iopsystems/rezolus v5.28.0

on GitHub

Added

• drivehealth reads the EC2 NVMe detailed performance statistics (log page
  0xD0) from EBS and instance store volumes on Nitro instances:
  drive_performance_exceeded_time (time demand exceeded the volume's or
  instance's IOPS or throughput limit) and drive_queue_length. Not yet run
  against a real EBS or instance store device. Backported from 6.0 (#1455)
  in #1456.

Changed

• task_cpu_usage series are no longer keyed by thread ID. With
  task_attribution = true, the id label and x<n> series-name suffix
  are a slot index; select a thread by its pid label. (#1451)
• Reduced overhead when threads with neighbouring thread IDs run on
  different CPUs, by indexing per-thread BPF arrays through a permutation of
  the thread ID. (#1451)
• Reduced overhead by giving each CPU its own slice of every BPF histogram,
  at about 4 KiB of memory per possible CPU per histogram (about 42 MiB for
  all 58 at 192 CPUs). A read sums the slices: agent CPU over 30 s of
  /metrics/json reads every 50 ms went from 8570-8600 ms to 8880-8930 ms
  on 56 vCPUs. (#1451)
• Reduced overhead when many CPUs count the same cgroup, by giving each
  per-cgroup BPF counter 16 copies: 512 KiB per per-cgroup map instead of
  32 KiB, 23.5 MiB with every cgroup-counting sampler enabled (19.3 MB
  measured with the default samplers, against 3.5 MB). With cgroup
  attribution off the maps are not created. (#1452)
  These BPF changes were measured on kernel 6.12 and have not yet been loaded
  on 5.8-5.10 kernels.

Fixed

• cpu_l3 reports unsupported on a machine without L3 cache perf counters,
  such as a VM without a PMU, so rezolus status no longer exits 1 there. A
  permission refusal is also unsupported and names perf_event_paranoid,
  CAP_PERFMON and seccomp; any other open error is still a failure, now
  with its errno. (#1459)
• Corrected the swapped memory_numa_miss and memory_numa_foreign
  descriptions. The NUMA dashboard's remote plot now uses
  memory_numa_other, a Memory Policy subgroup plots hit, miss and
  interleave, and MCP anomaly detection analyses local, other and miss.
  (#1459)
• The agent's start-up log gives the reason for a sampler that is
  unsupported at init; it printed nothing after the dash. (#1459)