Performance metrics and specifications of Block Storage - Elastic Compute Service

Compare IOPS, throughput, and latency specifications across cloud disks, local disks, and elastic ephemeral disks to select the right storage for your workload.

Note

For pricing and billing details, see Block Storage billing.
For features and use cases, see Overview of Block Storage.

Performance metrics

Block Storage performance is measured by IOPS, throughput, and latency. For some products, performance scales with capacity. For example, ESSDs at different performance levels require different capacity ranges.

I/O size
I/O size is the data volume per read or write operation (for example, 4 KiB). IOPS × I/O size = Throughput. The metric to prioritize depends on your application's I/O size.

IOPS: The number of read/write operations a block storage device processes per second.

Focus on IOPS for latency-sensitive random small I/O workloads, such as database applications.

Note

Databases perform frequent insert, update, and delete operations. High IOPS keeps the system efficient under heavy random I/O and prevents bottleneck-induced latency.

Common IOPS metrics

Metric	Description	Data access method
Total IOPS	The total number of I/O operations performed per second.	Discontinuous and continuous access to storage locations on the disk.
Random read IOPS	The average number of random read I/O operations performed per second.	Discontinuous access to storage locations on the disk.
Random write IOPS	The average number of random write I/O operations performed per second.	Discontinuous access to storage locations on the disk.
Sequential read IOPS	The average number of sequential read I/O operations performed per second.	Continuous access to storage locations on the disk.
Sequential write IOPS	The average number of sequential write I/O operations performed per second.	Continuous access to storage locations on the disk.

Throughput: The amount of data transferred per second, measured in MB/s.
Focus on throughput for sequential large-I/O workloads, such as Hadoop offline analytics.
Note
Offline computing services such as Hadoop process petabytes of data. Low throughput significantly increases processing time and delays results.
Latency: The time a block storage device takes to complete an I/O operation, measured in s, ms, or μs. High latency degrades application performance.
For latency-sensitive applications such as databases, use low-latency products such as ESSD AutoPL disks or ESSDs.
Capacity: The size of the storage space. The unit is TiB, GiB, MiB, or KiB.
Block Storage capacity uses binary units (1 GiB = 1,024 MiB). Capacity itself is not a performance metric, but disk performance scales linearly with capacity up to the per-disk maximum for each type.

Disk performance

The following table compares disk types by performance.

Important

A disk's actual performance is limited by both the disk specifications and the attached instance specifications. See Storage I/O performance.
Standard SSDs, ultra disks, and basic disks are previous-generation products being phased out in some regions. Use PL0 ESSDs or ESSD Entry disks to replace ultra disks and basic disks, and ESSD AutoPL disks to replace standard SSDs.

Performance category	ESSD (Zone-redundant)	ESSD AutoPL	PL3 ESSD	PL2 ESSD	PL1 ESSD	PL0 ESSD	ESSD Entry	Standard SSD	Ultra Disk	Basic disk
Capacity range per disk (GiB)	10 to 65,536	1 to 65,536	1,261 to 65,536	461 to 65,536	20 to 65,536	1 to 65,536	10 to 32,768	20 to 32,768	20 to 32,768	5 to 2,000
Max IOPS	50,000	1,000,000	1,000,000	100,000	50,000	10,000	6,000	25,000②	5,000	Hundreds
Max throughput (MB/s)	350	4,096	4,000	750	350	180	150	300②	140	30 to 40
IOPS formula per disk①	min{1,800 + 50 × Capacity, 50,000}	Baseline performance: max{min{1,800 + 50 × Capacity, 50,000}, 3,000} Provisioned performance: Capacity (GiB) <= 3: Provisioned performance cannot be set. Capacity (GiB) >= 4: [1, min{(1,000 IOPS/GiB × Capacity - Baseline IOPS), 50,000}] Performance burst^③: Actual final IOPS - Baseline IOPS - Provisioned IOPS	min{1,800 + 50 × Capacity, 1,000,000}	min{1,800 + 50 × Capacity, 100,000}	min{1,800 + 50 × Capacity, 50,000}	min{1,800 + 12 × Capacity, 10,000}	min{1,800 + 8 × Capacity, 6,000}	min{1,800 + 30 × Capacity, 25,000}	min{1,800 + 8 × Capacity, 5,000}	N/A
Throughput formula per disk (MB/s) ①	min{120 + 0.5 × Capacity, 350}	Baseline performance: max{min{120 + 0.5 × Capacity, 350}, 125} Provisioned performance: 16 KB × Provisioned IOPS / 1,024 Performance burst^③: Actual final throughput - Baseline throughput - Provisioned throughput	min{120 + 0.5 × Capacity, 4,000}	min{120 + 0.5 × Capacity, 750}	min{120 + 0.5 × Capacity, 350}	min{100 + 0.25 × Capacity, 180}	min{100 + 0.15 × Capacity, 150}	min{120 + 0.5 × Capacity, 300}	min{100 + 0.15 × Capacity, 140}	N/A
Data reliability	99.9999999999%	99.9999999%
Average latency of single-path random writes (ms), Block Size=4K	Millisecond-level④	0.2	0.2	0.2	0.2	0.3 to 0.5	1 to 3	0.5 to 2	1 to 3	5 to 10

Baseline performance: The IOPS and throughput included with the disk at purchase. Scales linearly with capacity and varies by disk type.
Provisioned performance: Configure performance independently of capacity, decoupling the two.
① Notes on the formulas for single-disk performance:
- Formula for the maximum IOPS of a PL0 ESSD: Starts at 1,800, increases by 12 per GiB, and is capped at 10,000.
- Formula for the maximum throughput of a PL0 ESSD: Starts at 100 MB/s, increases by 0.25 MB/s per GiB, and is capped at 180 MB/s.
② The performance of a standard SSD varies with the block size:
- When IOPS is constant, a smaller block size results in lower throughput.
- When throughput is constant, a smaller block size results in higher IOPS.
I/O size (KiB)
Max IOPS
Throughput (MB/s)
4
About 25,000
About 100
16
About 17,200
About 260
32
About 9,600
About 300
64
About 4,800
About 300
③ ESSD AutoPL disks also support performance bursts beyond baseline and provisioned performance. Use EBS Lens (CloudLens for EBS) to monitor burst details in real time, including burst duration and total burst I/O. See Cloud disk analysis.
④ Data on an ESSD (Zone-redundant) disk is distributed across multiple zones for an RPO of 0 through physical replication. Synchronous multi-zone writes increase latency beyond PL1 ESSD levels; actual latency varies by region. Test write latency by following Test for block storage devices.

Local disk performance

Warning

Local disks cannot be created independently. Their data reliability depends on the physical server, which introduces single-point-of-failure risk. A physical server failure can affect multiple running instances. Do not store critical long-term data on local disks. See Local disks.

NVMe SSD local disks

NVMe SSD local disk performance for the d3c instance family:

Metric	Performance per disk	ecs.d3c.3xlarge	ecs.d3c.7xlarge	ecs.d3c.14xlarge
Max read IOPS	100,000	100,000	200,000	400,000
Max read throughput	4 GB/s	4 GB/s	8 GB/s	16 GB/s
Max write throughput	2 GB/s	2 GB/s	4 GB/s	8 GB/s
Latency	Microsecond-level (μs)

NVMe SSD local disk performance for the i5e instance family:

NVMe SSD metric	ecs.i5e.2xlarge	ecs.i5e.4xlarge	ecs.i5e.8xlarge	ecs.i5e.12xlarge	ecs.i5e.16xlarge	ecs.i5e.32xlarge
Max read IOPS	1,400,000	2,900,000	5,800,000	8,700,000	11,600,000	23,200,000
Max read throughput	7 GB/s	14 GB/s	28 GB/s	42 GB/s	56 GB/s	112 GB/s
Max write throughput	4.5 GB/s	9 GB/s	18 GB/s	27 GB/s	36 GB/s	72 GB/s
Latency	Microsecond-level (μs)

NVMe SSD local disk performance for the i5 instance family:

NVMe SSD metric	ecs.i5.xlarge	ecs.i5.2xlarge	ecs.i5.4xlarge	ecs.i5.8xlarge	ecs.i5.12xlarge	ecs.i5.16xlarge
Max read IOPS	700,000	1,400,000	2,900,000	5,800,000	8,700,000	11,800,000
Max read throughput	3.5 GB/s	7 GB/s	14 GB/s	28 GB/s	42 GB/s	56 GB/s
Max write throughput	2 GB/s	4 GB/s	8 GB/s	16 GB/s	24 GB/s	32 GB/s
Latency	Microsecond-level (μs)

NVMe SSD local disk performance for the i5g instance family:
NVMe SSD metric
ecs.i5g.8xlarge
ecs.i5g.16xlarge
Max read IOPS
1,400,000
2,900,000
Max read throughput
7 GB/s
14 GB/s
Max write throughput
4 GB/s
8 GB/s
Latency
Microsecond-level (μs)

NVMe SSD local disk performance for the i5ge instance family:

NVMe SSD metric	ecs.i5ge.3xlarge	ecs.i5ge.6xlarge	ecs.i5ge.12xlarge	ecs.i5ge.24xlarge
Max read IOPS	1,400,000	2,900,000	5,800,000	11,800,000
Max read throughput	7 GB/s	14 GB/s	28 GB/s	56 GB/s
Max write throughput	4 GB/s	8 GB/s	16 GB/s	32 GB/s
Latency	Microsecond-level (μs)

NVMe SSD local disk performance for the i4 instance family:

NVMe SSD metric	ecs.i4.large	ecs.i4.xlarge	ecs.i4.2xlarge	ecs.i4.4xlarge	ecs.i4.8xlarge	ecs.i4.16xlarge	ecs.i4.32xlarge
Max read IOPS	112,500	225,000	450,000	900,000	1,800,000	3,600,000	7,200,000
Max read throughput	0.75 GB/s	1.5 GB/s	3 GB/s	6 GB/s	12 GB/s	24 GB/s	48 GB/s
Max write throughput	0.375 GB/s	0.75 GB/s	1.5 GB/s	3 GB/s	6 GB/s	12 GB/s	24 GB/s
Latency	Microsecond-level (μs)

Note

Values in the table represent optimal performance. To achieve these values, use the latest Linux image version, such as Alibaba Cloud Linux 3. This instance family supports only Linux images.

NVMe SSD local disk performance for the i4g and i4r instance families:

NVMe SSD metric	ecs.i4g.4xlarge and ecs.i4r.4xlarge	ecs.i4g.8xlarge and ecs.i4r.8xlarge	ecs.i4g.16xlarge and ecs.i4r.16xlarge	ecs.i4g.32xlarge and ecs.i4r.32xlarge
Max read IOPS	250,000	500,000	1,000,000	2,000,000
Max read throughput	1.5 GB/s	3 GB/s	6 GB/s	12 GB/s
Max write throughput	1 GB/s	2 GB/s	4 GB/s	8 GB/s
Latency	Microsecond-level (μs)

Note

Values in the table represent optimal performance. To achieve these values, use the latest Linux image version, such as the images described in Alibaba Cloud Linux 3 image release notes. This instance family supports only Linux images.

NVMe SSD local disk performance for the i3 instance family:

NVMe SSD metric	ecs.i3.xlarge	ecs.i3.2xlarge	ecs.i3.4xlarge	ecs.i3.8xlarge	ecs.i3.13xlarge	ecs.i3.26xlarge
Max read IOPS	250,000	500,000	1,000,000	2,000,000	3,000,000	6,000,000
Max read throughput	1.5 GB/s	3 GB/s	6 GB/s	12 GB/s	18 GB/s	36 GB/s
Max write throughput	1 GB/s	2 GB/s	4 GB/s	8 GB/s	12 GB/s	24 GB/s
Latency	Microsecond-level (μs)

Note

NVMe SSD local disk performance for the i3g instance family:

NVMe SSD metric	ecs.i3g.2xlarge	ecs.i3g.4xlarge	ecs.i3g.8xlarge	ecs.i3g.13xlarge	ecs.i3g.26xlarge
Max read IOPS	125,000	250,000	500,000	750,000	1,500,000
Max read throughput	0.75 GB/s	1.5 GB/s	3 GB/s	4.5 GB/s	9 GB/s
Max write throughput	0.5 GB/s	1 GB/s	2 GB/s	3 GB/s	6 GB/s
Latency	Microsecond-level (μs)

Note

NVMe SSD local disk performance for the i2 and i2g instance families:

NVMe SSD metric	Performance per disk		Overall instance performance①
NVMe SSD metric	ecs.i2.xlarge and ecs.i2g.2xlarge only	Other i2 and i2g instance types	Overall instance performance①
Max capacity	894 GiB	1,788 GiB	8 × 1,788 GiB
Max read IOPS	150,000	300,000	1,500,000
Max read throughput	1 GB/s	2 GB/s	16 GB/s
Max write throughput	0.5 GB/s	1 GB/s	8 GB/s
Latency	Microsecond-level (μs)

① Overall instance performance applies only to ecs.i2.16xlarge — the largest instance type in the i2 family.

NVMe SSD local disk performance for the i2ne and i2gne instance families:

NVMe SSD metric	ecs.i2ne.xlarge and ecs.i2gne.2xlarge	ecs.i2ne.2xlarge and ecs.i2gne.4xlarge	ecs.i2ne.4xlarge and ecs.i2gne.8xlarge	ecs.i2ne.8xlarge and ecs.i2gne.16xlarge	ecs.i2ne.16xlarge
Max capacity	894 GiB	1,788 GiB	2 × 1,788 GiB	4 × 1,788 GiB	8 × 1,788 GiB
Max read IOPS	250,000	500,000	1,000,000	2,000,000	4,000,000
Max read throughput	1.5 GB/s	3 GB/s	6 GB/s	12 GB/s	24 GB/s
Max write throughput	1 GB/s	2 GB/s	4 GB/s	8 GB/s	16 GB/s
Latency	Microsecond-level (μs)

NVMe SSD local disk performance for the i1 instance family:

NVMe SSD metric	Performance per disk	Overall instance performance ②
Max capacity	1,456 GiB	2,912 GiB
Max IOPS	240,000	480,000
Write IOPS ①	min{165 × Capacity, 240,000}	2 × min{165 × Capacity, 240,000}
Read IOPS ①	min{165 × Capacity, 240,000}	2 × min{165 × Capacity, 240,000}
Max read throughput	2 GB/s	4 GB/s
Read throughput ①	min{1.4 × Capacity, 2,000} MB/s	2 × min{1.4 × Capacity, 2,000} MB/s
Max write throughput	1.2 GB/s	2.4 GB/s
Write throughput ①	min{0.85 × Capacity, 1,200} MB/s	2 × min{0.85 × Capacity, 1,200} MB/s
Latency	Microsecond-level (μs)

① Notes on the formulas for single-disk performance:

Example of the formula for the write IOPS of a single NVMe SSD local disk: 165 IOPS per GiB, with a cap of 240,000 IOPS.
Example of the formula for the write throughput of a single NVMe SSD local disk: 0.85 MB/s per GiB, with a cap of 1,200 MB/s.

② Overall instance performance applies only to ecs.i1.14xlarge — the largest instance type in the i1 family.

SATA HDD local disks

SATA HDD local disk performance:

SATA HDD metric	d1, d1ne		d2c		d2s		d3s
SATA HDD metric	Performance per disk	Overall instance performance	Performance per disk	Overall instance performance	Performance per disk	Overall instance performance	Performance per disk	Overall instance performance
Max capacity	5,500 GiB	154,000 GiB	3,700 GiB	44,400 GiB	7,300 GiB	219,000 GiB	11,100 GiB	355,200 GiB
Max throughput	190 MB/s	5,320 MB/s	190 MB/s	2,280 MB/s	190 MB/s	5,700 MB/s	260 MB/s	8,320 MB/s
Latency	Millisecond-level (ms)

Note

Overall instance performance applies only to ecs.d1.14xlarge, ecs.d1ne.14xlarge, ecs.d2c.24xlarge, ecs.d2s.20xlarge, and ecs.d3s.16xlarge — the largest instance type in each respective family.

Elastic ephemeral disk performance

Note

Customize elastic ephemeral disk capacity for temporary data storage. See Elastic ephemeral disk.

Elastic ephemeral disks come in two categories: standard (large capacity, high throughput) and premium (small capacity, high IOPS).

Metric	Standard elastic ephemeral disks	Premium elastic ephemeral disks
Single-disk capacity range (GiB)	64 to 8,192	64 to 8,192
Maximum read IOPS per disk	Either 100 times the capacity or 820,000, whichever is smaller	Either 300 times the capacity or 1,000,000, whichever is smaller
Maximum write IOPS per disk	Either 20 times the capacity or 160,000, whichever is smaller	Either 150 times the capacity or 500,000, whichever is smaller
Maximum read throughput per disk (MB/s)	Either 0.8 times the capacity or 4,096, whichever is smaller	Either 1.6 times the capacity or 4,096, whichever is smaller
Maximum write throughput per disk (MB/s)	Either 0.4 times the capacity or 2,048, whichever is smaller	Either the capacity or 2,048, whichever is smaller
Write I/O density①	20	150
Read I/O density①	100	300

①: I/O density = IOPS / disk capacity, unit: IOPS/GiB, indicating the IOPS capability per GiB.

Test Block Storage performance

Test Block Storage performance with the following tools:

Troubleshoot slow disk I/O

Monitor disk performance in the ECS, EBS, or CloudMonitor console to identify bottlenecks. See View monitoring information for a disk.

Check whether the disk uses pay-as-you-go billing. If so, I/O speed is throttled when your account has an overdue payment and restores after you settle the balance.
Note: If you do not renew a pay-as-you-go disk within 15 days after your payment becomes overdue, the disk is automatically released, and its data cannot be recovered.
On Linux, see Troubleshoot high disk I/O usage on a Linux instance to identify high-IOPS programs.
During data import, both client and server performance affect read/write speeds.
Use the atop tool to monitor Linux system metrics and check resource usage logs in /var/log/atop to locate the problem.
If performance is still insufficient, see How to improve disk performance.

How to improve disk performance

If disk performance is insufficient, try the following methods:

Important

A disk's actual performance is capped by both the disk and the instance specifications. If the instance type's IOPS and bandwidth limits are lower than the disk's, upgrading only the disk does not help — you must also upgrade the instance type. See Instance families.

Scenarios	Method to improve performance
Your current disk type (for example, standard SSD) can no longer meet growing IOPS or throughput demands. Switch to a higher-performance type such as a PL1 ESSD for better IOPS and response times.	Change the category of a disk
Adjust the ESSD performance level based on workload changes.	Modify the performance level of an ESSD
For ESSD AutoPL disks, set provisioned performance or enable performance bursting.	Modify the performance configuration of an ESSD AutoPL disk
Scale out the disk to increase both capacity and baseline IOPS. For disk types like PL1 ESSDs, baseline IOPS grows with capacity. For example, a 40 GiB PL1 ESSD has 3,800 IOPS; scaling to 100 GiB increases IOPS to 6,800 (formula: min{1,800 + 50 × Capacity, 50,000}).	Scale out a disk
Use LVM to distribute data across multiple logical volumes for parallel I/O. Suitable for multi-threaded applications and databases requiring high-concurrency access.	Create a logical volume
Create a RAID array to improve IOPS and throughput with optional data redundancy. Use RAID 0 for speed, or RAID 1/RAID 10 for performance with redundancy.

I/O size (KiB)	Max IOPS	Throughput (MB/s)
4	About 25,000	About 100
16	About 17,200	About 260
32	About 9,600	About 300
64	About 4,800	About 300

NVMe SSD metric	ecs.i5g.8xlarge	ecs.i5g.16xlarge
Max read IOPS	1,400,000	2,900,000
Max read throughput	7 GB/s	14 GB/s
Max write throughput	4 GB/s	8 GB/s
Latency	Microsecond-level (μs)