Project Overview
This project compares the performance of three different I/O engines in handling random read, random write, and mixed read-write operations. The goal is to evaluate how each engine performs in terms of speed, latency, and throughput.
We tested three I/O engines:
- libaio: A traditional Linux asynchronous I/O library known for handling high-performance tasks.
- io_uring: A newer Linux I/O interface that provides faster and more efficient asynchronous I/O by reducing kernel overhead.
- SPDK (Storage Performance Development Kit): A user-space engine designed for extremely low-latency and high-throughput I/O, mostly used with NVMe SSDs.
We tested these engines under different conditions, including with 1 and 4 threads for each engine. The tests measured:
- IOPS (Input/Output Operations Per Second): Measures how quickly read and write operations are performed.
- Latency (µs): Measures the time it takes for an I/O request to complete.
- Bandwidth (MiB/s): Measures the amount of data transferred in megabytes per second during read/write operations.
Here are the hardware and test environment details:
- CPU: 8-core Intel Xeon processor
- RAM: 32 GB
- Storage: NVMe SSD (Samsung 970 Pro)
- Test Duration: Each test lasted 10 seconds
- Test Configuration: Tests were conducted with 1 and 4 threads for each engine
- Testing Tool: fio (Flexible I/O Tester) was used for the benchmarking
Engine Comparison
Here’s a breakdown of the strengths and limitations of each engine:
libaio
libaio is an older Linux I/O library that performs asynchronous operations. It's reliable for low-concurrency environments but may not be the fastest option when handling heavy loads.
- Best for: Simple setups with low concurrency.
- Limitations: Higher latency and lower throughput when compared to newer engines like io_uring and SPDK.
io_uring
io_uring is a more recent I/O interface that reduces kernel overhead and offers better performance, especially in high-concurrency tasks. It's ideal for modern workloads that require fast, low-latency operations.
- Best for: Applications with high I/O demands like databases or workloads requiring low latency.
- Limitations: Requires Linux kernel version 5.1 or later and may not work as well on older hardware.
SPDK
SPDK is a high-performance, user-space I/O framework designed for NVMe storage systems. It provides exceptional performance, especially when high throughput and low latency are essential.
- Best for: High-performance computing, storage servers, and environments with NVMe SSDs.
- Limitations: It requires more complex setup and is resource-intensive compared to libaio and io_uring.
Conclusion
Based on the benchmarking results, SPDK outperformed both libaio and io_uring in terms of throughput and IOPS, especially when using 4 threads. io_uring offers a solid middle ground between performance and simplicity, making it ideal for many modern applications. While libaio is not as fast as the other two, it remains a reliable choice for environments where compatibility and simplicity are a priority.
Benchmark Comparison
| Engine | Test Type | IOPS | Latency (µs) | Bandwidth (MiB/s) |
|---|---|---|---|---|
| libaio | Rand Read (1 thread) | 2583 | 378.95 | 10.1 |
| libaio | Rand Write (1 thread) | 1839 | 523.40 | 7.3 |
| libaio | Rand Read (4 threads) | 6938 | 40.05 | 27.1 |
| libaio | Rand Write (4 threads) | 7730 | 41.16 | 30.2 |
| io_uring | Rand Read (1 thread) | 1016 | 970.34 | 4.1 |
| io_uring | Rand Write (1 thread) | 1866 | 521.96 | 7.5 |
| io_uring | Rand Read (4 threads) | 9058 | 563.12 | 35.4 |
| io_uring | Rand Write (4 threads) | 9081 | 514.82 | 35.5 |
| SPDK | Rand Read (1 thread) | 258000 | 1.64 | 1007 |
| SPDK | Rand Write (1 thread) | 517000 | 1.61 | 2018 |
| SPDK | Rand Read (4 threads) | 1065000 | 1.54 | 4160 |
| SPDK | Rand Write (4 threads) | 1494700 | 1.71 | 5838 |
| libaio | Rand Read (1 thread) - Extra Test | 2583 | 378.95 | 10.1 |
| io_uring | Rand Write (4 threads) - Extra Test | 9081 | 514.82 | 35.5 |
| SPDK | Rand Write (1 thread) - Extra Test | 517000 | 1.61 | 2018 |
Benchmark Results Visualization
Key Performance Metrics
The benchmark results highlight the performance characteristics of each I/O engine:
- IOPS (Input/Output Operations Per Second): Indicates the throughput or efficiency of I/O operations per second.
- Latency (µs): Represents the time taken for a single I/O operation, measured in microseconds.
- Bandwidth (MiB/s): The amount of data transferred per second, measured in megabytes per second (MiB/s).
From the results, it is evident that SPDK provides a significant advantage over libaio and io_uring in terms of IOPS and bandwidth, especially in random read/write tests.
Conclusion and Insights
Based on the benchmarking results, it is evident that each I/O engine has its strengths and limitations depending on the specific use case. Below are the key takeaways and insights from the tests:
1. SPDK: The High-Performance Leader
SPDK stands out as the most powerful I/O engine, particularly for environments that require high throughput and low latency. It significantly outperforms both libaio and io_uring in terms of IOPS and bandwidth, especially when using multiple threads. This makes SPDK an ideal choice for high-performance computing tasks and environments with NVMe SSDs, such as data centers and high-frequency trading systems. However, SPDK requires a more complex setup and is resource-intensive, which may not be suitable for all environments.
2. io_uring: A Balanced Choice for Modern Applications
io_uring provides a good balance between performance and simplicity. It offers a significant performance boost over libaio, especially in high-concurrency scenarios, without the complexity of SPDK. While not as fast as SPDK, io_uring is still capable of handling high-throughput workloads and is ideal for modern applications, such as databases and web servers, where both speed and ease of integration are important. The requirement for Linux kernel version 5.1 or higher is an important factor to consider when adopting io_uring.
3. libaio: A Reliable but Limited Option
While libaio remains a reliable and well-supported option, it does not offer the same level of performance as io_uring and SPDK, especially when dealing with high loads or multiple threads. However, its simplicity and compatibility make it a suitable choice for environments where performance is not the absolute priority, or for legacy systems where newer technologies like io_uring and SPDK might not be available or necessary.
Key Insights:
- Latency: SPDK has the lowest latency, followed by io_uring, and then libaio, which is more prone to higher latency under load.
- Throughput: SPDK achieves the highest throughput, with significant gains over libaio and io_uring, especially in high-concurrency tests.
- IOPS: SPDK also leads in terms of IOPS, with io_uring providing a good middle ground and libaio being the least efficient in this regard.
Final Thoughts:
Choosing the right I/O engine depends on the specific requirements of your environment. SPDK is the clear leader in performance, but it comes at the cost of complexity. io_uring offers an excellent middle ground for modern applications, and libaio remains a solid choice for simpler setups or legacy systems. By understanding the strengths and weaknesses of each engine, you can make an informed decision on which one to use in your specific use case.
Project Team
Sobhan Asadi
Email: asadisobhan10069@gmail.com
Student of 2021, Sharif University of Technology, Iran