EFS Performance Modes and Throughput

Amazon EFS offers two performance modes—General Purpose (GP) and Max I/O—and two throughput modes—Bursting and Provisioned. These settings are chosen at file system creation (performance mode) or can be modified after creation (throughput mode). They exist because different workloads have vastly different access patterns: some need low latency for many small files (e.g., web serving, content management), while others need high aggregate throughput for large sequential operations (e.g., big data analytics, media processing). Choosing the wrong mode can lead to poor performance or unnecessary cost.

General Purpose (GP) is the default performance mode. It is optimized for latency-sensitive workloads that require strong consistency. Internally, GP mode uses a dual-metadata-server architecture. When a client performs a metadata operation (e.g., open, close, stat), the request goes to one of two metadata servers that maintain a strongly consistent view of the file system namespace. This dual-server setup provides low-latency metadata operations (typically single-digit milliseconds) and supports up to 7,000 file operations per second (ops/s) per file system. However, GP mode has a ceiling: it cannot scale beyond 7,000 metadata ops/s, making it unsuitable for workloads that require very high rates of small file access.

Max I/O mode is designed for workloads that need high levels of aggregate throughput and operations per second, especially for large files and sequential access. Instead of two metadata servers, Max I/O uses a distributed metadata architecture that scales horizontally. Each metadata operation is handled by one of many metadata servers, allowing the file system to achieve up to 500,000 ops/s or more. However, this scalability comes at a cost: Max I/O provides weaker consistency. While EFS always offers close-to-open consistency (i.e., once a file is closed, subsequent opens see the latest data), Max I/O may have slightly higher metadata latency on individual operations because the distributed system introduces coordination overhead. For workloads that can tolerate this (e.g., big data, genomics, media transcoding), Max I/O is ideal.

To create a file system with a specific performance mode via AWS CLI:

aws efs create-file-system --creation-token mytoken --performance-mode generalPurpose --throughput-mode bursting --region us-east-1

To change throughput mode to Provisioned:

aws efs update-file-system --file-system-id fs-12345678 --throughput-mode provisioned --provisioned-throughput-in-mibps 100

To view current settings:

aws efs describe-file-systems --file-system-id fs-12345678

The output includes PerformanceMode, ThroughputMode, and ProvisionedThroughputInMibps.

EFS performance interacts with several AWS services: - EC2 placement groups: For Max I/O, using a cluster placement group can reduce network latency between EC2 instances and EFS, improving throughput. - VPC endpoints: Using an EFS VPC endpoint (gateway endpoint) reduces data transfer costs and can improve latency by keeping traffic within the VPC. - EFS mount targets: Each mount target provides a network endpoint. For high throughput, distribute client traffic across multiple mount targets in different Availability Zones. - EFS replication: When using EFS replication (one-way), the destination file system inherits the performance mode of the source. Replication adds latency and can impact throughput. - AWS DataSync: For large-scale data transfers, DataSync can be used to migrate data to EFS, but it does not affect performance mode selection.

Bursting Throughput is the default and works on a credit model. The file system accumulates burst credits whenever it uses less than its baseline throughput. Credits are consumed when throughput exceeds the baseline. The burst rate is up to 100 MiB/s per TiB, but only for a limited duration. The burst credit bucket size is 2.1 TiB (approximately 2.2 hours of maximum burst for a 1 TiB filesystem). If credits are exhausted, throughput is throttled to the baseline. This model is cost-effective for workloads with variable throughput, such as periodic backups or batch processing.

Provisioned Throughput allows you to specify a fixed throughput independent of storage size. This is useful for workloads that consistently require high throughput but have small storage footprints (e.g., a 100 GiB filesystem needing 200 MiB/s). Without Provisioned, bursting would be limited because the baseline is only 50 MiB/s per TiB (i.e., ~5 MiB/s for 100 GiB). Provisioned costs more than Bursting for the same throughput level because you pay for the provisioned amount regardless of usage. You can change between Bursting and Provisioned at any time, but changes take effect within a few minutes.

Use Amazon CloudWatch metrics for EFS: - PermittedThroughput: The maximum throughput allowed based on bursting credits or provisioned setting. - BurstCreditBalance: The amount of burst credits remaining (in bytes). If this drops to zero, throughput is throttled. - MeteredIOBytes: The amount of data read/written (used for billing). - DataReadIOBytes, DataWriteIOBytes: I/O size metrics. - PercentIOLimit: How close you are to the metadata ops limit (for General Purpose).

If you see high latency or throttling (503 errors), check BurstCreditBalance and PercentIOLimit. If credits are low, consider switching to Provisioned Throughput or increasing storage (which increases baseline). If metadata ops are near 7,000 ops/s for General Purpose, consider moving to Max I/O (requires creating a new filesystem).

On the SOA-C02 exam, you will encounter questions that test your ability to:

Common wrong answers include: selecting Max I/O for low-latency workloads, thinking throughput mode is fixed at creation, or confusing burst credit calculations.

A large e-commerce company runs a web serving platform on hundreds of EC2 instances behind an Application Load Balancer. They use EFS to store static assets (HTML, CSS, JavaScript, images) that are frequently accessed by all instances. The workload involves millions of small files (average 10-50 KB) and requires low latency for each file open. They chose General Purpose performance mode because it provides low-latency metadata operations (single-digit milliseconds) and strong consistency. They use Bursting Throughput because traffic spikes during sales events (e.g., Black Friday) can be absorbed by burst credits. They monitor BurstCreditBalance closely and have a CloudWatch alarm to alert if credits drop below 20%. During peak, they temporarily switch to Provisioned Throughput to avoid throttling. A common misconfiguration is choosing Max I/O for this workload, which would increase metadata latency and cause poor user experience.

A biotech company processes genomic sequencing data using Apache Spark on EC2. Each sequencing run produces large files (1-10 GB) that need to be read and written sequentially. The cluster has 50+ nodes that frequently read and write these large files concurrently. They need high aggregate throughput (multiple GiB/s) to keep the cluster busy. They chose Max I/O performance mode because it can scale to hundreds of thousands of metadata ops/s and provides up to 10 GiB/s read throughput. They use Provisioned Throughput set to 5 GiB/s to ensure consistent performance regardless of storage size (which is only 5 TiB). Without Provisioned, Bursting would provide only 250 MiB/s baseline (50 MiB/s per TiB * 5 TiB), far below their needs. They also distribute mount targets across three Availability Zones to maximize network throughput. A pitfall is using General Purpose mode, which would limit metadata ops to ~7,000/s and throttle the Spark driver's directory listing operations.

A media company transcodes video files using a pipeline of EC2 instances. Input and output files are stored on EFS. Each transcoding job reads a large input file (2-5 GB) and writes multiple output files (hundreds of MB each). The pipeline runs continuously, processing hundreds of jobs per day. They chose Max I/O for high throughput and Provisioned Throughput (2 GiB/s) to handle the sustained load. They use EFS lifecycle management to move older files to Infrequent Access (IA) storage class to reduce costs. A common issue is misconfiguring mount options: using default rsize/wsize (128 KB) instead of 1 MB, which reduces throughput by 80%. They resolved this by setting rsize=1048576, wsize=1048576 in the mount command.