Amazon EBS I/O characteristics and monitoring - Amazon EBS

Amazon EBS I/O characteristics and monitoring

On a given volume configuration, certain I/O characteristics drive the performance behavior for your EBS volumes. SSD-backed volumes—General Purpose SSD (gp2 and gp3) and Provisioned IOPS SSD (io1 and io2)—deliver consistent performance whether an I/O operation is random or sequential. HDD-backed volumes—Throughput Optimized HDD (st1) and Cold HDD (sc1)—deliver optimal performance only when I/O operations are large and sequential. To understand how SSD and HDD volumes will perform in your application, it is important to know the connection between demand on the volume, the quantity of IOPS available to it, the time it takes for an I/O operation to complete, and the volume's throughput limits.

IOPS

IOPS are a unit of measure representing input/output operations per second. The operations are measured in KiB, and the underlying drive technology determines the maximum amount of data that a volume type counts as a single I/O. I/O size is capped at 256 KiB for SSD volumes and 1,024 KiB for HDD volumes because SSD volumes handle small or random I/O much more efficiently than HDD volumes.

When small I/O operations are physically sequential, Amazon EBS attempts to merge them into a single I/O operation up to the maximum I/O size. Similarly, when I/O operations are larger than the maximum I/O size, Amazon EBS attempts to split them into smaller I/O operations. The following table shows some examples.

Volume type Maximum I/O size I/O operations from your application Number of IOPS Notes
SSD 256 KiB 1 x 1024 KiB I/O operation 4 (1,024÷256=4) Amazon EBS splits the 1,024 I/O operation into four smaller 256 KiB operations.
8 x sequential 32 KiB I/O operations 1 (8x32=256) Amazon EBS merges the eight sequential 32 KiB I/O operations into a single 256 KiB operation.
8 random 32 KiB I/O operations 8 Amazon EBS counts random I/O operations separately.
HDD 1,024 KiB 1 x 1024 KiB I/O operation 1 The I/O operation is already equal to the maximum I/O size. It is not merged or split.
8 x sequential 128 KiB I/O operations 1 (8x128=1,024) Amazon EBS merges the eight sequential 128 KiB I/O operations into a single 1,024 KiB I/O operation.
8 random 32 KiB I/O operations 8 Amazon EBS counts random I/O operations separately.

Consequently, when you create an SSD-backed volume supporting 3,000 IOPS (either by provisioning an io1 or io2 volume with 3,000 IOPS, by sizing a gp2 volume at 1,000 GiB, or by using a gp3 volume), and you attach it to an EBS-optimized instance that can provide sufficient bandwidth, you can transfer up to 3,000 I/Os of data per second, with throughput determined by I/O size.

Volume queue length and latency

The volume queue length is the number of pending I/O requests for a device. Latency is the true end-to-end client time of an I/O operation, in other words, the time elapsed between sending an I/O to EBS and receiving an acknowledgement from EBS that the I/O read or write is complete. Queue length must be correctly calibrated with I/O size and latency to avoid creating bottlenecks either on the guest operating system or on the network link to EBS.

Optimal queue length varies for each workload, depending on your particular application's sensitivity to IOPS and latency. If your workload is not delivering enough I/O requests to fully use the performance available to your EBS volume, then your volume might not deliver the IOPS or throughput that you have provisioned.

Transaction-intensive applications are sensitive to increased I/O latency and are well-suited for SSD-backed volumes. You can maintain high IOPS while keeping latency down by maintaining a low queue length and a high number of IOPS available to the volume. Consistently driving more IOPS to a volume than it has available can cause increased I/O latency.

Throughput-intensive applications are less sensitive to increased I/O latency, and are well-suited for HDD-backed volumes. You can maintain high throughput to HDD-backed volumes by maintaining a high queue length when performing large, sequential I/O.

I/O size and volume throughput limits

For SSD-backed volumes, if your I/O size is very large, you may experience a smaller number of IOPS than you provisioned because you are hitting the throughput limit of the volume. For example, a gp2 volume under 1,000 GiB with burst credits available has an IOPS limit of 3,000 and a volume throughput limit of 250 MiB/s. If you are using a 256 KiB I/O size, your volume reaches its throughput limit at 1000 IOPS (1000 x 256 KiB = 250 MiB). For smaller I/O sizes (such as 16 KiB), this same volume can sustain 3,000 IOPS because the throughput is well below 250 MiB/s. (These examples assume that your volume's I/O is not hitting the throughput limits of the instance.) For more information about the throughput limits for each EBS volume type, see Amazon EBS volume types.

For smaller I/O operations, you may see a higher-than-provisioned IOPS value as measured from inside your instance. This happens when the instance operating system merges small I/O operations into a larger operation before passing them to Amazon EBS.

If your workload uses sequential I/Os on HDD-backed st1 and sc1 volumes, you may experience a higher than expected number of IOPS as measured from inside your instance. This happens when the instance operating system merges sequential I/Os and counts them in 1,024 KiB-sized units. If your workload uses small or random I/Os, you may experience a lower throughput than you expect. This is because we count each random, non-sequential I/O toward the total IOPS count, which can cause you to hit the volume's IOPS limit sooner than expected.

Whatever your EBS volume type, if you are not experiencing the IOPS or throughput you expect in your configuration, ensure that your EC2 instance bandwidth is not the limiting factor. You should always use a current-generation, EBS-optimized instance (or one that includes 10 Gb/s network connectivity) for optimal performance. Another possible cause for not experiencing the expected IOPS is that you are not driving enough I/O to the EBS volumes.

Monitor I/O characteristics using CloudWatch

You can monitor these I/O characteristics with each volume's CloudWatch volume metrics. Important metrics to consider include the following:

  • VolumeStalledIOCheck

  • BurstBalance

  • VolumeReadBytes | VolumeWriteBytes

  • VolumeReadOps | VolumeWriteOps

  • VolumeQueueLength

VolumeStalledIOCheck monitors the status of your EBS volumes to determine when your volumes are impaired. The metric is a binary value that will return a 0 (pass) or a 1 (fail) status based on whether or not the EBS volume can complete I/O operations. This check detects underlying issues with the Amazon EBS infrastructure, such as the following:

  • Hardware or software issues on the storage subsystems underlying the EBS volumes

  • Hardware issues on the physical host that impact reachability of the EBS volumes from your EC2 instance

  • Connectivity issues between the instance and EBS volumes

If the VolumeStalledIOCheck metric fails, you can either wait for AWS to resolve the issue, or you can take actions, such as replacing the affected volume or stopping and restarting the instance to which the volume is attached. In most cases, when this metric fails, EBS will automatically diagnose and recover your volume within a few minutes. You can use the Pause I/O action in AWS Fault Injection Service to run controlled experiments to test your architecture and monitoring based on this metric to improve your resiliency to storage faults.

You can measure Amazon EBS storage I/O latency using VolumeReadOps, VolumeWriteOps, VolumeTotalReadTime, and VolumeTotalWriteTime. You can use the following formula to monitor the average I/O latency of your volume:

Average I/O latency in ms/op = (VolumeTotalReadTime + VolumeTotalWriteTime) / (VolumeReadOps + VolumeWriteOps)

If your I/O latency is higher than you require, check your driven IOPS and make sure that your application is not trying to drive more IOPS than you have provisioned. You can use the following formula to monitor the average driven IOPS on your volume:

Estimated average IOPS in ops/s = (Sum(VolumeReadOps) + Sum(VolumeWriteOps)) / (Period - Sum(VolumeIdleTime))

If your application requires a greater number of IOPS than your volume can provide, you should consider using one of the following:

  • A gp3, io2, or io1 volume that is provisioned with enough IOPS to achieve the required latency

  • A larger gp2 volume that provides enough baseline IOPS performance

HDD-backed st1 and sc1 volumes are designed to perform best with workloads that take advantage of the 1,024 KiB maximum I/O size. To determine your volume's average I/O size, divide VolumeWriteBytes by VolumeWriteOps. The same calculation applies to read operations. If average I/O size is below 64 KiB, increasing the size of the I/O operations sent to an st1 or sc1 volume should improve performance.

Note

If average I/O size is at or near 44 KiB, you might be using an instance or kernel without support for indirect descriptors. Any Linux kernel 3.8 and above has this support, as well as any current-generation instance.

BurstBalance displays the burst bucket balance for gp2, st1, and sc1 volumes as a percentage of the remaining balance. When your burst bucket is depleted, volume I/O (for gp2 volumes) or volume throughput (for st1 and sc1 volumes) is throttled to the baseline. Check the BurstBalance value to determine whether your volume is being throttled for this reason. For a complete list of the available Amazon EBS metrics, see Amazon CloudWatch metrics for Amazon EBS and Amazon EBS metrics for Nitro-based instances.

Related resources

For more information about Amazon EBS I/O characteristics, see the following re:Invent presentation: Amazon EBS: Designing for Performance.