EFS Performance Modes and Throughput Modes

Amazon EFS offers two performance modes: General Purpose and Max I/O. These modes determine how the file system handles metadata operations and overall throughput. The choice impacts latency and scalability. General Purpose is the default and is ideal for latency-sensitive applications like web serving, content management, and home directories. Max I/O is designed for high-throughput, parallel workloads like big data analytics, media processing, and scientific computing.

EFS uses a distributed architecture with multiple storage servers. In General Purpose mode, metadata operations (e.g., file lookups, directory operations) are optimized for low latency. The file system uses a single metadata server to ensure consistency, which can become a bottleneck under high concurrent metadata operations. This mode supports up to 7,000 file operations per second (ops/s) and provides sub-millisecond latency for reads and writes.

In Max I/O mode, EFS scales horizontally by distributing metadata across multiple servers. This allows up to 50,000+ ops/s, but each metadata operation may have slightly higher latency (single-digit milliseconds). The trade-off is higher throughput for parallel workloads at the cost of increased latency for individual operations. Max I/O cannot be changed after creation.

EFS offers three throughput modes: Bursting, Provisioned, and Elastic. Throughput is the rate of data transfer (in MiB/s) the file system can sustain.

Bursting Throughput is the default. It uses a credit model similar to EBS gp2. Each file system accumulates throughput credits based on its size. The baseline throughput is 50 MiB/s per 1 TiB of storage. Credits are earned at the baseline rate and consumed when throughput exceeds the baseline. Burst credits allow short bursts up to 100 MiB/s per 1 TiB (for file systems up to 1 TiB). Larger file systems have a higher baseline, reducing burst dependency. The burst credit balance is visible in CloudWatch metrics (BurstCreditBalance). If credits are exhausted, throughput drops to the baseline.

Provisioned Throughput allows you to specify a fixed throughput independent of storage size. You pay for the provisioned amount, even if unused. This is useful for workloads that consistently need high throughput regardless of data stored. Provisioned throughput can be increased or decreased, but changes take effect within minutes.

Elastic Throughput (introduced in 2022) automatically scales throughput up and down based on workload demands. It uses machine learning to predict needs and adjusts without manual intervention. You pay only for the throughput used, with a per-GB-month charge. This mode is ideal for unpredictable workloads. It cannot be used with Provisioned mode.

Using the AWS CLI, you can create a file system with specific modes:

# Create a file system with General Purpose and Bursting (default)
aws efs create-file-system --creation-token my-fs --performance-mode generalPurpose --throughput-mode bursting --region us-east-1

# Create a file system with Max I/O and Provisioned throughput
aws efs create-file-system --creation-token my-fs2 --performance-mode maxIO --throughput-mode provisioned --provisioned-throughput-in-mibps 100 --region us-east-1

# Modify throughput mode (cannot change performance mode after creation)
aws efs update-file-system --file-system-id fs-12345678 --throughput-mode elastic

# Describe file system to see current settings
aws efs describe-file-systems --file-system-id fs-12345678

Output example:

{
    "FileSystems": [
        {
            "OwnerId": "123456789012",
            "CreationToken": "my-fs",
            "FileSystemId": "fs-12345678",
            "PerformanceMode": "generalPurpose",
            "ThroughputMode": "bursting",
            "ProvisionedThroughputInMibps": null,
            "SizeInBytes": {
                "Value": 1073741824,
                "Timestamp": "2025-01-01T00:00:00Z"
            }
        }
    ]
}

EFS performance modes and throughput modes interact with:

Burst credits are earned and consumed at a rate measured in MiB-seconds. The baseline throughput is (50 MiB/s) * (storage in TiB). For a 1 TiB file system, baseline = 50 MiB/s. If the file system uses 100 MiB/s for 10 seconds, it consumes 100 MiB/s * 10 s = 1000 MiB-s of credits. During that time, it earns 50 MiB/s * 10 s = 500 MiB-s, so net consumption is 500 MiB-s. The initial credit balance is 2.1 TiB-s (2.1 * 1024 * 1024 = 2,201,216 MiB-s). This allows bursting for (credit balance) / (burst rate - baseline). For a 1 TiB file system bursting at 100 MiB/s, burst duration = 2,201,216 / (100 - 50) = 44,024 seconds (~12.2 hours). For larger file systems, the baseline increases, so burst duration decreases for the same burst rate.