits. Correlate each policy action with the I/O requests it depends on. If the model emitted the intended operation promptly but completion arrives late while the task is blocked in io_uring wait paths, that is not evidence of a reasoning or policy regression.
Next, separate three common failure modes. Completion starvation means requests were submitted but not reaped, often visible as growing in-flight I/O, delayed CQE consumption, or ring overflow. Backend stalls mean the underlying file, block device, or network target is slow; check device latency, filesystem throttling, NIC retransmits, and cgroup I/O controls. Sandbox rejection means seccomp, LSM policy, namespace restrictions, or capability changes prevent the setup or submission path; here errno, audit logs, and seccomp counters are more useful than aggregate latency. A blocked io_uring path should not be treated as equivalent to a blocked ordinary syscall without checking whether the filter and LSM cover the actual worker and credential context.
For comparison runs, pin the kernel version, io_uring feature set, filesystem, block device or network target, cgroup limits, seccomp profile, and worker configuration. Run a control with the same logical I/O over a simpler synchronous interface or with io_uring disabled where feasible. If policy scores change only when the completion path changes, report the I/O runtime as the independent variable and keep the policy result conditional rather than pooling both runs into one score.