@@ -19,6 +19,7 @@ Monitoring Plugins:
 Monitoring Plugins:
 
 * disk-io: iowait is now reported as saturated CPU cores (e.g. `1.02 cores`) instead of a percentage that could exceed 100% on multi-core hosts, and gets its own Grafana panel; the alerting thresholds keep their meaning and defaults, so no reconfiguration is needed
+* disk-io: per-disk performance data now reports read/write throughput per second and disk busy percentage as ready-to-graph values instead of raw cumulative counters; re-import the disk-io Grafana dashboard after updating ([#320](https://github.com/Linuxfabrik/monitoring-plugins/issues/320))
 
 ### Fixed
 
 
@@ -7,7 +7,7 @@ Checks disk I/O bandwidth over time and alerts on sustained saturation, not shor
 
 On Linux, the check also monitors how much I/O wait the system accumulates (CPU time spent waiting for I/O). The machine-wide iowait reported by the kernel is normalized to a per-core basis and shown as *saturated cores*: `1.0 cores` means the equivalent of one CPU core spent the whole interval waiting for I/O, `2.0 cores` means two, and so on. This is the same mental model as the load average. Reporting per-core instead of as a percentage of the whole machine is deliberate: it lets you deploy one identical threshold to an entire fleet. `--iowait-warning` and `--iowait-critical` are expressed in percent of a single core (default 80/90), so `90` means "alert when 1 core is stuck 90% in I/O wait", and that means the same thing on a 4-core VM and a 64-core VM, even when the core count changes over time. Like bandwidth alerts, iowait alerts require COUNT consecutive threshold violations.
 
-Perfdata is emitted for each disk (busy_time, read_bytes, read_time, write_bytes, write_time) and for iowait, so you can graph trends. On Linux the check focuses on block devices with a mounted filesystem by default; use `--include-unmounted` to also include raw, unmounted devices such as multipath SAN volumes. On Windows it uses psutil's disk counters. Optionally, `--top` lists the processes that generated the most I/O traffic (read/write totals) to help identify offenders.
+Perfdata is emitted for each disk (read/write throughput per second and disk busy percentage) and for iowait, so you can graph trends. On Linux the check focuses on block devices with a mounted filesystem by default; use `--include-unmounted` to also include raw, unmounted devices such as multipath SAN volumes. On Windows it uses psutil's disk counters. Optionally, `--top` lists the processes that generated the most I/O traffic (read/write totals) to help identify offenders.
 
 This check is cross-platform and works on Linux, Windows, and all psutil-supported systems. The check stores its short trend state locally in an SQLite DB to evaluate sustained load across runs.
 
@@ -67,8 +67,8 @@ fleet-wide threshold works on every host regardless of its core count:
 --iowait-warning and --iowait-critical are given in percent of one core
 (default 80/90, i.e. alert near one saturated core). Like bandwidth alerts,
 iowait alerts require COUNT consecutive threshold violations. Perfdata is
-emitted for each disk (busy_time, read_bytes, read_time, write_bytes,
-write_time) and for iowait, so you can graph trends. On Linux the check
+emitted for each disk (read/write throughput per second and disk busy
+percentage) and for iowait, so you can graph trends. On Linux the check
 focuses on block devices with a mounted filesystem by default; use `--include-
 unmounted` to also include raw, unmounted devices such as multipath SAN
 volumes. On Windows it uses psutil's disk counters. Optionally, `--top` lists
@@ -200,15 +200,17 @@ Global:
 |----|----|----|
 | iowait | Number | Machine-wide I/O wait normalized to saturated CPU cores (`1.0` = one core fully waiting for I/O). Linux only. |
 
-Per matched disk, where `<disk>` is the block device name:
+Per matched disk, where `<disk>` is the block device name. The `1` suffix is the latest interval (like a "load1"), the `15` suffix is the average over the last `--count` runs (like a "load15"):
 
 | Name | Type | Description |
 |----|----|----|
-| `<disk>`\_busy_time | Continuous Counter | Time spent doing actual I/Os (in milliseconds). |
-| `<disk>`\_read_bytes | Continuous Counter | Number of bytes read. |
-| `<disk>`\_read_time | Continuous Counter | Time spent reading from disk (in milliseconds). |
-| `<disk>`\_write_bytes | Continuous Counter | Number of bytes written. |
-| `<disk>`\_write_time | Continuous Counter | Time spent writing to disk (in milliseconds). |
+| `<disk>`\_busy_percent | Percentage | Share of wall-clock time the device was busy with I/O over the last interval (iostat's %util). |
+| `<disk>`\_read_bytes_per_second1 | Bytes | Bytes read per second over the latest interval. |
+| `<disk>`\_read_bytes_per_second15 | Bytes | Bytes read per second, averaged over the last `--count` runs. |
+| `<disk>`\_throughput1 | Bytes | Read plus write bytes per second over the latest interval. |
+| `<disk>`\_throughput15 | Bytes | Read plus write bytes per second, averaged over the last `--count` runs. |
+| `<disk>`\_write_bytes_per_second1 | Bytes | Bytes written per second over the latest interval. |
+| `<disk>`\_write_bytes_per_second15 | Bytes | Bytes written per second, averaged over the last `--count` runs. |
 
 
 ## Troubleshooting
 
@@ -34,7 +34,7 @@ except ImportError:
 
 
 __author__ = 'Linuxfabrik GmbH, Zurich/Switzerland'
-__version__ = '2026061801'
+__version__ = '2026061802'
 
 DESCRIPTION = """Checks disk I/O bandwidth over time and alerts on sustained saturation, not
 short spikes. The check records per-disk read/write counters and then derives current (R1/W1)
@@ -50,7 +50,7 @@ host regardless of its core count: --iowait-warning and --iowait-critical are gi
 of one core (default 80/90, i.e. alert near one saturated core). Like bandwidth alerts, iowait
 alerts require COUNT consecutive threshold violations.
 
-Perfdata is emitted for each disk (busy_time, read_bytes, read_time, write_bytes, write_time)
+Perfdata is emitted for each disk (read/write throughput per second and disk busy percentage)
 and for iowait, so you can graph trends. On Linux the check focuses on block devices with a
 mounted filesystem by default; use `--include-unmounted` to also include raw, unmounted devices
 such as multipath SAN volumes. On Windows it uses psutil's disk counters. Optionally, `--top`
@@ -610,7 +610,7 @@ def main():
             lib.base.oao('Waiting for more data.', state)
 
         # calculate current rates (like "load1")
-        timediff, read_bytes_per_second1, write_bytes_per_second1, bandwidth1 = (
+        timediff1, read_bytes_per_second1, write_bytes_per_second1, bandwidth1 = (
             get_rate(
                 data[0]['timestamp'],
                 data[1]['timestamp'],
@@ -620,7 +620,7 @@ def main():
                 data[1]['write_bytes'],
             )
         )
-        if timediff <= 0:  # often happens after a reboot
+        if timediff1 <= 0:  # often happens after a reboot
             lib.db_sqlite.close(conn)
             lib.base.oao('Waiting for more data.', state)
 
@@ -644,7 +644,7 @@ def main():
         if len(data) != args.COUNT:  # not enough data yet
             continue
 
-        timediff, read_bytes_per_second15, write_bytes_per_second15, bandwidth15 = (
+        timediff15, read_bytes_per_second15, write_bytes_per_second15, bandwidth15 = (
             get_rate(
                 data[0]['timestamp'],
                 data[args.COUNT - 1]['timestamp'],
@@ -654,7 +654,7 @@ def main():
                 data[args.COUNT - 1]['write_bytes'],
             )
         )
-        if timediff <= 0:  # often happens after a reboot
+        if timediff15 <= 0:  # often happens after a reboot
             lib.db_sqlite.close(conn)
             lib.base.oao('Waiting for more data.', state)
 
@@ -682,55 +682,66 @@ def main():
             }
         )
 
-        # perfdata
-        try:
-            perfdata += lib.base.get_perfdata(
-                f'{bd}_busy_time',
-                data[0]['busy_time'],
-                uom='c',
-                warn=None,
-                crit=None,
-                _min=0,
-                _max=None,
-            )
-            perfdata += lib.base.get_perfdata(
-                f'{bd}_read_bytes',
-                data[0]['read_bytes'],
-                uom='c',
-                warn=None,
-                crit=None,
-                _min=0,
-                _max=None,
-            )
-            perfdata += lib.base.get_perfdata(
-                f'{bd}_read_time',
-                data[0]['read_time'],
-                uom='c',
-                warn=None,
-                crit=None,
-                _min=0,
-                _max=None,
-            )
-            perfdata += lib.base.get_perfdata(
-                f'{bd}_write_bytes',
-                data[0]['write_bytes'],
-                uom='c',
-                warn=None,
-                crit=None,
-                _min=0,
-                _max=None,
-            )
-            perfdata += lib.base.get_perfdata(
-                f'{bd}_write_time',
-                data[0]['write_time'],
-                uom='c',
-                warn=None,
-                crit=None,
-                _min=0,
-                _max=None,
-            )
-        except (KeyError, TypeError):
-            pass
+        # disk utilization (iostat's %util): share of wall-clock time the
+        # device was busy with I/O over the last interval. busy_time is a
+        # cumulative millisecond counter, so derive the value from the delta
+        # here rather than emitting the raw counter (see CONTRIBUTING: no
+        # continuous counters in perfdata).
+        busy_delta = data[0]['busy_time'] - data[1]['busy_time']
+        busy_percent = round(
+            max(0.0, min(100.0, busy_delta / (timediff1 * 1000) * 100)),
+            1,
+        )
+
+        # perfdata: emit per-second rates and utilization as gauges, never the
+        # raw cumulative counters. "throughput" is read + write per second; its
+        # _15 average carries the per-disk alerting threshold (a percentage of
+        # the observed maximum bandwidth).
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_read_bytes_per_second1',
+            read_bytes_per_second1,
+            uom='B',
+            _min=0,
+        )
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_read_bytes_per_second15',
+            read_bytes_per_second15,
+            uom='B',
+            _min=0,
+        )
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_write_bytes_per_second1',
+            write_bytes_per_second1,
+            uom='B',
+            _min=0,
+        )
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_write_bytes_per_second15',
+            write_bytes_per_second15,
+            uom='B',
+            _min=0,
+        )
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_throughput1',
+            bandwidth1,
+            uom='B',
+            _min=0,
+        )
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_throughput15',
+            bandwidth15,
+            uom='B',
+            warn=int(bandwidth_max * args.WARN / 100),
+            crit=int(bandwidth_max * args.CRIT / 100),
+            _min=0,
+        )
+        perfdata += lib.base.get_perfdata(
+            f'{bd}_busy_percent',
+            busy_percent,
+            uom='%',
+            _min=0,
+            _max=100,
+        )
 
     # check iowait alerting (only if enough consecutive values collected)
     iowait_state = STATE_OK