AWS EBS CloudWatch Metrics: Monitoring Guide

Introduction

An EBS volume rarely fails loudly. It slows down, queues I/O, burns through burst credits, or sits idle while you pay for it.

Watch IOPS, throughput, latency, queue length, and provisioned limits separately, and you'll miss the pattern. Lucidity's State of Cloud Block Storage research puts average enterprise disk utilization at 30%, and 8-10% of disks lead to downtime.

This guide is for AWS infrastructure, ITOps, DevOps, and FinOps teams monitoring volumes through the AWS/EBS namespace. It covers the core metrics, dashboard and alarm setup, TM90 versus P90, pricing and retention, and troubleshooting workflows.

Key Takeaways

  • AWS/EBS metrics give per-volume visibility that EC2 and CloudWatch Agent metrics don't.
  • Read IOPS, throughput, latency, queue length, and burst or idle signals together.
  • Alarm on sustained breaches against each volume's provisioned limits, not isolated spikes.
  • Native publication is free; dashboards, alarms, and API calls are not.

What AWS EBS CloudWatch Metrics Measure

What are AWS EBS CloudWatch metrics?

AWS/EBS metrics are published by Amazon EBS for individual volumes, selected by the VolumeId dimension. Certain Multi-Attach metrics also support InstanceId. The separate AWS/EC2 namespace holds instance-level metrics, including instance-side EBS figures.

Native metrics reveal I/O volume, latency, saturation, and burst behavior. They don't show filesystem free space or database health. For free space, use the CloudWatch Agent, which publishes disk_used_percent and related metrics to the CWAgent namespace by default.

How to identify the correct volume and metric scope

Metrics are scoped by namespace, metric name, and dimensions, and live in one Region at a time. Before graphing, confirm three things:

  • The console Region matches the volume's Region
  • The account is the one that owns the volume
  • The volume ID maps to the right environment and owner

Eligibility for newer metrics varies. All volume types publish 1-minute metrics while attached, but average IOPS and latency metrics need Nitro-based instances. Check the AWS EBS metrics documentation for the current list.

How to read CloudWatch statistics and periods

A statistic aggregates samples over a period. Sum, Average, and Maximum answer different questions:

  • Sum of VolumeReadOps gives operations in the period
  • Maximum of a latency metric exposes the worst minute
  • Average smooths readings across the full window

EC2 monitoring graphs often display a 5-minute period, but publication stays at one minute. Retrieval periods should be at least the collection period.

Don't apply one threshold across volume types. A gp3 database volume and an st1 log volume behave nothing alike.

Which AWS EBS CloudWatch Metrics Matter?

Performance and workload-demand metrics

VolumeReadOps and VolumeWriteOps count completed operations. VolumeReadBytes and VolumeWriteBytes count data moved. Together they separate IOPS-heavy workloads (transactional databases) from throughput-heavy ones (sequential processing).

Derive the rates yourself: IOPS is operations divided by period seconds, and throughput is bytes divided by period seconds. You can also use the newer native metrics.

VolumeAvgIOPS (ops/s) and VolumeAvgThroughput (KiB/s) launched in October 2025 and require Nitro instances. They return zero when no relevant operations occurred.

Then compare against limits. Both the volume and the instance have ceilings:

Volume type Max IOPS Max throughput
gp3 80,000 2,000 MiB/s
gp2 16,000 250 MiB/s
io1 64,000 1,000 MiB/s
io2 Block Express 256,000 4,000 MiB/s
st1 500 500 MiB/s
sc1 250 250 MiB/s

An m5.large, for example, has a baseline of 650 Mbps and 3,600 IOPS, and can burst to 4,750 Mbps and 18,750 IOPS for 30 minutes at least once every 24 hours. A fast volume on a small instance is still slow.

Latency and exceeded-performance metrics

VolumeTotalReadTime and VolumeTotalWriteTime report time spent completing operations. On Nitro, read latency in milliseconds is 1000 * Sum(VolumeTotalReadTime) / Sum(VolumeReadOps).

Native VolumeAvgReadLatency and VolumeAvgWriteLatency give you this directly, with Minimum and Maximum statistics.

Three Nitro-only checks act as correlation signals. VolumeIOPSExceededCheck and VolumeThroughputExceededCheck return 1 when an application consistently tries to exceed provisioned performance. VolumeStalledIOCheck returns 1 on failure.

If latency rises while an exceeded check reads 1, the volume is undersized for the workload.

Saturation, queueing, and burst metrics

VolumeQueueLength counts requests awaiting completion. AWS suggests roughly one average queued I/O per 1,000 provisioned IOPS for consistency, so a 3,000 IOPS volume should sit near a depth of 3. Read it alongside I/O size and latency. Large sequential I/O on HDD volumes naturally produces high queue length.

BurstBalance shows remaining credits as a percentage: I/O credits on gp2, throughput credits on st1 and sc1. A steady decline means demand exceeds the sustained rate, and the volume is borrowing time.

AWS EBS saturation signals showing queue length and burst balance

Health, status, and utilization signals

VolumeStatus is not a CloudWatch metric. Volume status checks live in the EC2 console (Volumes, then Status checks) or in describe-volume-status, returning ok, impaired, warning, or insufficient-data.

Insufficient data can simply mean checks are ongoing. I/O performance status applies only to attached io1, io2, and gp3 volumes.

VolumeIdleTime reports seconds with no I/O submitted. It flags potentially overprovisioned volumes, but an idle volume can also mean a broken application or dependency.

Selecting a monitoring set by workload

Workload Start with Companion signal
Transactional database Avg IOPS, read/write latency Queue length, IOPS exceeded check
Sequential processing Avg throughput Throughput exceeded check, BurstBalance on st1/sc1
General application Avg IOPS, latency BurstBalance on gp2
Cost review VolumeIdleTime, ops counts Provisioned size, owner tags

Pick metrics by the question:

  • Slow? Latency plus queue length
  • Saturated? Demand against limits
  • Unhealthy? Status checks
  • Oversized or idle? Idle time and low I/O

How to Set Up AWS EBS Monitoring in CloudWatch

Find and graph per-volume EBS metrics

  1. Open CloudWatch, then Metrics, then All metrics.
  2. Select the EBS namespace, then Per-Volume Metrics.
  3. Search by volume ID and choose the metric.
  4. Confirm the Region, period, statistic, and unit before drawing conclusions.

You can also reach graphs from EC2, then Volumes, then the Monitoring tab.

Build a useful EBS dashboard

Dashboards are global, not Region-specific. Group widgets by demand, throughput, latency, saturation, and health. Put the volume ID, owner, environment, and provisioned limits in a text widget so anyone reading it has context.

Keep all widgets on the same time range so you can line up a latency spike against queue growth or a breached threshold.

Use metric math for derived views. IDs start lowercase, functions uppercase. With Sum metrics:

  • IOPS: (r+w)/PERIOD(r)
  • Bytes per second: (rb+wb)/PERIOD(rb)
  • Read latency: IF(r>0,1000*t/r,0)

Guard against division by zero, or data points will drop. AWS published a CloudFormation dashboard example creating 12 widgets for latency, IOPS, throughput, I/O size, idle time, and queue length.

Create CloudWatch alarms

Go to Alarms, then All alarms, then Create alarm. A math alarm must return a single time series. Key settings:

  • Sustained breaches: set EvaluationPeriods (N) and DatapointsToAlarm (M), for example 5 of 5 minutes
  • Missing data: notBreaching, breaching, ignore, or missing. Pick deliberately; detached volumes stop publishing
  • Baselines: set thresholds from observed workload and the volume's limits, not a generic number
  • Paired signals: alarm when IOPS is high and latency or queue length is also elevated

AWS EBS CloudWatch alarm configuration with sustained breach settings

Send notifications through SNS on ALARM, OK, or INSUFFICIENT_DATA. Be cautious with automated actions. Automatic resizing or failover needs validation, scoped permissions, cooldowns, and a rollback path.

Extend monitoring through the AWS CLI or APIs

aws cloudwatch get-metric-statistics --namespace AWS/EBS \
  --metric-name VolumeReadOps --dimensions Name=VolumeId,Value=VOLUME_ID \
  --statistics Sum --period 60 --start-time UTC_START --end-time UTC_END \
  --region us-east-1

Alarms and dashboards can be managed as code with AWS::CloudWatch::Alarm and AWS::CloudWatch::Dashboard in CloudFormation. Standardize tags, naming, ownership, and alert routing across accounts and Regions.

How to Interpret EBS Metrics and Troubleshoot Problems

Diagnose IOPS or throughput constraints

Work down this path:

  1. Compare observed IOPS and throughput to the volume's provisioned values.
  2. Compare the same figures to the instance's EBS limits.
  3. Check I/O size and the read/write pattern.
  4. Only then consider a resize, type change, or instance change.

A volume can show spare IOPS and still be throughput-bound. On SSD volumes, a 1,024-KiB application I/O splits into four 256-KiB operations, and at 256-KiB I/O, 1,000 IOPS equals 250 MiB/s.

Instance aggregate bandwidth may cap you either way. Validate current limits for your volume type first.

Diagnose latency with correlated signals

Overlay latency, queue length, IOPS, throughput, and the exceeded checks over the same window. Rising latency with an exceeded check at 1 points at the volume. Rising latency with flat IOPS and normal queue depth points elsewhere: the application, OS, network, or instance bandwidth.

AWS EBS latency troubleshooting correlation matrix for performance signals

Snapshot restores are another cause. Blocks initialize lazily, so latency rises until complete. Remedies include full-device reads, fast snapshot restore, or provisioned-rate initialization.

Troubleshooting sequence:

  1. Start with the affected application and its volume.
  2. Confirm dimensions, Region, and time range.
  3. Check volume status and I/O performance status.
  4. Expand to the instance and dependent services.

Use statistics appropriately: TM90 versus P90

A trimmed mean averages the data after discarding extremes. A percentile returns the value below which a share of observations fall.

CloudWatch notation matters: tm90 is valid and equals TM(:90%), which excludes the highest 10%. Uppercase TM90 is not valid. p90 is the 90th-percentile value, not a mean of retained samples.

  • p90 exposes tail behavior against a service-level target
  • tm90 shows typical performance without outlier noise

They aren't interchangeable. One caveat: CloudWatch's statistics definitions list EC2, not EBS, for advanced statistics, and native EBS latency is a per-minute average. Verify support in your account before building alarms on either.

Identify idle or overprovisioned storage

Combine VolumeIdleTime, low I/O, provisioned size, and ownership tags. AWS Compute Optimizer treats a volume as idle if it has fewer than one read+write operation per day over the lookback (14 days by default, up to 32), excluding root volumes. Unattached means no attachment for 32 days.

Never assume idle means safe to delete. Backup, disaster-recovery, staging, and periodic-use volumes look idle by design. Confirm ownership and retention, and snapshot first.

Move from observation to optimization

CloudWatch tells you what is happening. Acting on those signals through resizing, type changes, or performance adjustments needs a controlled process.

For enterprise fleets that need broader storage visibility and autonomous optimization across volumes and clouds, Lucidity adds that layer on top. CloudWatch remains the native AWS source of truth.

CloudWatch Costs, Retention, and Limitations

Are AWS EBS CloudWatch metrics free to use?

Native AWS/EBS metric publication is free. Other CloudWatch features are not. Examples from CloudWatch pricing for US East (N. Virginia):

  • Dashboards: 3 custom dashboards free (up to 50 metrics each), then billed
  • Standard alarms: $0.10 per alarm-month; math alarms bill underlying metrics
  • High-resolution alarms: $0.30 per alarm-month
  • Custom metrics: $0.30 per metric-month
  • GetMetricData: charged, and excluded from the API free tier

Rates vary by Region, feature, and usage. Check current pricing before rollout.

How long does CloudWatch retain metric data?

Resolution Retention
1 minute 15 days
5 minutes 63 days
1 hour 455 days (15 months)

Data rolls to coarser granularity, so 455 days does not mean minute-level history. Metrics with no new data for two weeks drop from console discovery. Alarm history, dashboards, and exported data in external systems follow their own retention rules.

Operational blind spots and governance

  • Metrics are Region-scoped; cross-account observability shares them within a Region
  • Detached volumes stop publishing, so alarms may see missing data
  • Newer metrics depend on Nitro and exclude Outposts, Local Zones, and Wavelength Zones
  • No native filesystem utilization metric exists

Add the CloudWatch Agent or application telemetry when you need filesystem capacity, mount health, or database latency. Enforce consistent tagging, ownership, alert routing, and periodic review so monitoring leads to action.

Frequently Asked Questions

What are AWS EBS CloudWatch metrics?

They're per-volume measurements in the AWS/EBS namespace. They cover IOPS, throughput, latency, queue length, burst balance, idle time, and performance checks.

What is the difference between TM90 and P90 for AWS EBS CloudWatch metrics?

tm90 is a trimmed mean excluding the top 10% of values; p90 is the 90th-percentile value. Trimmed means show typical behavior, percentiles show the tail, so they diverge on spiky latency.

Are AWS EBS CloudWatch metrics free?

Native publication is free. Dashboards beyond the free tier, alarms, custom metrics, and GetMetricData calls are charged. Check current pricing for your Region.

How long does AWS EBS CloudWatch keep metric data?

Retention depends on resolution: 15 days at 1 minute, 63 days at 5 minutes, and 455 days at 1 hour. Older data is coarser.

Which AWS EBS CloudWatch metrics should I monitor first?

Start with average IOPS, average throughput, read/write latency, and queue length. Add exceeded checks on Nitro, BurstBalance on gp2/st1/sc1, and VolumeIdleTime for cost reviews.

How do I create an alarm for an EBS volume in CloudWatch?

Select the metric, set the period and statistic, and pick a baseline-derived threshold. Set M-of-N evaluation, choose missing-data treatment, attach an SNS topic, and test it.